NATO ASI Series

Advanced Science Institutes Series

A series presenting the results of activities sponsored by the NATO Science
Committee, which aims at the dissemination of advanced scientific and technological
knowledge, with a view to strengthening links between scientific communities.
The Series is published by an international board of publishers in conjunction with the
NATO Scientific Affairs Division

A Life Sciences Plenum Publishing Corporation

B Physics London and New York

C Mathematical and Physical Sciences Kluwer Academic Publishers

D Behavioural and Social Sciences Dordrecht, Boston and London
E Applied Sciences

F Computer and Systems Sciences Springer-Verlag

G Ecological Sciences Berlin Heidelberg New York Barcelona
H Cell Biology Budapest Hong Kong London Milan
Global Environmental Change Paris Santa Clara Singapore Tokyo

PARTNERSHIP SUB-SERIES
1. Disarmament Technologies Kluwer Academic Publishers
2. Environment Springer-Verlag
3. High Technology Kluwer Academic Publishers
4. Science and Technology Policy Kluwer Academic Publishers
5. Computer Networking Kluwer Academic Publishers

The Partnership Sub-Series incorporates activities undertaken in collaboration with

NATO's Cooperation Partners, the countries of the CIS and Central and Eastern
Europe, in Priority Areas of concern to those countries.

NATO-PCO DATABASE
The electronic index to the NATO ASI Series provides full bibliographical references
(with keywords and/or abstracts) to about 50000 contributions from international
scientists published in all sections of the NATO ASI Series. Access to the NATO-PCO
DATABASE compiled by the NATO Publication Coordination Office is possible in two
ways:

- via online FILE 128 (NATO-PCO DATABASE) hosted by ESRIN,

Via Galileo Galilei, 1-00044 Frascati, Italy.

- via CD-ROM "NATO Science & Technology Disk" with user-friendly retrieval software
in English, French and German (© wrv GmbH and DATAWARE Technologies Inc.
1992).

The CD-ROM can be ordered through any member of the Board of Publishers or
through NATO-PCO, Overijse, Belgium.

Series F: Computer and Systems Sciences, Vol. 143

Springer
Berlin
Heidelberg
New York
Barcelona
Budapest
Hong Kong
London
Milan
Paris
Santa Clara
Singapore
Tokyo
Batch Processing
Systems Engineering
Fundamentals and Applications
for Chemical Engineering

Edited by

Gintaras V. Reklaitis
School of Chemical Engineering, Purdue University
West Lafayette, IN 47907, USA

Aydin K. Sunol
Department of Chemical Engineering, College of Engineering
University of South Florida, 4202 East Fowler Avenue, ENG 118
Tampa, FL 33620-5350, USA

David W. T. Rippint
Laboratory for Technical Chemistry, Eidgenossische
Technische Hochschule (ETH) Zurich, Switzerland

Oner Hortagsu
Department of Chemical Engineering, Bogazi<;i University
TR-80815 Bebek-Istanbul, Turkey

Springer
Published in cooperation with NATO Scientific Affairs Division
Proceedings of the NATO Advanced Study Institute on Batch Processing
Systems Engineering: Current Status and Future Directions, held in Antalya,
Turkey, May 29 - June 7,1992

Library of Congress Cataloging-in-Publication data applied for

CR Subject Classification (1991): J.6, 1.6, J.2, G.1, J.7, 1.2

ISBN-13: 978-3-642-64635-5 e-ISBN-13: 978-3-642-60972-5

DO I: 10.1007/978-3-642-60972-5
Springer-Verlag Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights oftranslation, reprinting, reuse of illustrations, recitation, broadcast-
ing, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this
publication or parts thereof is permitted only under the provisions of the German Copyright Law of
September 9, 1965, in its current version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable for prosecution under the German Copyright Law.

© Springer-Verlag Berlin Heidelberg 1996

Softcover reprint of the hardcover 1st edition 1996

Typesetting: Camera-ready by editors

Printed on acid-free paper
SPIN: 10486088 45/3142 - 5 4 3 210
Preface

Batch Chemical Processing, that ancient and resilient mode of chemical manufacture, has in the
past decade enjoyed a return to respectability as a valuable, effective, and, indeed, in many
instances, preferred mode of process operation. Batch processing has been employed in the past
in many sectors of chemical processing industries including food, beverage, pharmaceuticals,
agricultural chemicals, paints, flavors, polymers, and specialty chemicals. The batch mode is
increasingly being rediscovered by sectors that neglected it as the industry is focusing on more
specialized, application tailored, small volume but higher margin products. Moreover, as
information and control technologies have become both more technically accessible and
economically affordable, the operation of batch facilities has become more efficient, gradually
shifting from the conservative and simple operating strategies based on dedicated and cyclically
operating trains to more sophisticated and complex operating strategies involving flexibly
configured production lines using multi-functional equipment and employingjust-in-time inventory
management strategies.
The effects of these trends on the process systems engineering community has been a
renewed intensity of efforts in research and development on computational approaches to
modeling, design, scheduling, and control problems which arise in batch processing. The goal of
the NATO Advanced Study Institute (ASI), held from May 29 to June 7, 1992, in Antalya,
Turkey, was to review state-of-the-art developments in the field of batch chemical process systems
engineering and provide a forum for discussion of the future technical challenges which must be
met. Included in this discussion was a review of the current state of the enabling computing
technologies and a prognosis of how these developments would impact future progress in the
batch domain.
The Institute was organized into two interrelated sections. The first part dealt with the
presentations on the state of the batch processing in the Chemical Process Industries (CPI),
discussion of approaches to design and operation of more complex individual unit operations,
followed by the reviews of the enabling sciences. This four-day program served to set the stage
for a five-day program of discussions on the central problem areas of batch processing systems
 VI

engineering. That discussion was preceded by a one-day interlude devoted to software

demonstrations, poster sessions, and small group meetings. A unique feature of this ASI was the
presence of a computer room at the hotel site equipped with an llM RISC workstation, terminals,
and personal computers which could be used for application software demonstrations and trials.
The Institute opened with industrial and academic perspectives on the role of batch
processing systems engineering in the CPI. Two presentations on the status of batch processing
systems engineering in Japan and Hungary provided perspectives on developments in the Far East
and the former eastern block countries. The Japanese innovations in batch plant organization using
moveable vessels offered insights into materials handling arrangements particularly suitable for
multiproduct, smaIl-batch production environments. These presentations were followed by a suite
of papers describing applications in CPI sectors such as polymer processing, food and beverages,
biochemical, specialty chemicals, textile, and leather industries.
The more complex batch unit operations which give rise to special modeling, design, and
control problems were given attention in separate lectures. These included batch distiIlation,
reactors with complex reacting systems, and sorptive separation systems. These presentations
were complemented by expositions on the estimation and unit control issues for these more
complex systems.
The three categories of enabling technologies which were reviewed were simulation,
mathematical programming, and knowledge based systems. The simulation component included
discussion of solution techniques for differential algebraic systems, the elements of
discrete/continuous simulation, and available simulation environments, as well as prospects
offered by advanced computer architectures. The mathematical programming review included a
critical assessment of progress in nonlinear optimization and mixed integer programming domains.
The knowledge based systems program consisted of a review of the field, continued with its
elements and closed with more advanced topics such as machine learning including neural
networks.
During the fifth day, attendees divided into small discussion groups on specific topics,
participated in the software demonstrations and workshops, and participated in the poster sessions.
The software demonstrations included the DICOPT MlNLP solver from Carnegie Mellon
University, the BATCHES simulation system from Batch Process Technologies, and the BATCH-
KIT system (a knowledge based support systems for batch operations scheduling) developed at
ETHZurich.
 VII

The central problem areas in batch process systems engineering are those of plant and
process design and plant operations. One day was devoted to the former topic, focusing especially
on retrofit design as well as approaches to incorporating uncertainty in the design of processing
systems. The second day was devoted to scheduling and planning, including consideration of the
integration issues associated with linking the control, scheduling, and planning levels of
operational hierarchy. The Institute concluded with plenary lectures on future of batch processing
systems engineering and an open forum on questions which arose or were stimulated during the
course of the meeting.
The ASI clearly could not have convened without the financial resources provided by the
Scientific and Environmental Affairs Division of NATO. The support, advice, and understanding
provided by NATO, especially through the Division Director Dr. L. V. da Cunha, is gratefully
acknowledged. The additional financial support for specific attendees provided by the NATO
offices of Portugal and Turkey and by the US National Science Foundation is highly appreciated.
The enthusiastic and representative participation of the batch processing systems
engineering community was important for the realization of the goals of the ASI. Fortunately, such
participation was realized. Indeed, since the participation represented all the main research groups
in this domain, at one point the meeting concerns were voiced about the dire fate of the field if
some calamity were to visit the conference site. Fortunately, these concerns were abated the next
morning when the participants were greeted by maneuvers of NATO naval forces in Antalya bay.
Without question, the active participation of the distinguished lecturers, session chairs, reviewers,
and participants made this Advanced Study Institute a great success. Thanks are due to all!
Most of the manuscripts were updated considerably beyond the versions made available
to attendees during the Institute and we thank the authors for their diligent work. We sincerely
appreciate Springer-Verlag's understanding with unforeseeable delays with the manuscript as well
as their kind assistance throughout this endeavor. Special thanks are due to Dr. Hans Wossner and
1. Andrew Ross.
Finally, the organizers would like to recognize the help of the following individuals and
organizations without whom the Institute would have considerably diminished if not ineffective:
Sermin Gonen~ (now Sunol), Muzaffer Kapanoglu, Praveen Mogili, <;:agatay Ozdemir, Alicia
Balsera, Shauna Schullo, Nihat Giirmen, C. Chang, and Burak Ozyurt for assistance with
brochures, program, re-typing, indexing, and correspondence; Dean M. Kovac and Chairman R.
Gilbert of University of South Florida for supplementary financial support; Bogazi~i Turizm Inc.
 VIII

and Tamer Tours for local arrangements in Turkey and social programs; IBM Turkey, especially
Mtinire Ankol, for the RISC Station and personal computers; Canan Tamerler and Vildan Din~bR§
(ASI's Angels) for tireless help accompanied by perpetual smiles throughout the AS I; and Falez
Hotel management and staff, especially Filiz Giiney, for making our stay a very pleasant one.
The idea of organizing a NATO ASI on systems engineering goes back to 1988 and was
partially motivated by AKS's desire to do something in this domain at home for Turkey. However,
its realization was accompanied by personal losses and impeded by unanticipated world events.
A week before the proposal was due AKS lost his best friend, mentor, and mother, Mefharet
Sunol. The Institute had to be postponed due to the uncertainties arising from the Gulf crisis. A
few months before finalization ofthis volume, our dear friend and esteemed colleague, Prof David
W. T. Rippin passed away. It is fitting that this proceedings volume be dedicated to the memories
of Mefharet Sunol and David Rippin.

Gintaras V. Reklaitis and Aydm K. Sunol

West Lafayette, Indiana and Tampa, Florida
September 1996
List of Contributors and Their Affiliation

Organizing Committee and Director

Dner Horta~su, Chemical Engineering Department, Bogazi~i Universitesi, Istanbul, Turkey

Gintaras. V. Reklaitis, School of Chemical Engineering, Purdue University, USA
David W.T. Rippin, Technical Chemistry Lab, ETH Zurich, Switzerland

Director: Aydm K. Sunol, Chemical Engineering Department, University of South Florida, USA

Main Lecturers and Their Current Affiliation

Michel Lucet, Rhone Poulenc, France

Sandro Macchietto, Imperial College, UK
Rodger Sargent, Imperial College, UK
A1irio Rodriguez, University of Porto, Portugal
John F. MacGregor, McMaster University, Canada
Christos Georgakis, Lehigh University, USA
Arthur W. Westerberg, Carnegie Mellon University, USA
Ignacio E. Grossmann, Carnegie Mellon University, USA
Giresh S. Joglekar, Batch Process Technologies, USA
Jack W. Ponton, University of Edinburgh, UK
Kristian M. Lien, Norwegian Institute of Technology, Norway
Luis Puigjaner, Catalunya University, Spain

Special Lecturers

Mukul Agarwal, ETH Zurich, Switzerland

RIdvan Berber, Ankara Universitesi, Turkey
Cristine Bernot, University of Massachusetts, USA
 x

Ali Cmar, lIT, USA

Shinji Hasebe, Kyoto University, Japan
Laszlo Halasz, ETH, Switzerland
Gyula Kortvelyessy, SZEVIKI R&D Institute, Hungary
Joe Pekny, Purdue University, USA
Dag E. Ravemark, ETH, Switzerland
Nilay Shah, Imperial College, University of London, UK
Eva Sorensen, University ofTrondheim, Norway
Venkat Venkatasubramanian, Purdue University, USA
Zentner M. G., Purdue University, USA
Denis L.J. Mignon, Universite Catholique de Louvain, Belgium

Session Chairs (In Addition to Organizers and Lecturers)

Yaman Arkun, Georgia Tech, USA

Tiirker Giirkan, METU, Turkey
lisen Onsan, Bogazi~i Universitesi, Turkey
Canan Ozden, METU, Turkey
L. H. Garcia-Rubio, University of South Florida, USA

Additional Poster Contributors

Bela Csukas, Vezsprem University, Hungary

Bilgin Klsakiirek, METU, Turkey
Table of Contents

Plenary Papers

Current Status and Challenges of Batch Process Systems Engineering

David W T. Rippin

Future Directions for Research and Development in Batch Process

Systems Engineering ........................................................................................................ 20
Gintaras V Reklaitis

Status of Batch Processing Systems Engineering in the World

Role of Batch Processing in the Chemical Process Industry ........................................... 43

Michel Lucet, Andre Charamel, Alain Chapuis, Gilbert Guido, and Jean Loreau

Present Status of Batch Process Systems Engineering in Japan 49

Shinji Hasebe and 10ri Hashimoto

Batch Processing Systems Engineering in Hungary 78

Gyula K6rlvelyessy

Design of Batch Processes

Design of Batch Processes ............................................................................................... 86

L. Puigjaner, A. Espuiia, G. Santos, and M Graells

Predesigning a Multiproduct Batch Plant by Mathematical Programming 114

Dag E. Ravemark and D. W T. Rippin

The Influence of Resource Constraints on the Retrofit Design of Multipurpose

Batch Chemical Plants ..................................................................................................... 150
Savoula Papageorgaki, Athanasios G. Tsirukis, and Gintaras V Reklaitis

Design of Operation Policies for Batch Distillation 174

Sandro Macchietto and LM Mujtaba

Sorption Processes ...................... ... ............... ............. .... ... ......... ........ .......... ...... ... ... ..... .... 216
Alirio E. Rodrigues and Zuping Lu

Control of Batch Processes

Monitoring Batch Processes ............................................................................................. 242

John F MacGregor and Paul Nomikos
 XII

Tendency Models for Estimation, Optimization and Control of Batch Processes ........... 259
Christos Georgakis

Control Strategies for a Combined Batch Reactor / Batch Distillation Process .............. 274
Eva Sr/H"ensen and Sigord Skogestad

A Perspective on Estimation and Prediction for Batch Reactors ..................................... 295

Mukul Agarwal

A Comparative Study of Neural Networks and Nonlinear Time Series

Techniques for Dynamic Modeling of Chemical Processes ............................................ 309
A. Raich, X Wu, H F. Lin, and Ali (:mar

Enabling Sciences: Simulation Techniques

Systems of Differential-Algebraic Equations 331

R. W H Sargent

Features of Discrete Event Simulation ............................................................................ . 361

Steven M Clark and Girish S. Joglekar

Simulation Software for Batch Process Engineering ....................................................... 376

Steven M Clark and Girish S. Joglekar

The Role of Parallel and Distributed Computing Methods in

Process Systems Engineering ........................................................................................... 393
Joseph F. Pekny

Enabling Sciences: Mathematical Programming

Optimization .. .................... .... ........ ........... ... ....... ..... ............. ..... ...... ..... ... ............ ....... ...... 417
Arthur W Westerberg

Mixed-Integer Optimization Techniques for the Design and Scheduling of

Batch Processes ................................................................................................................ 451
Ignacio E. Grossmann, Ignacio Quesada, Ramesh Raman, and Vasilios T Voudouris

Recent Developments in the Evaluation and Optimization of Flexible

Chemical Processes .......................................................................................................... 495
Ignacio E. Grossmann and David A. Straub

Enabling Sciences: Knowledge Based Systems

Artificial Intelligence Techniques in Batch Process Systems Engineering ..................... 517

Jack W Ponton

Elements of Knowledge Based Systems - Representation and Inference ....................... 530

Kristian M Lien

Selected Topics in Artificial Intelligence for Planning and Scheduling Problems,

Knowledge Acquisition, and Machine Learning .............................................................. 595
Aydm K. Sunol, MuzaJfer Kapanoglu, and Praveen Mogili
 XIII

Integrating Unsupervised and Supervised Learning in Neural Networks for

Fault Diagnosis ................................................................................................................. 631
Venkat Venkatasubramanian and Surya N Kavuri

Scheduling and Planning of Batch Processes

Overview of Scheduling and Planning of Batch Process Operations ....... ...... ....... ...... .... 660
Gintaras V Reklaitis

GanttKit - An Interactive Scheduling Tool ..................................................................... 706

L. Halasz, M Hofmeister, and David W T Rippin

An Integrated System for Batch Processing ..................................................................... 750

S. Macchietto, C. A. Crooks, and K. Kuriyan

An Interval-Based Mathematical Model for the Scheduling of

Resource-Constrained Batch Chemical Processes .... ............... ... ........ ............ ....... .......... 779
M G. Zentner and Gintaras V Reklaitis

Applications of Batch Processing in Various Chemical Processing Industries

Batch Processing in Textile and Leather Industry ............................................................ 808

L. Puigjaner, A. Espufza, G. Santos, and M Graells

Baker's Yeast Plant Scheduling for Wastewater Equalization ........................................ 821

Neyyire (Renda) Tilmsen, S. Giray Velioglu, and Oner Hortar;su

Simple Model Predictive Control Studies on a Batch Polymerization Reactor ............... 838
Ali Karaduman and Ridvan Berber

Retrofit Design and Energy Integration of Brewery Operations .. ............. ....... ................ 851
Denis J Mignon

List of Participants 863

Index ................................................................................................................................ 867

Current Status and Challenges of Batch Processing
Systems Engineering

David W.T. Rippin

Labor fur Technische Chernie, E. T.H. Zurich, Switzerland

Abstract: The field offine and speciality chemicals exhibits an enonnous variety in the nature of
the products and the character and size of their markets, in the number and type of process
operations needed for production, in the scale and nature of equipment items and in the
organizational and planning procedures used.
This introductory paper draws attention to the need for a careful analysis of a batch situation
to identify the dominant features. Some methods and tools are identified to aid design, planning
and operations and some challenges or growing points for future work are suggested. These
themes will be taken up in more detail in later papers.

Keywords: Recipe, batch size, cycle time, design, multi product plant, multi plant, multipurpose
plant, scheduling, equipment capacity

Factors Characterizing the Batch Production of Chemicals

Any system for producing chemical products has three necessary components (Figure 1)
1. A market
2. A sequence of process tasks whereby raw materials are converted into products
3. A set of equipment items in which the process tasks are carried out
For the production ofa single product in a continuous plant, the links between these components
are :finnIy established at the design stage. The process task sequence is designed to serve a specific
market capacity for the product and the equipment is selected or specially designed to perfonn
precisely those necessary tasks most effectively.
 2

Process

Plant _ 0 - - - -__ Market

Figure 1: Components of a processing system

In a batch production system, the components themselves are much less rigidly defined and
the links between them are subject to change or fuzzy. For example, the market will not be for a
defined amount of a single product but may be for a range of products of which the relative
demands are likely to vary and to which new products are likely to be added. A variety of
processes may have to be considered which cannot be characterized with the precision demanded
for a continuous single product plant. Available equipment must be utilized or new equipment
selected from a standard range to serve multiple functions rather than specially designed. Similarly,
the allocation of process tasks to equipment items must be changed frequently to match the
changing requirements of the market.
A much wider range of choices may be available to the operators of batch plants than in a
continuous plant environment.
Furthermore, there is an enormous diversity of batch processing situations determined by
factors such as:
1. Scale of production
2. Variability of market demand
3. Frequency of "birth" of new products and "death" of old ones
4. Reproducibility of process operations
5. Equipment breakdowns
6. Value of materials in relation to cost of equipment
7. Availability and skill of process workers
8. Skill and experience of planning personnel
9. "Company culture"
Thus, the dominating factors may be quite different in one batch environment compared with
 3

another. No single sequential procedure can be called upon to solve all problems. A variety of
tools and practices will be needed which can be matched as necesslll)' to each situation.
The diversity of batch situations makes it important that, before starting on a particular plan
of action, an analysis of the situation should be made to assess problems and opportunities, to
identify the potential for improvement and to determine where effort and resources can most
effectively be invested.
In a busy production environment, the natural tendency, particularly if the overall situation is
not fully appreciated, is to treat the most pressing problems first, perhaps missing much more
profitable, but less obvious opportunities. Obviously, a right balance should be sought between
solving pressing problems, anticipating new problems, and grasping new opportunities. An overall
view is needed to do this effectively.
For example, a large chemical company I visited had set itself the task in a 5-year program for
its batch processes to reduce labor, cycle time and lost or waste material all by 35%. This certainly
provided a focus for activity, although it may not, of course, be the best focus in another
environment.

Analyzing Batch Processes

Some directions in which action might be taken to improve batch processing can be illustrated by
considering a series of questions:
1. What has to be done? - a breakdown of necesslll)' activities
2. How is achievement measured?
3. How to assess potential for improvement?
4. How to make improvement?
5. What facilities/tools are needed or advantageous?
6. Where to begin?

What Has to Be Done?

l. Batch recipe definition

2. Coordination of process tasks
 4

3. Design of new equipment or realization in existing equipment

4. Allocation of process tasks to equipment items
5. Sequencing of production
6. Timing of production
7. Realization in production with appropriate supervision and control measures
8. Matching the customer's requirements

How Is Achievement Measured?

Measures of performance can be associated with each of the activities that have been defined.
They may indicate the need for change in the corresponding activity, or in other activities earlier
in the list.
The batch recipe definition may be evaluated in terms of the raw materials and reaction
sequence, waste products, requirements for utilities and other resources, proportion of failed
batches and more quantitatively in the yield or productivity of reaction stages, the achieved
recovery of separations.
Coordination of process tasks at the development stage may be judged by how diverse are the
cycle time requirements of the different process tasks (although equipment and organizational
measures may be able to counter this diversity, if necessary). Overall optimization of a chain of
process tasks may give different results than attention only to isolated units.
Choice of new equipment items may be judged by investment cost. For existing equipment,
charges may be made for utilization or additional equipment may be called for. Choice of
equipment may also be judged by frequency of breakdown or availability.
Process definition and choice of new equipment will also be judged by whether the plant
performs to design specifications and achieves its design capacity.
Allocation of process tasks to equipment items may be judged by the size and frequency of
batches which can be produced of a single product or in some cases of a number of different
products.
The sequence in which products are produced may influence the total time and cost required
to produce a group of products, if the change over time and cost are sequence dependent.
The timing of production combined with the sequencing detennines if the delivery
 5

requirements and due dates of the customers can be satisfied.

Due to the intrinsic variability of the manufacturing situation, a variety of supervision and
control procedures may be required to counter both recognized and unrecognized changes. They
will be applied and their effectiveness assessed at different levels:
1. Process control on individual process tasks/equipment items
2. Sequence control on individual items and between interconnecting items
3. Fault detection and methods of statistic process control to detect changes in the process
or its environment
4. Quality control to monitor the quality of products delivered to the customer or perhaps
also raw material deliveries or environmental discharges
The success of the plant as a whole will be assessed by the overall production costs and the
extent to which the market requirements can be met. Customer satisfaction will reflect delivery
times and the quantity and quality of products delivered. Production costs will be ascribed to
different components or activities of the system and may indicate dominant factors.

How to Assess Potential for Improvement?

First indications of where action can profitably be taken can often be obtained by making extreme
or limiting assumptions about the changes which might be effected.
For example:
• At the process definition stage, what benefit would accrue if complete conversion/selectivity
could be achieved in the reactor or complete recovery in the separator? - or some less extreme
but conceivable assumption
• What consequence would there be for equipment requirement or utilization if the
interdependencies between the processing tasks were neglected and each equipment item sized
or assigned assuming 100% occupation for those tasks it had to perform?
• What benefits would there be in terms of operating costs, inventory costs, customer
satisfaction, iffailed batches or equipment breakdowns could be eliminated?
• What benefits in terms oftirneliness of production, customer satisfaction, inventory reduction
or production capacity by improvement of scheduling?
• What benefits if variability could be reduced or eliminated in terms of the costs of supervision
and inspection, product quality give away?
 6

The discipline of assessing how achievement is measured and what potential for improvement
may be identified should draw attention to those features ofthe system where further effort may
best be invested. If properly applied, such an analysis should be equally valid for exploring
measures which might be taken to solve a current problem, to preempt a potential problem or to
exploit a new opportunity.

How to Make Improvement?

Attention should be directed, initially, to the most promising (or threatening!) aspects of the
situation. These may vary widely from case to case, so no general recipe can be given. An obvious
defect or lack of reliability in the way a particular operation is carried out may claim immediate
attention and resources. However, sight should not be lost of the fact that equal or greater benefits
may derive from some higher level coordination or planning activity of which the effects may not
be immediately obvious to the plant operator or manager.
To appreciate the potential benefits of many of the possible actions, a quantitative
characterization is required of the processing requirements of a product, how these can be
matched to available or potentially available equipment, and how the requirements of different
products can be accommodated.
At the level of individual process tasks, there may be potential for improvement, for example,
with respect to reactor conversion, selectivity, separator recovery, waste production, time
requirements, in the light of the current objectives. This may be realized by developing a
mathematical model to be used for determining the optimal levels for operating variables. In some
circumstances there may be further advantage in making optimal choice not only of the levels of
operating variables, but also of the time profile of these variables during the course of the batch.
Reviews are available of where such benefits may be expected.

Quantitative Characterization of Batch Processing Requirements

Quantitative characterization of the processing requirements at each step of the process allows the
 7

determination of the perfonnance which can be achieved (amount of product per unit time) when
any step is allocated to an equipment item of appropriate type and a specified size. A process step
is any physical or chemical transfonnation which could be carried out in a separate equipment
item. For example heating, dissolving a solid, reacting, cooling, crystallization, are all regarded as
separate steps in the process specification, even though, at a later stage, several steps may be
carried out in sequence in the same equipment item.
The processing requirement of a step can be represented in different levels of detail by models
of widely differing complexity. The minimal specification of capacity is the size factor Sij that is
the volume (or other measure of capacity) required in a step j for each unit of product i to be
produced at the end of the batch and the batch time Tij required to complete the step.
Since the size factors of the processing steps are expressed relative to the amount of product
produced at the end of the batch, their calculation requires a complete balance of material and
corresponding capacity requirements encompassing all steps of the process.
If the size factor and cycle time are regarded as fixed, then the selection and allocation of
equipment and the determination of the overall production capacity of single and multi product
plants are relatively straightforward. Such a specification corresponds to a fixed chemical recipe

Table 1: Suggested Hierarchy of models

Item Model Type Function Derivation Use

A Comprehensive Perfonnance as function of Mechanistic Effect of
model time and all operating understanding individual unit on
variables whole plant
B Perfonnance Perfonnance as function of Reduced model Coordination of
model time or empirical fit cycle time of
sequence
C Model of time Fixed perfonnance, Reduced model Equipment sizing
and capacity requirement may depend or external and task
requirement on equipment or batch size specification assignment
D Model of time Reduced model Simple sequencing
requirement or independent and production
only specification
E Stochastic Superposition
model on any of
above
 8

from which all change in operating conditions is excluded. It may be appropriate when a fixed
manufacturing specification is taken over from conditions established, for example, in a
development laboratory or in previous production experience.
The allowance for variation in operating conditions calls for more complex models to predict
the effects ofthese variations on the processing requirements and probably iterative recalculation
of equipment allocation and overall production capacity determination. A hierarchy of models can
be envisaged for different purposes with information being fed from one stage ofthe hierarchy to
another as required (Table 1). However, in many practical applications for the planning and
operation of batch facilities the assumption of constant size factor and batch time will be
acceptable.
The simple quantitative characterization of batch processing requirements is used to determine
production capability when a defined process is allocated to a set of equipment items of given size.
No account is taken here of uncertainty or variability in processing requirements.

Production Capability of a Set of Equipment Items

If a process for product i is allocated to a set of equipment items of specified capacity Vj the
production capability is determined by:

The Limiting Batch Size BJi

An upper limit for the amount of product I which can be manufactured in a batch is imposed by

the equipment at the stage j with the smallest value of the ratio

equipment capacity =~
size factor (capacity requirement per unit product) Sij

B = Min "V;
Li . S
] ij
 9

The Limiting Cycle Time Tu

A lower limit to the interval between producing successive batches of product i is imposed by the
process stage with the largest batch time

The maximum production rate of product i per unit time is then

limiting batch size BLi

limiting cycle time TLi

~ ~ ~
Rl R2
R3

I T I

TAl

TAJ

Figure 2: Limiting batch size and cycle time

 10

The effect of the batch size and cycle time limitations are best illustrated graphically by a simple
example with three equipment items (Figure 2). The cycle time limitation occurs on the first item
and the batch size limitation on the second. Ifit is desired to increase the production rate of the
product, this can be done by addressing either of the two bottlenecks in capacity or cycle time.
The following measure:; can be taken to increase production rate, or to make underutilized
equipment available for other-purposes:
• Add parallel equipment items
which operate in phase to increase batch size
which operate out of phase to reduce cycle time
• Merge neighboring tasks to carry them out in the same equipment with potential saving of
equipment items or split a task, allowing it to be carried out in two consecutive items thus
reducing the cycle time
• Insert storage provision between two consecutive tasks allowing different batch sizes and
cycle times up and downstream ofthe storage.

Design of Single Product Plant

If a batch plant is to be designed to meet a specified demand of a single product I with a known
process, the number of batches which can be manufactured is given by

available time
limiting cycle time

The required batch size is then

B = total product demand

I number of batches

The necessary size of each equipment item is

~ = (batch size) x (size factor) =B j Sij

The design ofa single product plaut can be carried out immediately, if the configuration of the
plant is specified.
 11

However, if the designer is free to choose different structural alternatives of the type discussed
in the previous section, consideration may be given to
• The installation and function of parallel equipment items
• Changing the allocation of process tasks to equipment items
• The installation of intermediate storage
An acceptable design may be arrived at by examination of alternative cases or by optimization
over the range of discrete choices available, for example to minimize total equipment cost as is
discussed in later papers.

Multiproduct Operation

In multiproduct operation, several products are to be produced in sequence on the same plant with
a defined allocation of process tasks to equipment items for each product. For each product, the
limiting cycle time can be calculated and hence the production rate. It is then easy to check if
specified demands for a given set of products can be met within the available production time.
Bottlenecks to increased production for a particular product can be relaxed in ways already
discussed. However, in a multiproduct plant, the situation is likely to be complicated by the fact
that bottlenecks for different products will be differently located.
This is illustrated in Figure 3. The first product, as considered previously, is limited in batch
size by the equipment item on the second stage and its limiting cycle time is on the first stage. The
second product has its capacity limitation on the first stage and the cycle time limitation on the
second.
It is not immediately clear what is the most cost effective way of increasing the capacity of
such a plant. For example, increasing the size of the second equipment item would increase the
batch size of the first product enabling more to be manufactured in the same number of batches,
or the same amount to be manufactured in fewer batches leaving more time available for the
production of the second product. There are alternative ways, already discussed, for either product
by which its production rate can be increased.
The best design, for example, to satisfY specified demands for a set of products at minimum
equipment cost can be determined by solving an optimization problem. The optimization can be
 12

J ~ ~
RI Rz
R3

l I I

t TAl ~

TAl

TAl
'--
---
_ TA _

R,-----===,...._---1===r_
R3 ____________________ ~~~ __________ ==~

Figure 3: Multi-product production

formulated as choosing the best sizes (and possibly also the configuration) of the equipment items
to satisfy the product demands within the available production time, or alternatively it can be
viewed as choosing how the available time is divided between production of the different products.
With a fixed equipment configuration, when the time aIlocation to a product has been fixed the
equipment sizes to accommodate that product are also determined. For a particular time allocation
to the complete set of products the necessary size of any equipment item is determined by
scanning over the equipment size requirement at that stage for all products and choosing the
largest.
In a practical situation where exact cost optimization is not of major importance, a plausible
design can be arrived at by allocating the available time to products or groups of products
 13

detennining the equipment cost and checking its sensitivity to time allocation or equipment
configuration.
Bottlenecks may be removed by changing configuration as for the single product.

Discrete Equipment Sizes

Much batch equipment is not available in continuously variable sizes. For example volumes of
available reaction vessels may increase by a factor of -1.6. A discrete optimization procedure,
such as a branch and bound search, may be used to make an optimal selection of equipment items
in these circumstances, if required.

Partly Parallel Production (See Figure 4)

Products being produced on the same plant may differ substantially in the number of equipment
items needed. In a plant equipped to produce a complex product with many transformations it may
be possible at another time to produce several less demanding products in parallel. Capacity
evaluation or design can be carried out in ways similar to those already discussed.

Alternative Plant Configurations

1. Multiproduct

2. Partly parallel

3. Multiplant

4. Multiplant and
partly parallel

Figure 4: Partly parallel production

 14

Multiplant Production

When a number of products have similar equipment requirements, it may appear attractive to
manufacture as many of them as possible in a single large multiproduct plant. The product will be
made in large batches and, by economy of scale, the equipment cost will be less than the total cost
of a number of smaller plants operating in parallel to produce subgroups of products or even
individual products.
However, other factors may militate against the very large multiproduct plant. Manufacturing
many products in the same plant will call for frequent product changeovers with associated loss
of time and material and other expenses. In addition, if a product is in continuous demand, but is
only manufactured for short periods, as will be the case if it shares the plant with many other
products, then the inventory level to maintain customer supplies will be much higher than for a
plant in which only a few products are manufactured. For high value products the cost of such
inventories may be far more significant than "economy of scale" savings on equipment costs. If
in addition, purity requirements are high leading to very stringent changeover procedures, the use
of parallel plants for single, or a few products, may easily be justified on economic grounds.
Discrete optimization procedures can be set up to assist the grouping of products into subsets for
production together.

Multipurpose Operation

Many batch products are made in multipurpose plants. A set of equipment items is available in
which a group of products is manufactured. The products may change from time to time calling
for the plant to be reconfigured. Several products may be manufactured in the plant at one time
and the same product may follow different routes through the plant at different times, perhaps
depending upon which other products are being produced at the same time.
One way of assessing the capacity of such a plant is to. plan its activity in campaigns. A
campaign is the assignment of all the equipment in the plant to the production of a subgroup of
products for a period. The total demand for a set of products over a planning period can be
 15

satisfied by selecting and assigning times to an appropriate set of campaigns chosen from a larger
set of possible campaigns. This selection can be made by a linear programming procedure.
The candidate campaigns can be built up by assigning equipment to the manufacture of
batches of the same or different products in parallel (Figure 5). Candidates which do not make
good use ofthe equipment can be screened out, leaving a modest number of efficient campaigns,
from which the linear programming procedure can make an appropriate selection corresponding
to any distribution of demand.
The method described seeks to make the best use of the capacity of an existing multipurpose
plant. It could theoretically be extended to consider the selection of equipment items to include
in the plant, but at the cost of enormously more computing. It is also not at all clear how to specify
the potential product demand for such a plant at the design stage. In fact, many multipurpose
plants are built with limited knowledge of the products to be made in them, or even with the
intention that as yet unknown products will be produced there.
In those circumstances it is not surprising that the selection of items to include in a

multipurpose plant is commonly based on past experience rather than any mathematical
procedure. The choice of items may be based on the requirements of one, or a group of
representative products, possibly with special provision for other unusual products. Alternatively,
a typical mixture of equipment items may be chosen that has given good service in the past and
has sufficient flexibility to accommodate a range of types and amounts of products. Of course, the
selection of equipment items made in this way may be heavily dependent on the particular
application field being considered.

The Choice of Production Configuration and Short Term Scheduling

The choice of a production configuration depends on the character of the processes and the
market situation. A group of products with very similar equipment requirements can often be
produced conveniently in a multiproduct configuration. Very diverse products may be produced
in independent production lines or, depending on the availability of equipment and the level and
variability of demand in mixed multipurpose campaigns.
The structuring of campaigns described in the previous section is a device for identifying
favorable combinations of products. It may be useful for medium term and capacity planning, but
 16

Plant Specifications Production Process

Multi Purpose Batch Plant Product A

Production Requirements

00 llal [.
~ < planning period >

t t t
c;J @ loo
(kaJ
450 300 A 5000
B4000
Clooo
B3000
C7000
D2000 [kg]
E6000

.~ (kg)
500 ~ ~ lOO

Economic Data
- Sales prices
- raw mat. costs
..
- storage costs

Subdivision of
Planning period
Camp II Camp 21 3 5

t
into production
campaigns

Module-Task Production line

Allocations (Product A)

G~~
Campaign I Production line
(Product B)

Gantt Chart
S I

Prod. line I l----1-t----,r----_-1..t_.-_____

t t
I --t

Figure 5: Multi-purpose planning

 17

it will not be rigidly adhered to in day to day production planning. There must be flexibility to
adapt, for example, to short-term changes in equipment availability, local working conditions or
utility capacity. There will probably be less freedom to assign process tasks to equipment items
than in the earlier construction of multipurpose campaigns. However, there must be the possibility
to make local adjustment to the timing and allocation of individual batches and to propagate the
consequences of these changes.
In practice, this is often done manually on a bar chart. A similar graphical facility can be made
available in a computer, or a more comprehensive computer program could not only propagate
the consequences of a change, but also examine alternative measures to adapt most profitably to
the change. Various aspects of scheduling will be reviewed later.

Accommodating the Variety of Batch Processing

If the manufacture of fine and speciality chemicals is to be undertaken, the following questions
should be considered:
1. Which products, in consideration of their process and market characteristics are suitable
for batch processing?
2. Which products might be produced together in shared equipment?
3. Which groups of products have sufficiently homogeneous processing requirements that
they might be produced consecutively in a single multi-production line?
4. Which products have diverse or fluctuating requirements suggesting that they might be
produced in a general mUlti-purpose facility?
5. On the basis of past experience and anticipated developments, what range of equipment
items should be installed in a multipurpose facility?
Whatever decisions are taken about product assignments to particular types of production and
choice and configuration of equipment, there will be continual need for monitoring performance,
for scheduling and rescheduling production, and for evaluation the effect of introducing new
products and deleting old ones.
Harnessing the inherent flexibility of batch processing to deal effectively with change and
uncertainty is a problem which is solved routinely in practice. However, the mathematical
representation of this problem and how it can be solved in a truly optimal way, are still the subject
of study, some of which is reported later.
 18

What Facilities I Tools Are Needed or Advantageous?

• The ability to assess relevant overall data and make rapid order of magnitude estimates of the
effect of constraining factors or potential benefits
Hence, to identify elements exercising the dominant constraints on improvements
• The ability to predict the effect of measures to relieve the constraints and hence expected
improvements resulting from suggested changes
• In some cases, optimization capability to extract the maximum benefit from certain defined
types of changes may be justified
• Packages to perform some of these tasks may be available: overall or detailed simulation,
design, scheduling, optimization
• Flexible capability to do a variety of simple calculations and easy access to the necessary
basic data may often be important

Some Challenges and Opportunities in Batch Processing

1. Because of the great diversity of batch processing, measures are needed to characterize
a batch processing situation to enable computer and other aids to be matched to requirements.
2. Quick estimation procedures to assess whether features of the system are sufficiently
significant to be considered in greater detail.
3. Integration of knowledge to make available as appropriate the totality of knowledge about
the system including hierarchical model representations.
4. Batch process synthesis including co-ordination of the performance of individual stages
with overall requirements of assignment and scheduling, perhaps also coupled with
multi-product considerations.
5. Non-linear and integer programming - efficient problem formulation, significance and
exploitation of problem structure.
6. Catalogue of potential benefits of profile optimization for different batch operations,
reaction kinetics and objective functions.
 19

7. Guidelines for potential usefulness of adaptive control, on-line estimation, optimization

as a function of the batch operation, the conditions'to which it is exposed and the wider plant
environment in which it is situated.
8. Single/multi-product plant design - potential benefits of recipe adjustment. Is the more
detailed examination of scheduling at the design stage beneficial?
9. Effect of wide ranging uncertainty on multi-product or multi-purpose design. When is it
sensible to drop explicit consideration of product demands and what should be done then?
10. Are there significant advantages in coordinating the dynamic simulation of individual
batch units over the whole process and coupling with sequence control?
11. What can be achieved in scheduling, and where will further progress be made, for example
with reference to problem size, interaction with the user, relative merits of algorithmic
(integer programming) versus heuristic, knowledge-based methods?
12. What is really new - pipeless plants? anything else?

Conclusion

Industrial interest in fine and speciality chemicals has increased substantially in recent years, not
least because these often seem to be the most profitable parts ofthe chemical industry. Over the
same period academic work has produced a number of models which have been refined in various
ways to express details of how batch processing could be carried out.
There is certainly scope for further interaction between industry and university to match
modeling and optimization capabilities to industrial requirements. One benefit of the present
institute could be further moves in this direction.

Addendum

For further details and a comprehensive reference list, the reader is directed to D. W. T. Rippin,
Batch Process Systems Engineering: A Retrospective and Prospective Review, ESCAPE-2,
Supplement to Comput. & Chern. Engineering, 17, Supplement, S I-S 13 (1993)
Future Directions for Research and Development in
Batch Process Systems Engineering

Gintaras V. Reklaitis

School of Chemical Engineering, Purdue University, West Lafayette, IN 47907-1283, USA

Abstract: The global business and manufacturing environment, to which the specialized and
consumer products segments of the CPI are subjected, inexorably drive batch processing into the
high tech forefront. In this paper the features ofthe multipurpose batch plant of the year 2000 are
reviewed and the implications on its design and operation summarized. Research directions for
batch process systems engineering are proposed, spanning design applications, operations
applications and tool developments. Required advances in computer aided design encompass task
network definition, preliminary design of multipurpose plants, retrofit design, and plant layout.
Needs in computer support for operations include integration of the application levels of the
operational hierarchy as well as specific developments in scheduling, monitoring and diagnosis,
and control. Advances in tools involve improved capabilities in developing and testing algorithms
for solving structured 0-1 decision problems and interpreting their results, further enhancements
in capabilities for handling large scale differential algebraic simulation models with implicit
discontinuities, and creation of flexible data models for batch operations.

Keywords: Algorithm adversary, computer integrated manufacturing, continuous/discrete

simulation, control, data mode~ heat integration, manufacturing environment, materials handling,
mixed integer optimization, monitoring and diagnosis, multiplant coordination, multipurpose plant,
plant layout, preliminary design, reactive scheduling, resource constrained scheduling, retrofit
design, task networks, uncertainty

Introduction
The lectures and presentations of this Advanced Study Institute have amply demonstrated the
vigor and breadth of contemporary systems engineering developments to support batch chemical
processing. Much has been accomplished, particularly in the last ten years, to better understand
 21

the design, operations, and control issues relevant to this sector of the chemical industry. New
computing technologies have been harnessed to solve very challenging and practical engineering
problems. Yet, given the explosive growth in hardware capabilities, software engineering and
tools, and numerical and symbolic computations, further exciting developments are within our
reach. In this presentation, we will attempt to sketch the directions for continuing developments
in this domain. Our proposals for future research are based on the conviction that the long term
goal for batch process systems engineering should be to fully realize computer integrated
manufacturing and computer aided engineering concepts in the batch processing industry in a form
which faithfully addresses the characteristics of this mode of manufacturing.

Trends in the Batch Processing Industry

Any projections of process systems engineering developments targeted for the batch processing
industry over the next five to ten years necessarily must be based on our anticipation of the
manufacturing challenges which that industry will need to face. Thus, the point of departure for
our discussion must lie in the assumptions that we make about the directions in which the batch
processing industry will evolve. Accordingly, we will first present those assumptions and then
launch into our discussion of necessary batch process systems developments.

Chemical Manufacturing Environments 2000

As noted by Loos [12] and Edgerly [6] the chemical industry in the year 2000 will evolve to four
basic types of manufacturing environments: the consumer products company, the specialized
company, the utility, and the megacompany.
The consumer products company, of which 3M, Procter & Gamble, and Unilever are present
day precursors, features captive chemical manufacturing which supports powerful consumer
franchises. The manufacture offine chemicals, polymers, and petrochemicals within this type of
company will be driven by consumer demands and as these peak and wane chemical processing
will have to change to accommodate. The market life of such products is often two years or less
[24]. Many food processing companies are tending in this direction.
The specialized companies, of which Nalco and Lubrizol are illustrative, will be midsized
organizations that possess unique technical capabilities and marketing/customer access. These
organizations will be involved in continuous technical innovation and intensive focus on customer
 22

service and thus their manufacturing functions will subject to continuous product tum-overs,
pressures for quick start-up and rapid response to market needs. Pharmaceutical companies are
evolving in this direction.
The utility, of which SABIC is a precursor, will be the low cost converter of chemical
feedstocks into basic building block chemicals for the other sectors of the CPI. This type of
organization will flourish by virtue of its leading-edge technology, world-class scale of production,
and advantaged access to raw material or energy sources. Manufacturing in such an organization
will be highly automated and optimized for operation with minimum upsets and quality deviations.
The fourth category, the megacompany will be the leader in diverse segments of the chemical
market, encompassing some or all of the above manufacturing models. These organizations will
operate on a global basis with great technical depth, and financial and marketing strength. DuPont,
Hoechst, and ICI could be precursors of such companies. Manufacturing in the megacompany
will be subject to the same factors as the above three types of companies depending upon the
sector in which that particular arm of the organization competes.
It is clear that the specialized and consumer products companies and analogous arms of the
megacompanies will be the CPI components in which batch processing will continue to grow and
flourish. These sectors of the processing industry will increasingly share in the same business
environment as experienced in discrete manufacturing: a high level of continuous change in
products and demands, close ties to the customer, whether consumer or other organization, strong
emphasis on maintaining quality and consistency, accelerating demands for worker, product, and
community safety and prudent environmental stewardship, and relentless competitive pressures
to be cost effective.

Consequences of the Changing Environment

The consequences of these factors are that batch processing, the most ancient mode of chemical
manufacturing, will be increasingly driven into the high technology forefront. The batch plant of
the year 2000 will be a multipurpose operation which uses modularized equipment and novel
materials handling methods and is designed using highly sophisticated facilities layout tools. It will
feature a high degree of automation and well integrated decision support systems and,
consequently, will require significantly lower levels of operating staff than is present practice. The
batch processing based firm will employ a high degree of integration of R&D, manufacturing and
business functions, with instantaneous links to customers, suppliers, as well as other cooperating
 23

plant sites on a global basis. It will employ computer aided process engineering tools to speed the
transition from the development of a product to its manufacture, without the protracted learning
curves often now encountered.

Design Implications: Short product life and intensely competitive markets will impose major
challenges on both the manufacturing and the product development processes. Responsiveness to
the customer needs for tailored formulations, generally, will lead to increasing specialization and
multiplication of products, resulting in increased focus on very flexible, small batch production.
The multipurpose chemical plant will become the workhorse of the industry. At the same time, in
order to reduce the operational complexity often associated with multipurpose plants,
standardization and modularization of the component units operations will be employed, even at
the cost of higher capital requirements and possibly lower capacity utilization levels. As in the
manufacture of discrete parts, streamlining of flexible, small batch production will require
increased focus on the materials handling aspects of batch operations. With small batch operation,
the traditional pipes, pumps, and compressors can, depending upon recipe details, loose their
effectiveness as material transfer agents. Other modes of transfer of fluids and powders such as
moveable bins, autonomous vehicles, and mobile tanks can become more efficient alternatives.
Factors such as efficient materials handling logistics, reduction of in-process inventories,
minimization of cross-contamination possibilities and increased operating staff safety will dictate
that consideration of physical plant layout details be fully integrated into the design process and
given much greater attention than it has in the past.
The industry will also need to give increased emphasis to reducing the product development
cycle. This means employing sophisticated computational chemistry tools to guide molecule design
and exploiting laboratory automation at the micro quantity level to accelerate the search for
optimum reaction paths, solvents and conditions. The use of large numbers of automated parallel
experimental lines equipped with robotic aids and knowledge based search tools will become quite
wide-spread. To insure that recipe decisions take into account available production facilities early
in product development, process chemists will need to be supported with process engineering tools
such as preliminary design and simulation software. Use of such tools early in the development
process will identifY the time, resource, and equipment capacity limiting steps in the recipe,
allowing process engineering effort to be- focused on steps of greatest manufacturing impact. The
models and simulations used in development will need to be transferred to production sites in a
 24

consistent and usable fonn to insure that processing knowledge gained during development is fully
retained and exploited.

Operational Implications: Since maintenance of tight product quality standards will be even more
of a necessity, sophisticated measurement and sensing devices will be required. The need to
control key product quality indices throughout the manufacturing process will put high demands
on the capabilities of regulatory and trajectory tracking control systems. Early prediction and
correction of recipe deviations will become important in order to reduce creation of off-spec
materials and eliminate capacity reducing reprocessing steps. Thus, integrated process monitoring,
diagnosis and control systems will be widely employed. The needs to largely eliminate operator
exposure to chemical agents and to contain and quickly respond to possible releases of such
materials will further drive increased use of automation and robotic devices. Indeed, all routine
processing and cleaning steps will be remotely controlled and executed. To allow the reduced
number of operating staff to effectively manage processes, intelligent infonnation processing and
decision support systems will need to be provided. Effective lateral communications means
between corporate staff world-wide will be employed to facilitate sharing of manufacturing
problem solving experience, leading to continued manufacturing quality improvements. Rapid and
predictable response to customer orders will require development of simple, reliable operating
strategies and well integrated scheduling tools. Manufacturing needs for just-in-time arrival of raw
materials, key intennediates, and packaging supplies will drive the development of large scale
planning tools that can encompass multiple production site and suppliers. The realization of
computer integrated process operations will require extensive and realistic operator and
management training using high fidelity plant training simulations. These models will further be
used in parallel with real plant operation to predict and provide benchmarks for manufacturing
perfonnance.
The inescapable conclusion resulting from this view of trends in batch chemical processing is
that the needs for infonnation management, automation, and decision support tools will accelerate
dramatically over the next decade. The marching orders for the process systems community are
thus to deliver the concepts and tools that will ~ncrease the cost-effectiveness, safety, and quality
of multipurpose batch operations. The greatest challenge will be to use these tools to discover
concepts and strategies that will lead to drastic simplifications in the design and operation of batch
facilities without loss in efficiency or quality. Simplifications and streamlining of potentially
 25

complex manufacturing practices will be the key to maximum payoffs in safety, reliability, and
competitiveness.

Research Directions for Batch Process Systems Engineering

In this section, we will outline specific areas in which productive process systems developments
should be made. Our projection of research directions will be divided into three areas: design
applications, operations applications, and tool development. This division is admittedly artificial
since in the batch processing domain design and operation are closely linked and successful
developments in both domains depend critically on effective tools for optimization, simulation, and
information processing and solution comprehension. However, the division is convenient for
discussion purposes.

Advances in Design

While considerable methodological progress has been made since the review of computer aided
batch process design given at FOCAPD-89 [20], a number ofissues remain unexplored. These can
be divided into four categories: task network definition, preliminary design methodology, retrofit
design approaches, and plant layout.

Task Network Definition: The key process synthesis decisions made early in the development of
a product center on the definition of the precise product recipe and the aggregation of subsets of
the contiguous steps ofthe recipe into tasks which are to be executed in specific equipment types.
These decisions define the task network which is the basis for selecting the number and size of the
process equipment. Recipe definition is usually made by process chemists in the course of
exploring alternative synthesis paths for creating the product or molecule of interest. The
decisions involved include the selection of the most effective reaction path which has direct impact
on solvents to be employed, reaction conditions, by-product formation, and the types of unit
operations which will be required. Task selection involves a range of qualitative and experiential
information which incorporates choices of the broad types of equipment which will be selected to
execute the tasks. The overall task network definition problem would be greatly facilitated if a
knowledge based framework could be developed for task network synthesis which incorporate
both the recipe definition and task selection components. To date the proprietary PROVAL
package [1] remains the only development which addresses some aspects of this synthesis
problem.
 26

Preliminary Design ofMultipurpose Plants: The recent work ofPapageorgaki [16] and Shah
and Pantelides [23] does address the deterministic, long campaign case, while Voudouris and
Grossmann [27] offer approaches to incorporating discrete equipment sizes. However, one of the
key issues in grass roots design of such facilities is the treatment of uncertainties. Shah and
Pantelides [22] do suggest an approach for treating multiple demand scenarios within a
deterministic formulation of the problem, an idea which had previously been advanced by Reinhart
and Rippin [19] in the multiproduct setting. Moreover, it would appear that the staged expansion
concept, initially explored by Wellons [28,29] for longer term demand changes, merits
consideration in the multipurpose case, especially in the context of modular plant expansion. Yet
missing is a framework for handling likely changes in the product slate, in other words
uncertainties in recipe structures, since one of the key reasons for the existence of multipurpose
plants is the adaptability of the plant in accommodating not only demand but also product changes.
The latter aspect of flexibility needs to be given a quantitative definition.
The increasing interest in alternative material handling modes raises questions of under what
recipe conditions these various alternatives are most cost effective. For instance, the vertical
stacker crane concept appears to be advantageous for the short campaign, reaction dominated
recipe, while the tracked vessel concept is said to be appropriate for mixinglblending type recipes.
Clearly, depending upon recipe structure and campaign length, different combinations of these
handling modes together with conventional pipe manifold systems might be most economical. The
incorporation of these material handling options within an overall process design framework
would appear to be highly desirable as a way of allowing quantitatively justified decisions to be
made at the preliminary design stage.
While mathematical programming based design formulations are adequate at the preliminary
design stage, detailed investigation of designs requires the use of simulation models. Simulation
models do allow consideration of the dynamics of key units, step level recipe details, complex
operating strategies, as well as stochastic parameter variations. Ideally such simulations should
also form the basis for detailed design optimizations. However, while batch process simulation
capability does exist (see [3]), the optimization of dynamic plant models with many state and time
event discontinuities continues to present a challenging computational problem. Although some
interesting developments in the optimization of differentiaVaigebraic systems involving applications
 27

such as batch distillation columns [2] have been reported, further investigation of strategies for
the optimization ofDAE systems with implicit discontinuities are clearly appropriate.

Retrofit Design: A MINLP formulation and decomposition based solution approach for the
retrofit design of multipurpose plants operating under the long campaign operating strategy was
reported in [15]. This formulation included consideration of changes in product slate and demands
as well as addition of new and deletion of old units with the objective of maximization of net
profit. An extension of this work to accommodate resource constraints was reported at this
conference [17]. Incorporation of the effects of campaign changeover and startup times is
straightforward in principle, although it does introduce additional 0-1 variables. Developments
which merit investigation include incorporation of continuous units and investigation of key state
variable trade-offs during retrofit. In principle, the latter would require inclusion of the functional
dependence of key recipe parameters on state and design variables such as temperature,
conversion, and recovery. Since these nonlinear dependencies may need to be extracted from
simulations of the associated processing tasks, a two level approach analogous to the SQP strategy
widely employed in steady state flowsheet optimization may be feasible.
Further key factors not treated in available retrofit design approaches include consideration
of space limitations for the addition of new equipment and changes in the materials handling
requirements which the addition of new equipment and elimination of old equipment impose.
These factors clearly can only be addressed within the framework of the plant layout problem
which will be discussed in a later section ofthis paper.
Heat integration is a feasible retrofit option for batch operations, especially under long
campaign operation, and has been investigated in the single product setting [25]. Recent work has
led to an MILP formulation which considers stream matches with finite heat exchange times and
batch timing modifications to minimize utilities consumption [11]. Interesting extensions which
should be pursued include scheduling of multifunctional heat exchange equipment so as to
minimize the number of required exchanger units as well as consideration of the integrated use of
multiple intermediate heat transfer fluids.

Plant Layout: Once the selection of the number and capacities of the plant equipment items has

been made, the next level of design decision involves the physical layout of the process equipment.
The physica1layout must take into account (1) the sizes/areas/volumes ofthe process equipment,
(2) the unit/task assignments, which together with the recipe fix the materials transfer links
 28

between process vessels, (3) the materials transfer mechanisms selected to execute these links, and
(4) the geometry of the process structure within which the layout is imbedded. Clearly, safety
considerations, maintenance access requirements, cross-contamination prohibitions, and vibrational
and structural loading limitations will further serve to limit the placement of process equipment.
While these aspects of plant layout have been handled traditionally via rules of thumb and the
evolving practices of individual engineering firms, the trend toward truly multipurpose facilities
which employ a variety of material handling options will require that the plant layout problem be
approached in a more quantitative fashion. This is particularly the case with layouts involving
enclosed multilevel process buildings which are increasingly being employed for reasons of
esthetics, safety, and containment of possible fugitive emissions.
As discussed in [7], the plant layout problem can be viewed as a two level decision problem
involving the partitioning of the equipment among a set oflevels and the location of the positions
of the equipment assigned to each level. The former subproblem can be treated as a constrained
set partitioning problem in which the objective is to minimize the cost of material transfers
between units and the constraints involve limitations on additive properties such as areas and
weights of the vessels assigned to each level. Because of the effects of gravity, different cost
structures must be associated with transfers in the upwards, downwards, and lateral directions.
As shown in [7], the problem can be posed as a large MILP and solved using exact or heuristic
partial enumeration schemes. The subproblem involving the determination of the actual positions
of the equipment assigned to a level is itself a complex decision problem for which only rather
primitive heuristic approaches have been reported [8]. The integrated formulation and solution of
these subproblems needs to be investigated as such a formulation could form the basis for
investigating alternative combinations of material handling strategies. mtimately, the layout
problem solution methodology should be linked to a computer graphics based 3D solids modeling
system which would allow display and editing of the resulting layout. Further linkage of the 3D
display to a plant simulation model would allow animation of the operation of the multipurpose
plant, especially of the material transfer steps. Such virtua1 models of plant operation could be very
effectively used for hazards and operability analysis, operator training, and design validation
studies.

Advances in Operations

The effective operation of the batch processing based enterprise of the year 2000 will require full
 29

exploitation of computer integrated manufacturing developments and the adaptation of these

developments for the specific features of batch operations. While descriptions of the information
flows in process oriented CIM systems have been formulated [30] and implementation standards
oriented to the chemical industry are being formalized [31], no concrete implementations have
actually been realized to date [18]. Given the state of the art, the required research must proceed
along two tracks: first, investigation of integrated process management frameworks spanning the
planning, scheduling, contro~ diagnosis and monitoring functions and, second, basic developments
in the component applications themselves. In this section we will briefly outline research thrusts
in both of these tracks.

Application Integration: From the perspective of a process scheduling enthusiast, the batch
processing CIM framework can be viewed as a multilevel integrated scheduling problem, as shown
in Figure 1. At the top-most level of the hierarchy is the coordination of the production targets and
logistics involving multiple plant sites. This level interfaces with the master schedules for individual
plant sites which treat resource assignment and sequencing decisions of a medium term duration.
The master schedule in turn must be continuously updated in response to changes on the plant
floor and in the business sector. The need for these changes is identified by the process monitoring
and diagnosis system which includes key input from the operating staff The actual changes in the
timing and linkage to the required resources are implemented through the control system. The
entire framework is of course linked to business information systems, including order entry and
inventory tracking systems.
One of the key impediment to assembling such a framework is the difference in level of
aggregation of the information and decisions employed at each level of the hierarchy. As noted
recently by Macchietto and coworkers, the master schedule operates at the level of tasks, the
reactive scheduling function deals with timing of steps and material transfers, while the control
system operates at the even more detailed level of valve operations, interlock checks, and
regulatory loops. An initial and instructive experiment at integrating three specific functional
blocks, namely, master scheduling, reactive scheduling, and sequential control translation blocks
has been reported [5] . The principal focus of that work was on reconciling the differences in the
procedural information models employed by each of these blocks. Further work is clearly needed
to examine the implications of linking the process monitoring and diagnosis functionalities into a
comprehensive manufacturing control system, as shown in Figure 2 (after [18]). As envisioned,
 30

Coordinate Multi-Site
Multi-Plant
Scheduling Production &. Logis~ics

Medium term
Plant Master Assignment, sequencing
Schedul ing
&. timing

t
Reactive
Response to changes
,--;
Scheduling on plant floor

t Identify deviations
Diagnosis
from master schedule

t
L( Control)
Implement timing
&. set-point changes

Figure 1. Batch processing elM levels

the control system generates process information which is filtered and tested to eliminate gross
errors. The intelligent monitoring system extracts important qualitative trends, processes these
trends to provide near term forecasts, and provides qualitative description of process behavior.
The fault detection system detects deviations from expected behavior and recommends corrective
actions. These recommendations are offered to the user through a high level interface and, if
validated, are presented to the supervisory control system which selects appropriate control
configurations, algorithms, and settings, including changes in timing of actions. Neural networks
and knowledge based, especially rule based, systems would appear to be the relevant technologies
for these functions.
 31

)
r

l Process

r--:
( Regulatory COjJtrol 1
System )
\.
,.I
( Data Processing
System
J
Intelligent Monrtorlng
System

Intelligent Supervisory Fault Diaanosis

Control System System
4

I Intelligent
User-lntertace

,•I

Figure 2. Integrated process monitoring, diagnosis and control

A more fundamental question which underlies the very structure of the above conceptual
integration framework is how to quantitatively and rigorously deal with the uncertainty which is
inherent to the manufacturing environment. Under present methodology, each component of the
hierarchical framework shown in Figure 1, employs a deterministic approach to dealing with
uncertainly at its level of aggregation. The multiplant coordinator looks over longer time scales
than the plant master scheduler and thus deals with demand uncertainties through the rolling
horizon heuristic. The master schedule again operates on representative deterministic information,
relies on short term corrections applied by the reactive scheduler to correct infeasibilities and
resolve conflicts, and again is employed under the rolling horizon heuristic. In other words, the
master schedule is totally recomputed when the departures from the current master schedule
become too severe or at some predetermined time interval, whichever arises first. Finally, the
 32

monitoring and control systems account for the smallest time scale variations which are
encountered between actions of the reactive scheduler.
The key research question which must be addressed is, thus, what is the best way to reflect
the uncertainties in demands, order timing and priorities, equipment availabilities, batch quality
indices, resource availabilities, and recipe parameter realizations at each level of the hierarchy.
Clearly, if the multiplant schedule incorporates sufficient slack time, the plant master schedule
gains more flexibility. If the master schedule incorporates sufficient slack time, then the reactive
scheduler will be able to make do with less severe corrective actions and may prompt less frequent
master scheduling reruns. Ofcourse, if too much slack is allowed, manufacturing capacity will be
under-utilized. By contrast, simple use of expected values at each level may lead to infeasible
schedules and excessive continual readjustment or "chattering" of plans at each level, disrupting
the orderly functioning of shipping, receiving, shift scheduling, and materials preparation
activities. At present there is no guidance which can be found in the literature on how "slack"
should be distributed among the levels in a way which adequately reflects the underlying degree
of uncertainty in the various manufacturing inputs. Exploratory work in this area would be quite
valuable in guiding the development of CIM systems for batch operations.

Advances in Operations Applications

While integration of the various decision levels of the CIM hierarchy is the highest priority thrust
for research in the operations domain, that integration is only as effective as the methodology
which addresses the main application areas of scheduling, monitoring/diagnosis, and control.
Important work remains to be carried out in all three of these areas.

Scheduling: The three key issues in scheduling research are investigation of more effective
formulations for dealing with resource constraints, approaches to the reactive scheduling problem
which address a broader range of decision mechanisms for resolving scheduling conflicts, and
formulations and solution strategies for multiplant applications. These application areas are
especially important for the short campaign operating mode likely to be favored in multipurpose
plants of the future.
The key to effective treatment of globally shared resource constraints is to effectively handle
the representation of time. The classical approach of discretizing time in terms of a suitably smaIl
time quantum (see [10]) can be effective. However, in order to be able to accommodate problems
 33

of practical scope, considerably more work needs to be invested on solution algorithms which can
exploit the fine structure of the problem. This is essential in order to have any hope of treating
sequence dependent set-ups and clean-outs. Such refinements must go beyond the conventional
reformulations and cuts typically employed with MILP's. The interval based approach, explored
in [32], has considerable potential but needs further development and large scale testing. Again
careful analysis of structure is required in order to generate a robust and practical solution tool.
Furthermore, since the interval elimination logic, which is part of that framework, appears to have
promise as a preanalysis step for the uniform discretization approach as well as for reactive
scheduling approaches, it is worthwhile to investigate this approach in more detail and generality.
While the work of Cott and Macchietto[ 4] and Kanakamedala et al [9] provide a useful start
further research is required to exploit the full range of reactive scheduling decision alternatives,
which are shown in Fig. 3. In particular, it is important to investigate interval based mathematical
programming formulations which would allow simultaneous adjustments using all of the possible
decision alternatives, while permitting treatment of step details and the realistic timing of materials
transfers. It would appear that it would also be useful to formulate more realistic scheduling
criteria than minimization of deviations from the master schedule. Furthermore, both reactive
scheduling and master scheduling formulations need to be expanded to include consideration of

Resequence ,

ReaSSign
Resources

Reassign
Equipment

Revise
Timing

Figure 3. Reactive scheduler structure

 34

alternative material handling modes. While the movable processing vesseVrigid station
configuration can be treated as another shared resource, the transfer bin concept appears to
introduce some new logistics considerations into the problem. Furthermore, as noted earlier, a
theoretical and computational framework also needs to be developed for linking the master
scheduling and reactive scheduling functions.
Finally, the upper level of the integrated scheduling hierarchy which deals with the coordinated
scheduling of multiple plant sites needs to be investigated. An important consideration at this level
is the incorporation of the logistical links between the plants. Thus, the geographical distribution
of plant sites, the geographic distribution of inventories, and the associated transport costs and
time delays need to be addressed in the scheduling formulation. Moreover, in order to effectively
deal with the interchanges of products and feeds which enter and leave a plant at various points
in the equipment network, as in the case of the plant of Figure 4, reasonably detailed models of
the individual plants must be employed. The conventional lumping of the details of an entire plant
into a single black box can not adequately reflect the actual processing rates which are achieved
when production units are shared among products. Thus, considerable scope exists for large
enterprise scale models and solution methods.

MPP Plant 1
G
J
Packaging x
y
Plant 4
6
R z

i=8-
InvA I w
.---"---- + Inv C

Plant 3 - -8-------+ Inv 8

ln v

Figure 4. Multiplant example with interplant intermediates and inventory

Monitoring and Diagnosis: A key prerequisite for any identification of process deviations is the
ability to identify process trends. In the case of batch operations, such trends will give clues to the
progress of a batch and, if available in a timely fashion, can lead to on-line corrections which can
save reprocessing steps or wasted batches. Effective knowledge based methods for identifying
 35

trends from raw process data need to be developed, directed specifically at the wide dynamic
excursions and trajectories found in batch operations.
Although considerable work has been focused on fault diagnosis in the continuous process
setting, attention needs to be directed at the special opportunities and needs of batch operations
with their step/task structure. For instance, timely forecast of the delayed or early completion of
a task can lead to corrective action which minimizes the impact of that delay or exploits the
benefits of early completion. For this to occur, the diagnosis system must be able to extract that
forecast from the trend information presented to it. As noted earlier, although the monitoring,
diagnosis, and control blocks must be integrated to achieve maximum benefit, such integrated
frameworks remain to be developed.

Control: The control of nonlinear batch operations such as batch reaction remains a major
challenge since typically such batch reaction steps involve complex kinetics, and parameter values
which evolve over time and, thus, are not well understood or rigorously modeled, Consequently,
the key to effective batch control is to develop nonlinear models which capture the essential
elements of the dynamics. Wave propagation approaches which have been investigated in the
context of continuous distillation and tubular reactors offer promise in selected batch operations.
Schemes for identitying and updating model parameters during the course of regular operations
and for inferring properties from indirect measurements are as important in the batch domain as
they are in the continuous. The use of neural networks and fuzzy logic approaches appear to offer
real promise [26].
Although the trajectory optimization problem has been the subject of research for several
decades, the numerics of the optimization ofDAE systems with discontinuities remains an area
for fruitful research. Recent progress with batch distillation is very encouraging (see [13] in these
proceedings) but the routine optimization of the operation of such complex subsystems remains
a challenge given that such applications often involve complex mixtures, multiple phases, and
poorly characterized vapor liquid properties.

Advances in Tools

The design and scheduling applications discussed in the previous sections rely critically on the
availability of state-of-the-art tools for discrete optimization, process simulation, and intensive
input/output information processing. Indeed, the scope and complexity of the applications which
 36

must eventually be handled in order to fully exploit the potential for computer aided batch process
and plant design and computer integrated operations are beyond the capabilities of existing
methodology and software implementations. Therefore, the process systems engineering
community will need to take a leadership role not only in applications development but also in the
design and creation of the enabling tools. In this section, we briefly review essential tool
development needs in the areas of optimization, simulation, and information processing.

Optimization Developments: The preliminary design, retrofit, plant layout, scheduling, and
trajectory optimization applications all are at root large scale 0-1 decision problems with linear and
nonlinear constraints. Indeed, the solution of high dimensionality MINLP and MILP problems with
various special structures is a pervasive and key requirement for batch process systems
engineering. Unfortunately, the limitations of contemporary general purpose algorithms make the
routine solution of problems with over 200 0-1 variables impractical. Indeed as shown in Figure
5, although computing power has grown considerably in the last two decades the capabilities of
general purpose solvers for discrete mathematical programming problems have not kept pace.
Thus, since applications with hundreds of thousands of 0-1 variables can readily arise in practice,
it is clear that general purpose solvers are not the answer. Instead as shown by recent
accomplishments within and outside of chemical engineering, a high degree of exploitation of
problem structure must be undertaken in order to achieve successful, routine solution. Such
enhancements of solution efficiency typically involve not only reformulation techniques,
exploration of facets, cut exploitation, and decomposition techniques but also use of special
algorithms for key problem components, specialized bounding techniques, primal/dual
relationships, graph theoretic constructions and very efficient implementations of key repetitive
calculations. Since the software development effort involved in designing and implementing a
special purpose solver, tailored for a specific application, which employs all of these enhancements
is very large, it is essential that a framework and high level tool kit for algorithm developers be
created for efficiently building and verifying tailored algorithms [18]. The core framework of such
a solver would consist of the branch and bound structure as this lies at the root of all 0-1 problem
solution strategies but integrated within this framework would be a range of algorithmic tools,
including features for exploiting parallel and distributed computing, data compression techniques,
and caching techniques. In view of the major effort such a development would entail, it is essential
that all software components which are available or extractable from the existing commercial and
 37

c:mouter
caoaoility

material &
resource planning
caoability

data
rec:nciliation
capaoility

sc:,eduling
and planning
caoability

Time (Years)
Figure 5. Gap between optimization capability and computer capability

academic inventory be incorporated in the proposed framework.

A key feature of the proposed framework would be the provision of a capability for
systematically testing the performance of any particular tailored algorithm and thus discovering
and exposing its weak points. In the literature, developers of specialized 0-1 solution algorithms
typically only report computational results for a small number of problems, perhaps those which
exhibit the most favorable performance, and from these draw broad conclusions about the
potential of the strategy for a whole class of problems. Unfortunately, for combinatorial problems
such generalizations are in almost all cases invalid. Clearly, in studying an algorithm it is important
not only to identify the structural and data features which make it particularly effective but also
to identify those for which its performance will substantially deteriorate. This is especially
important for industrial application in an operating environment where reliability and predictability
are critical for acceptance and continued use of a technology. To facilitate such rigorous testing,
Pekny et al [18] propose the creation of an adversary, possibly built using AI methods and genetic
algorithms, which would purposely attempt to find data instances that would lead to algorithm
performance deterioration. In view of the practical implications of such a capability, its
investigation should be accorded the highest priority for future research. It may indeed be an
excellent opportunity for collaborative work between several systems research groups.
Finally, in addition to providing a framework for efficient construction of 0-1 solution
 38

algorithms for use by an expert, a shell needs to be provided which will allow the application user
to employ the tailored algorithm without concern for its fine technical detail. This shell should also
provide the user with capabilities for interpretation of the quality and robustness ofthe solution.
Although a general LP type sensitivity analysis is not available for discrete optimization problems,
intelligent bounds and procedures for generating approximate solutions should be developed which
might generate sensitivity-like information under the control of a rule based system.

Process Simulation Developments: While analytical and algebraic models of the MILP and
MINLP form can be extremely powerful tools for design and schedule optimization, such models
genera11yare simplification and approximations of more complex physical phenomena and decision
processes. Thus the solutions generated using these models must be viewed as good estimates
which ultimately must be refined or at least validated using more detailed models described in
terms of differential algebraic equations, stochastic elements, and detailed operating procedures.
In the continuous processing domain, process simulation systems have served as vehicles for the
creation of such more detailed models. The BATCHES system (see [3] as reported at this AS!)
offers such a tool for the simulation of combined continuous/discrete batch operations and recent
developments at Imperial College also point in that direction [14]. While BATCHES is an
effective, practical tool in its present state, it is limited in three aspects: efficient solution oflarge
scale DAE systems with frequent discontinuities, flexible description of tailored batch operating
decision rules, and optimization capability.
BATCHES marches through time by integrating the currently active set ofDAE's from one
state/time event to the next using a widely available DAE solver. Since in principle the set of active
DAE's changes with each event, the previous solution history can not be directly utilized in
restarting the integration process at the completion of the logic associated with the current event.
The resulting continual restarting of the solver can be quite demanding of computer time,
especially for larger scale nonlinear models. Research is thus needed on more efficient ways of
taking advantage of previous solution history during restart, say, in the form of suitably modified
polynomials, for those equations sets that remain unchanged after an event. Further continuing
research is of course also needed in developing more efficient ways of solving large structured
DAE's.
One of the key differences between batch process simulation and conventional dynamic
simulation is that in the batch case one must model the operational decisions along with the
 39

processing phenomena themselves. In BATCHES, a wide range of options is provided under

which unit to task allocations, resource allocations, and materials transfers, and batch sequencing
choices are defined and executed. However, since any finite choice of options can not encompass
all possibilities, the need does arise for either approximating the desired logic using combinations
of the available options or developing special purpose decision blocks. Because of the extensive
information needs of such decision blocks, creation of such blocks is beyond the scope of a typical
casual user. The need thus exists for developing a high level language for describing tailored
decision blocks which could be employed in the manner in which "in-line" FORTRAN is now used
within several of the flowsheeting systems. A natural language like rule based system would
appear to be the most likely direction for such a development.
Finally, once a batch process simulation model is developed and exercised via a set of case
studies, a natural next step would be to use it to perform optimization studies for design, retrofit,
or operational improvements. This idea is, of course, a natural parallel to developments in the
steady state simulation domain. Regrettably, combined continuous discrete simulations do not
directly lend themselves to the SQP based strategies now effectively exploited for steady state
applications because of three features: the presence of state and time event discontinuities, the
frequent model changes which are introduced as the active set of equipment or their modes of
operation changes over simulation time, and the discontinuities introduced by the Monte Carlo
aspects of the simulation. As a result the optimization of combined continuous-discrete simulations
has to date only been performed using direct search methods which treat the simulation model as
a black box. The challenge to the optimization community is to develop strategies which would
allow more direct exploitation of the structure of the simulation model (the "gray" box approach)
for more effective optimization.

Infornwtion Processing Developments: One of the key characteristics of batch operations is the
large amount of information required to describe a design or scheduling application. This
information includes detailed recipe or task network specifications for each product, the equipment
specifications, task suitabilities and inter-unit connectivities, the operating decisions and logic, the
production requirements and the initial condition of the entire plant. The quantitative description
of the operation of a batch plant over a specified time period is also quite information intensive
as such a description must cover the activity profile of each processing unit, transfer line or
mechanism, and resource over that time period. One of the challenges to the systems community
 40

is to develop effective means of generating, validating, maintaining, and displaying this mass of
information in a way which enhances understanding of the operation or design. Graphical
animation as is provided in BATCHES does help in qualitative assessment that the plant is
operating in a reasonable way. The colorful Gantt charts and resource profile charts made available
in contemporary scheduling support software such as [21] are certainly helpful. Nonetheless these
display tools provide the information about the operation essentially as a huge flat file and, thus,
overload the user with detail. Intelligent aids are needed that would help in identifying key
operational features, bottlenecks, and constraints and thus focus the users attention on critical
problem elements. An object oriented approach which allows one to traverse in the information
domain both in extent and in depth, as dictated by analysis needs, may be a useful model.
A further research need is to develop a flexible data model of batch operations which would
provide a structured, common information framework for all levels and tools employed in the
batch operations CIM hierarchy. A prototype data model for batch scheduling applications was
proposed by Zentner [33] and a simulation specific data model implemented using a commercial
data base is employed within BATCHES. The need for such a data model which is application
independent was recognized in [5] in the course of executing a limited integration study. The key,
of course, is application independence. The step level description required in a BATCHES
simulation differs only in some further details from the description required for a sequencing
control implementation. The step level description could also be employed in a rescheduling
application, while a task level aggregation might suffice for master scheduling purposes or for a
retrofit design application. This data model shOirld be supported with generalized consistency
checking and validation facilities which are now scattered across various applications and tools
such as the BATCHES input processor, the input preprocessors developed for various scheduling
formulations, and the detailed sequencing control implementation codes provided by control
vendors. Such unified treatment of process and plant information clearly is an essential prerequisite
for computer aided engineering developments as a whole and CIM implementations in particular.

Summary

In this paper, the future directions oftechnicai developments in batch process systems engineering
have been motivated and outlined. In the process design domain, methodology to support the
synthesis and definition of task networks, approaches for quantitatively balancing plant flexibility
 41

with demand and product uncertainties, retrofit design aspects including heat integration, and
quantitative approaches to plant layout were proposed for investigation. In the operations domain,
integration of the levels of the CIM hierarchy, especially of the multiplant, individual plant and
plant reactive scheduling levels and ofthe monitoring, diagnosis and control levels were offered
as high priority developments. The general problem of explicit treatment of uncertainty in the CIM
hierarchy is a highly appropriate subject for basic study at the conceptual and quantitative levels.
Operations applications requiring further attention include treatment of time within resource
constrained formulations, a broader investigation of reactive scheduling strategies, and the study
of multiplant scheduling formulations. Intelligent trend analysis to support diagnosis and further
developments in low order nonlinear modeling for control purposes also offer significant promise
for batch operations. In the area of tool development, the need for a flexible and well integrated
framework for discontinuous optimization was proposed, including provisions for both a
developer's and an algorithm user's view of the tool and the provision of an adversary feature for
algorithm testing. In the simulation domain, requirements for improvements in the solution of
DAE systems, flexible description of operational rules, and optimization capabilities were noted.
Finally, in the information processing area, the case was made for intelligent aids for plant and
process data analysis, visualization, and interpretation as well as the need for a batch operations
data model, which would form the basis for computer aided engineering developments. The scope
of these themes is such as to offer challenging and fruitful research opportunities for the process
systems engineering community well into the next decade.

Acknowledgment

This presentation benefited considerably from the ideas on these topics which have been developed
by my colleagues in the Purdue Computer Integrated Process Operations Center, namely, Profs.
Ron Andres, Frank Doyle, Joe Pekny, Venkat Venkatasubramanian, Dr. Mike Zentner, our
collective graduate student team, and our supportive industrial partners.

References

I. S. Bacher: Batch and Continuous Process Design. Paper 33d, AlChE National Mtg., Houston (April, 1989)
2. 1. Biegler: Tailoring Optimization algorithms to Process Applications. Comput. Chem. Eng., ESCAPE I
supplemental volume (\992)
3. S. Clark and G. Ioglekar: General and Special Purpose Software for Batch Process Engineering. This volume p.
376
 42

4. BJ. Cott and S. Macchietto: A General Completion Time Detennination Aigorit1un for Batch processes. AlChE
Annual Meeting, San Francisco, (Nov. 1989)
5. C.A. Crooks, K Kuriyan, and S. Macchietto: Integration of Batch Plant Design, Automation, and Operation
Software Tools. Comput Chern. Eng., ESCAPE-l supplernental volume (1992)
6. 1. B. Edgerly: The Top Multinational Chemical Companies. Chemical Processing, pp.23-31, (Dec. 1990)
7. S. Jayakumar and G.Y.Reldaitis: Graph Partitioning with Multiple Property Constraints for Multifloor Batch Plant
Layout Paper 133d, AlChE Annual Mtg. Los Angeles (Nov., 1991). See also, Comput. Chern. Eng., 18, 441-458
(1994)
8. S. Jayakumar: Chemical Plant Layout via Graph Partitioning. PhD. Dissertation, Purdue University, May, 1992
9. KB. Kanakameda1a, V. Venkatasubramanian, and G.Y.Reldaitis: Reactive Schedule Modifications in Multipurpose
Batch Chemical Plants, Ind. Eng. Chern. Res., 32, 3037-3050 (1993)
10. E. Koodili, C.C. Pantelides, andR W.H Sargent: A General Aigorit1un for Scheduling Batch Operations. Comput.
Chern. Eng., 17,211-229 (1993)
II. 1. Lee and G. V.Rekiaitis: Optimal Scheduling of Batch Processes for Heat Integration. I: Basic Formulation.
Comput. Chern. Eng., 19,867-882, (1995)
12. KB. 1oos: Models of the Large Chemical Companies of the Future. Chemical Processing, pp. 21-34 (Jan. 1990)
13. S. Macchietto and 1. M. Mujtaba: Design of Operation Policies for Batch Distillation. This volume, p. 174
14. C.C. Pantelides and PJ. Barton: The Modeling and simulation of Combined Discrete/Continuous Processes.
PSE'91, Montebello, Canada (August, 1991)
15. S. Papageorgaki and G. V.Rekiaitis: Retrofitting a General Multipurpose Batch Chemical Plant. Ind. Eng. Chern.
Res. 32, 345-361 (1993)
16. S. Papageorgaki and G. V.Rekiaitis: Optimal Design of Multipurpose Batch Plants: Part 1, Formulation and Part
2, A Decomposition Solution Strategy. Ind. Eng. Chem. Res., 29, 2054-2062, 2062-2073 (1990)
17. S. Papageorgaki, A.G. Tsirukis and G. V.Rekiaitis: The Influence of Resource Constraints on the Retrofit Design
of Multipurpose Batch Chemical Plants. This volume, p. 150
18. 1. Pekny, V. Venkatasubramanian, and G. V.Rekiaitis: Prospects for Computer Aided Process Operations in the
Process Industries. Proceedings of COPE-91 , Barcelona, Spain (Oct, 1991)
19. H. 1. Reinhart and D.W.T. Rippin: Design of Flexible Batch Plants. Paper 50e, AlChE Nat'l Mtg, New Orleans
(1986)
20. G. V. Reklaitis: Progress and Issues in Computer Aided Batch process Design. In Proceedings of Third Int'l
Conference onFotmdations of Computer Aided Process Design, CACHE-Elsevier, New York, pp.24 1-276 (1990)
21. Scheduling Advisor, Stone & Webster Advanced Systems Development Services, Boston, MA 202210. (1992)
22. N. Shah and C. C. Pantelides: Design of Multipurpose Batch Plants with Uncertain Production ReqUirements. Ind.
Eng. Chern. Res., 31,1325-1337 (1992)
23. N. Shah and C. C. Pantelides: Optima1Long Term Campaign Planning and Design of Batch Plants. Ind. Eng. Chern.
Res., 30, 2308-22321(1991)
24. K Tsuto and T. Ogawa: A Practical Example of Computer Integrated Manufacturing in Chemical Industry Japan.
PSE'91, Montebello, Canada (August, 1991)
25. J.A. Vaselnak, I.E. Grossmann, and A.W. Westerberg: Heat Integration in Batch Processing. Ind. Eng. Chern.
Process Des. Dev., 25, 367-366(1986)
26. V. Venkatasubramanian: Purdue University, School of Chemical Engineering, private communication (May, 1992)
27. V.T. Voudouris and I.E. Grossmann: Mixed Integer Linear Programming Reformulations for Batch Process Design
with Discrete Equipment Sizes. Ind. Eng. Chern. Res., 31,1315-1325, (1992)
28. H.S. Wellons and G.V.Rekiaitis: The Design of Multiproduct Batch Plants under Uncertainty with Staged
Expansion. Comput. Chern. Eng., 13, 115-126 (1989)
29. HS. Wellons: The Design of Multiproduct Batch Plants under Uncertainty with Staged Expansion, PhD
Dissertation, Purdue University, School of Chemical Engineering, December, 1989
30. T. 1. Williams: A Reference Model for Computer Integrated Manufacturing: A Description from the Viewpoint of
Industrial Automation. ISA, Research Triangle Park, N.C.(1989)
31. T. J. Williams: Purdue Laboratory for Applied Industrial Control, private communication (April, 1992)
32. M. Zentner and G.V.Rekiaitis: An Interval Based Mathematical Formulation for Resource Constrained Batch
Scheduling. This volume, p. 779
33. M Zentner: An Interval Based Frameworldor the Scheduling of Resource Constrained Batch Chemical Processes.
PhD. Dissertation, Purdue University, School of Chemical Engineering, May, 1992
Role of Batch Processing in the Chemical Process Industry

Michel Lucet, Andre Charamel, Alain Chapuis, Gilbert Guido, Jean Loreau

RMne-Poulenc Industrialisation, 24 Avenue Jean Jaures, 69151 Decines, France

Abstract: As the importance of batch processing increases in the Chemical Process Industry,
plants are becoming more specialized, equipment is standardized, computer aided process
operations methods are being improved and more widely used by manufacturers.
In 1980, the management of Rhone-Poulenc decided to develop fine chemical, specialty
chemistry, pharmaceutical, and agrochemical products rather than petrochemicals. Twelve years
later, Rhone-Poulenc has become an important producer of small volume products and has
acquired certain skins in this domain.

Keywords: Batch equipment; standardization; batch plant types; operations sequences; flexibility.

Batch process for low tonnage

A majority of chemical products whose production rates are less than 1000 t/y are unable to
support either significant amounts of research and development or major capital investments by
themselves. Therefore, they are processed in existing batch plants in a very similar mode to the
laboratory experiments involved in their invention.

Different kinds of batch plants

We distinguish four kinds of batch plants depending upon the average amount of each product
processed during one year.
 44

1. Pilot batch plants (zero to 30 t/y of each product)

These plants are devoted to new products: samples are made to test the market.
Products are ordered in very small quantities.
2. Flexible and polyvalent batch plants (30 to 300 t/y of each product)
These plants are used to process a large number of products.
The recipes of the different products may vary significantly from one product to another.
3. Multiproduct batch plants (300 to 700 t/y of each product)
These plants run a small number oflong campaigns. Often, the recipes are very similar from
one campaign to another.
4. Specialized batch plants (700 t/yand above)
The plant processes the same product all year long.

Standardization of equipment

We need a maximum level of flexibility to respond to the random demand for chemical products.
We have to be able to process a given recipe in a maximum number of different plants. So, we
have defined a number of standard equipment that will be the same in different plants; they are:
- Reactor under pressure
- Reactor for corrosive compounds
- Reactor at atmospheric pressure
- Distillation linked with reactor
- Rectification
- Crystallization
- Forming of solids, casting
- Liquid-liquid extraction
In Figure 1, a schematic diagram of a standard pressure reactor is given while Table 1 exhibits
the statistics about the frequency of use of standard equipment types which arises in the processing
of 87 of our products.
This standardization of equipment allows us to have uniform control procedures transferable
from one plant to another.
 45

EF cold wster
Ell chiliad waler
V6 6 bars liteam
C6 6 bars condensate

Figure 1: Pressure reactor

Table 1. Statistics for over 87 products - Use of equipments

Pressure reactor 23
Corrosion resistant reactor 25
Atmospheric reactor ......................................................... 62
Distillation linked with reaction 74
Rectification ..................................................................... 52
Crystallization .................... '................................................ 46
Flaking, casting .............. ................................................. 35
Liquid-liquid extraction 35
Phases ................................................................................ 345
Average phases/product 4
 46

Favored links between standard equipment

In Figure 2, a statistical overview ofthe sequences of equipment use is presented. Some favored
links are immediately apparent. The number of standard equipment in batch plants have to reflect
these figures, so that the plant is technically adaptable to give a good response to the market
demand. Moreover, sometimes, some phases of production of new products are slightly changed
to fit correctly the equipment that is present in pilot batch plants. When the product is developed
independently, the constraint of adaptability to the equipment becomes less and less effective.

CD Reactor under pressure

(2) Corrosive reactor
Q) Atmospheric pressure reactor
@ Distillation over a reactor
® Rectification
@ CrystallizatIOn
(J) liqUid-liquid extractIOn
@ Miscellaneous

~
~~~~--------------------------~®
t 40
final
~
product

Figure 2. Statistics over 345 processing phases (percentage of consecutive uses of equipments)

The batch diagram logic

The sequence of operations in the production of a product is shown in block diagram form in
Figure 3.
A mass balance is made, to yield the quantities transferred from one block to the following one.
This information is displayed phase by phase.
 47

Next, a Gantt Chart is drawn showing task by task, subtask by subtask, the occupation time in
each equipment. There is also an option to compute the demand on resources such as manpower,
electricity, steam, etc ... if these constraints are active in the plant. The program does not handle
the constraints on the availability resources. It only shows which subtask requires a given resource
and allows the user to slightly modify the chart.

A-------.,
+ - - - - t..~1 REACTION

C L
.....0 - - - - L_D_R_Y_IN_G_...l.
\flunATIOtJ/
II-L 1--11

Figure 3. Sequence of operations

The control of batch processes

The sequence of operations is also described at the level of the subtask, for example:
- open the output valve
- wait until the mass of the reactor is less than 1 t
- wait 10 more minutes
- close the output water
- etc ...
So the whole operation is logically represented by block diagrams, then tasks, then subtasks,
then sequence of operations.
 48

Optimization of batch processes

There are different levels of optimization. For new products, involving small quantities, the
optimization involves in the better use of existing plants. For products more mature that are
processed in bigger quantities and more often, there is a need to optimize the process by itself
This optimization is mainly obtained byiroproving the reaction part of the process from one run
to the following one. We need automatic data collection from the process and computer assisted
analysis of the present and past data to achieve better running parameters.

Industrial life of product - flexibility

Processing in different batch plants

If frequently happens that some phases of processing of a given product are made in different
plants through the world. It also happens that some phases are processed in plants of contractors.
Thus the planning manager has to take in account a large number of possibilities in processing
costs and time delay.

Flexibility in policy storage

There is a real storage policy for products run in long campaigns. The storage of the final products
has a cost and this depends upon the storage capacity for this kind of products. For intermediate
storage during a short campaign we sometimes use temporary storage - a cart for example.

Conclusion

As chemical products tend to be precisely tailored to sharp specifications. the number of small
products is increasing, the processing of these batch products is at the moment far from being
optimized as it is for continuous large products. Even if some standardization is made at the
moment, each product by itself cannot justify extensive studies. So we have to develop and
improve automatic method to optimize these processes.
Present Status of Batch Process Systems Engineering
in Japan

Shinji Hasebe and lori Hashimoto

Department of Chemical Engineering, Kyoto University, Kyoto 606-01, Japan

Abstract: Rapid progress in computer technology has had tremendous effect on batch plant
operation. In this paper, the present status of batch plant operation in Japan is reponed first by
referring to the questionnaires. The main purpose of the introduction of CIM in chemical plants
is to produce various kinds of products with a shon lead time without increasing inventory. In
order to accomplish this purpose, the development of a sophisticated scheduling system is vital.
The role of the scheduling system in ClM is discussed next. In addition to the development of
computer systems, development of hardware for the batch plant suitable for flexible
manufacturing is also imponant to promote CIM. Recently, a new type of batch plant called a
"pipeless batch plant" has received great attention from many Japanese companies. The
characteristics of pipeless batch plants and their present status are explained, and a design
method and future problems are discussed.

Keywords: Batch plant, computer integrated manufacturing, scheduling, pipeless plant

1. Introduction

An increasing variety of products have been produced in batch plants in order to satisfy
diversified customer needs. The deadline requirements for the delivery of products have also
become increasingly severe. In order to deliver various kinds of products by a given due date,
each product has to be stocked or frequent changeovers of the plant operation are required to
produce required products just in time. As a result, the inventory cost and the changeover cost
increase and the productivity of the plant decreases.
In the 1980s, the rapid progress of computer technology accelerated the introduction of
computer control systems even into the many small- to medium-sized batch plants. And it
contributed to the reduction of manpower. In recent years, in order to cope with increases in
product types, the development of Computer Integrated Manufacturing (CIM) system is being
 50

promoted actively in both continuous and batch plants. The dominant purpose of the
development of CIM is to produce various kinds of products under severe time constraints
without increasing the amount of inventory or decreasing the plant efficiency.
In this paper, the present status of batch plant operation in Japan is reponed first by
ref«rring to the questionnaires which were distributed by the Society of Chemical Engineers,
Japan. Then the problem of CIM in batch chemical plants is discussed from the viewpoint of
the Just-in-Time om production system which is succ~ssfully used in assembly industries.
Next, considering the imponant role a scheduling system plays in CIM, the present status of the
study on scheduling problems in Japan and future problems related to scheduling are discussed.
In addition to the development of computer systems, development of hardware for the
batch plant suitable for flexible manufacturing is also imponant to promote CIM. Recently, a
new type of batch plant called a "pipeless batch plant" has received great attention as a new-
generation production system. The pipeless plant has a structure which is suitable for the
production of various kinds of products. The difference between the ordinary batch plant and
the pipeless plant, and the present status of the introduction of the pipeless plant in Japanese
chemical industries are reponed.

2 Present Status of Batch Plant Operation

In order to obtain the information on the present status of batch plant operation, three
questionnaires were distributed by the plant operation engineering research group of the Society
of Chemical Engineers, Japan. The first questionnaire was distributed in 1981 and the purpose
was to obtain information on the present status of batch plant operation and on future
trends.[l4] Plant operation using cathode-ray tubes (CRT operation) became familiar in the
'80s instead of operation using a control panel. The second and third questionnaires were
distributed in 1987 and 1990 to obtain information on the present status of CRT operation
[15],[16], and the problems and future roles of plant operators. In this chapter, the present
status of batch plant operation in Japan is discussed by referring to the results of these
questionnaires.
Questionnaire # I was sent to 51 leading companies that use batch plants; 60 plants from 34
companies replied. The classification of the plants is shown in Fig. 1. Questionnaires #2 and
#3 were sent to companies that have continuous and/or batch plants in order to investigate the
present status of and future trends in CRT operation. The classification of the plants is shown
in Table 1.
 51

Figure l. Classification of Batch Plants (Ques.#l) f Batch Plants (Oues #1)

Table l. Classification of the Plants (Ques. #2 and #3)

Type of plants Number of plants

Questionnaire #2 Questionnaire #3

(Group 1)
continuous chemical plant 18 18
(Group 2)
batch chemical, pharmaceutical, 23 42
or food-processing plant
(Group 3)
oil refining or coal-gas 22 20
generation plant

Usually, a product developed at the laboratory is first produced by using a small-size batch
plant. Then a continuous plant is used according to increases in the production demand.
Should a batch plant be regarded as a transitional plant in the change from a pilot plant to a
continuous plant? In order to clarify the present status of batch plants. questionnaire #1 first
asked about the possibility of replacing a batch plant with a continuous plant if the current batch
plant would be rebuilt in the near future. For only 18% of the plants, replacement by a
continuous plant was considered; most of them were pharmaceutical and resin plants.
Figure 2 shows the reasons why batch plants will still be used in the future. Except in
cases where the introduction of the continuous plant is technically difficult, all of the reasons
 52

show that batch plants have some advantages compared with continuous plants. This means
that the batch plant is used even if the continuous plant is technically feasible.
The dominant advantages of batch plants are their flexibility in producing many kinds of
products, and their suitability for the production of high-quality and high value-added products.
As the material is held in the vessel during processing, it is easy to execute a precise control
scheme and perform complicated operations compared with the continuous plant. Technical
problems hindering the introduction of the continuous plant are the difficulty of handling
materials, especially powders, and the lack of suitable sensors for measuring plant conditions.
The main reasons for considering the replacement of a batch plant with a continuous plant
were low productivity and difficulty in automating batch plants. In order to increase the
productivity of multiproduct or multipurpose batch plants, generation of an effective operation
schedule is indispensable. However, in the 60 batch plants that responded to questionnaire #1,
mathematical methods were used only for a quarter of the plants to determine the weekly or
monthly production schedule. A Gantt chart was used for a quarter of the plants, and for half
of the plants an experienced plant manager generated the schedule by hand. The situation has
changed drastically during the last decade due to progress in elM discussed in chapter 3.
In order to automate batch plant operation, introduction of computer control system is
indispensable. In questionnaire #1, computer control systems had been introduced at 56% of
the plants. The dominant purposes of the introduction of computers were manpower reduction,
quality improvement, and safer and more stable operations, as shown in Fig. 3. Productivity
improvement did not gain notice, because the introduction of computers was limited to only

Multiple products
can be produced
High·quality products
can be produced
Introduction of continuous
plant is technically difficult {:;~~==:;:=:=;--'--_---' a: Resins
b : Fine chemicals
Manufacturing cost
c : Pharmaceuticals and
is cheap agricultural chemicals
Contamination can d : Foodstuffs
easily be avoided e : Oil refining
f : Steel and coke
Working period is g : Glass and insulators
not so long h : Paints and dyes
i: Other
Other
o 10 20 30 40
Number of plants

Figure 2. Reasons for Using Batch Plant

 53

batch equipment or batch plants. In order to improve plant productivity by introducing a

computer, it is required to develop a factory-wide computer system, which may be upgraded to
CIM.
Factors obstructing the introduction of computer control in 1981 were the lack of suitable
software. high computer costs, and lack of suitable sensors and actuators as shown in Fig. 4.
Due to rapid progress in computer technology, all of the obstacles shown in Fig. 4 may be
resolved now except for the lack of suitable sensors. Development of methods which can
estimate unmeasurable variables by using many peripheral measurable variables remains as an
interesting research area

Reduction of h
manpower

Improvement of
product quality
Improvement of a: Resins
plant reliability b : Fine chemicals
c : Pharmaceuticals and
Energy agricultural chemicals
conservation d : Foodstuffs
e : Oil refining
Improvement of f : Steel and coke
productivity 9 : Glass and insulators
h : Paints and dyes
Other i: Other

o 10 20 30 40
Number of plants
Figure 3. Purpose of Using Computers (Ques.#I)

Lack of suitable software I

High hardware costs
I
Difficulty of automation I
Analog-type controller
has sufficient ability I
Low reliability of hardware I
t::=::::::;-l ---l
Other
o 10 20
Number of plants
Figure 4. Factors Obstructing the Introduction of Computers (Ques#l)
 54

Questionnaires #2 and #3 were premised on the introduction of computer control system.

From Fig. 5, it is clear that the introduction of CRT operation has advanced significantly since
1982. Especially in plants of group B in Table 1 (batch plants), CRT operation was introduced
earlier than in plants of other groups.
Due to increases in the number of products, it became a crucial problem to improve
sequence control systems so that the addition of new sequences and sequence modification can
be easily executed by operators. From Fig. 6, showing the purposes of introduction of CRT
operation, it is clear that CRT operation was introduced to many batch plants in order to
improve sequence control function.
The other chief purpose of the introduction of the CRT operation was the consolidation of
the control rooms in order to reduce the number of operators. However, for batch plants the
consolidation of the control rooms was not often achieved. This suggests that automation of
the batch plant was very difficult and the plant still requires manual work during production.
Figure 7 shows the effect of the introduction of CRT operation. By introducing CRT
operation, manpower can be reduced significantly in batch plants. Improvements in safety and
product quality are other main beneficial effects of the introduction of CRT operation.
In order to reduce manpower still further and to improve plant efficiency, factory-wide
computer management system must be introduced. Figure 8 shows the state of the art of the
introduction of factory-wide computer management systems in 1987. Although implementation
of such a system had not been completed yet, many companies had plans to introduce such a
system. This system is upgraded to CIM by reinforcing the management function. The
purposes of the introduction of factory-wide computer management systems were reduction of
manpower and speed-up of information processing.

Number of plants
30

o :Group C
20 I!I : Group B
• :Group A

10

o
qS ~-n N -~ ~-M ffi - ~
year

Figure 5. Introduction of CRT Operation

 55

Consolidation of control rooms

Replacement of analog-type
controller

tmprovement of the !8:l : Group A

man-machine interface
o :Group B
Centralized management of
process data
o :GroupC
IlllI : Total
Improvement of operability
in unsteady state

Improvement of management
function of sequence control and ~;;~;;=:J
distributed control ~

Introduction of advanced control

schemes
Other

o 10 20 30 40 50

Figure 6_ Purpose of CRT Operations

Productivity is increased o :Group A

o :GroupB
Quality is increased
o :GroupC
Energy consumption is IlllI : Total
decreased

Manpower can be
reduced

Plant safety is
increased

Other

o 10 20 30 40 50

Figure 7_ Effect of CRT Operation

Automation of plant operation reduces the number of operators. For a batch plant. because
the plant condition is always changing. the possibility of malfunctions occurring is larger than
in a continuous plant. Therefore. the contribution of the operator to the plant operation of a
batch plant is larger than in a continuous plant. In other words. the role of operator becomes
very imponant.
 56

Group A

Group B
Group C
Total
o 20 40 60 80 100
(%)
III : Factory-wide computer system has been
introduced .
• : Introduction of factory-wide computer system
is being contemplated.
~ : Each plant is computerized but total system
has not been introduced.
o There
: is no need to introduce factory-wide
computer system.

Figure 8. Introduction of Factory-wide Computer System

Start-up
26

Cleaning
20

Figure 9. Occurrence of Malfunctions

Figure 9 shows when the malfunctions occurred. It is clear from this figure that
malfunctions occurred during unsteady state operations such as start-up and cleaning. Half of
these malfunctions were caused by operator errors, and 20 % were caused by trouble in the
control systems. The importance of the continued training of operators and the preparation of
revised, more precise operation manuals were pointed out by many companies.
 57

In the early days of CRT operation, the necessity of the control panel as a back-up for CRT
operation had been discussed widely and often. Many managers feared operators would not be
able to adapt to CRT operation. However, they have been happily disappointed, and recently
CRT operation without control panels has become common. It is clear that plant operation has
become more difficult and complicated, because a sophisticated control scheme has been
introduced and interaction between plants strengthened. What is the future role of operators in
an advanced chemical plant? Figure 10 shows future plans for plant operation. For 60 to 65%
of the plants. it is expected that the operation of the plant will be executed by engineers who are
university graduates and have sufficient knowledge of the plant and control systems. For about
20% of the plants. plant operation is expected to become easier as a result of automation and the
introduction of operation support systems.
It becomes clear from questionnaire #3 that most of the batch chemical plants are operated
24 hours per day in order to use the plant effectively. And at almost all the plants. the number
of operators is the same for day and night operations. Nowadays. the general way of thinking
of Japanese young people is changing. Most of the young workers do not want to work at
night. Furthennore, young workers have come to dislike entering manufacturing industries.
causing labor shonages. By taking these facts into account, much effort should be devoted to
reducing night operation. Progress in automation and development of sophisticated scheduling
systems may be keys for reducing the number of operators during the night without decreasing
productivity.

GrOuPAEi~
Group B
Group C
Total
o 20 40 60 80 100
(%)

: Unskilled operators will operate the plant.

~ : University graduates with sufficient knowledge
of the plant will operate it.
0: Multi-skilled workers will operate and
maintain the plant.

Figure 10. i'wure Plant Operation

 58

3. Computer Integrated Manufacturing in Batch Plant

In recent years, due to increases in the number of products, inventory costs have become
considerably large in order to satisfy rapid changes in production demand. Figure 11 shows an
example of the increase of the number of products in the field of foodstuffs [2]. In order to
cope with this situation, CIM is being promoted actively in both continuous and batch plants.
Figure 12 shows the purpose of the introduction of CIM in manufacturing industries [4].
From this figure, it may be concluded that the purpose of introducing elM is to produce more
kinds of products with a shon lead time without increasing inventory. In order to realize such a
system. automation of production is essential, and yet it is not enough. It is equally, or even
more imponant to further promote computerization of the production management system,
including the delivery management systems.
First, let us consider the lust-in-Time (111) production system, which is actively used in
assembly industries and 'is used successfully to produce various products with a small
inventory. The operational strategy of JIT is to produce just the required amount of product at
just the required date with a small inventory. In order to produce various products at just the
required date, frequent changeovers of operation are required. As a result, the changeover time
and the changeover cost increase and the working ratio of machinery decrease. In order not to
decrease the productivity of the plant, the following efforts have been undertaken in assembly
industries:

1250

Number of product 4
1000
CD
In
OJ E
a. :J
?;o 3 o>
U 750
:J III

ea.
"tJ CD
a;
1/1

'5 2 '0
.
Q; 500
.0 .~
E a:
:J
z
250
(1.0)

0
1983 1988

Figure II. Trend in Number of Products and Sales Volume of Frozen Foods at a Japanese Company
 59

Multiple-product and
sma1l-quantity production I
Reduction of lead time I
Integration of production and
delivery I
Inovation of management system I
Reduction of management costs I
Quick responce to customers
I
Reduction of intermediate products
---J
=r
Closer conection between research
and production sections

Precise market research

Reduction of labor costs
-
Improvement of product quality
Closer connection between research
and delivery sections

~
Reduction of raw material costs

Other

o 20 40 60

Figure 12. Purpose of the Introduction elM

I) Improvement of machinery so that changeover time from one product to another is cut
greatly.
2) Development of multi-function machines which can execute many different kinds of
operations.
3) Training of workers to perform many different kinds of tasks.

In JIT in assembly industries, the reduction of changeover time is realized by introducing

or improving hardware, and considerable reduction of inventory is achieved. By introducing
multi-function machines. it becomes possible to maintain a high working ratio even if the
product type and the amount of products to be produced are changed. It is expected that the
benefits obtained from inventory reductions exceed the investment costs necessary for
improving hardware.
In assembly plants. a large number of workers are required. and production capacity is
usually limited by the amount of manpower and not by insufficient plant capacity. Therefore.
in an assembly plant. variations of the product type and the amount of products are adjusted by
using the abilities of the multi-skilled workers and by varying the length of the working period.
On the other hand. chemical plants require few workers. but a great deal of investment in
 60

equipment. Thus having machinery idle is the significant problem. In order to keep high
working ratio of machinery, the inventory is used effectively to absorb the variation of the
production demand. This is one of the reason why extensive reduction of inventory has not
been achieved in chemical industries.
In chemical batch plants, reactors have to be cleaned to avoid contamination when product
type is changed. The need for the cleaning operation increases as product specifications
become stricter. Funhermore, cleaning of pipelines as well as batch units is required when the
product is changed. Effons for promoting the automation of the cleaning operation have been
continued. However, it will be difficult to completely automate the Cleaning operation and to
reduce cleaning time drastically. Increases in changeover frequency decrease productivity and
also increase the required amount of manpower. Therefore, in batch plants much effon has
been devoted to reducing changeover time by optimizing the production schedule rather than by
improving plant hardware.
The reduction of the amount of inventory increases the changeover time and the
changeover cost. Therefore, a reasonable amount of inventory has to be decided by taking into
account inventory and changeover costs. In order to accomplish this purpose, the development
of a sophisticated scheduling system is vital. And in order to rapidly respond to variations in
production demand, inventory status and customer requirements must be transferred to the
scheduling system without delay. In other words, systems for scheduling, inventory control,
and production requirement management must be integrated. For these reasons, the
development of company-wide information systems has been the main issue discussed in the
study of ClM in chemical batch plants. The role of the scheduling system in CIM is discussed
in the next chapter.
Recently two types of batch plants have been developed to reduce the time and cost of the
cleaning operation. One involve the introduction of a "multipurpose batch unit" in which
several kinds of unit operations, such as reaction, distillation, crystallization, and filtration can
be executed [1]. By introducing multipurpose units, the frequency of material transfer between
units can be reduced. However, it should be noted that in a mUltipurpose unit, only one
function is effectively performed during each processing period. For example, the equipment
used for distillation and filtration is idle during the reaction period. This means that the actual
working periods of many of the components which compose a multipurpose unit are very short
even if the entire unit is used without taking any idle time. Therefore, the beneficial
characteristics of the multipurpose unit, such as the reduction of pipelines, should be fully
exploited in order to compensate for this drawback when a plant using mUltipurpose units is
designed.
The other method of reducing the time and cost of cleaning operation is to reduce pipelines
by moving the reactors themselves. Such a plant is called a "pipeless batch plant." The
pipeless batch plant consists of a number of movable vessels; many types of stations where
 61

feeding, processing, discharging, and cleaning operations are executed; and automated guided
vehicles (AGV) to carry vessels from one station to another. Many Japanese engineering
companies are paying much attention to pipeless plants from the viewpoint of the flexibility.
The characteristics of the pipeless plants and the present status of their development are
discussed in chapter 5.

4. Scheduling System of Batch Plants

A scheduling system is one of the dominant subsystems of the production management system.
And the computerization of the scheduling system is indispensable to promote elM. By
regarding a scheduling system as an element of elM, the functions which the scheduling
system should provide become clearer. In this chapter, the relationships between the
scheduling system and other systems which compose elM are first considered to make clear
the purpose of the scheduling system in elM. Then, a scheduling system which has sufficient
flexibility to cope with changes in various restrictions is briefly explained.

Scheduling System in elM

When the scheduling system is connected to other systems, what is required of the scheduling
system by these other systems? For a plant where customer demands are met by inventory, a
production schedule is decided so as to satisfy the production requirement determined by the
production planning (long-term scheduling) system. The scheduling system must indicate the
feasibility of producing the required amount of product by the due date. The response from the
scheduling system is usually used to determine the optimal production plan. Therefore, quick
response takes precedence over optimality of the derived schedule.
Information from the scheduling system is also used by the personnel in the product
distribution section. They always want to know the exact completion time of production for
each product, and the possibility of modifying the schedule each time a customer asks for a
change in the due date of a scheduled material or an urgent order arrives.
If information on the condition of a plant can be transferred to the scheduling system
directly from the control system, information on unexpected delays occurring at the plant can be
taken into the scheduling system and the rescheduling can be executed immediately.
From the above discussion, it becomes clear that the following functions are required of
the scheduling system when it is connected with other systems.
One is the full computerization of scheduling. The scheduling system is often required by
other systems to generate schedules for many different conditions. Most of these schedules are
 62

not used for actual production but rather to analyze the effect of variations in these conditions.
It is a very time-consuming and troublesome task to generate all these schedules by hand.
Therefore, a fully computerized scheduling system that generates a plausible schedule quickly
is required when the scheduling system is connected with many other systems. This does not
decrease the importance of the manual scheduling system. The manual scheduling system can
be effectively used to modify and improve the schedule, and it increases the flexibility of the
scheduling system.
The other function is to generate schedules with varying degrees of precision and speed. In
some cases a rough schedule is quickly required. And in some cases a precise schedule is
needed that considers, for example, the restriction of the upper bound of utility consumption or
of the noon break. Computation time for deriving a schedule depends significantly on the
desired preciseness. Therefore, a schedule suitable to the request in terms of precision should
be generated.
The objective of scheduling systems is twofold: one is to determine the sequence in which
the products should be produced (sequencing problem), and the other is to determine the
starting moments of various operations such as charging, processing, and discharging at each
unit (simulation problem).
There are two ways to solve the scheduling problem. One is to solve both the sequencing
and the simulation problems simultaneously. Kondili, Pantelides, and Sargent [11] formulated
the scheduling problem as an MILP and solved both problems simultaneously. They proposed
an effective branch-and-bound method but the problems that can be treated by this formulation
are still limited because the necessary computations are time-consuming. For cases where
many schedules must be generated, the time required for computation becomes very great. The
other way is to solve the sequencing and the simulation problems separately.
From the viewpoint of promoting CIM, many scheduling systems have been developed by
Japanese companies, and some of them are commercially sold [9],[18],[20). Most of them
take the latter approach. In some systems, backtracking is considered to improve the schedule,
but the production sequence is determined mainly by using some heuristic rules. In order to
determine the operation starting moments at each batch unit, it is assumed that a batch of
product is produced without incurring any waiting time. That is, a zero-wait storage policy is
taken in many cases.
In these systems, creation of a user-friendly man-machine interface is thoroughly
considered, and the optimality of the schedule is not strongly emphasized. That is, the
schedule derived by computer is regarded as an initial schedule to be improved by an
experienced plant operator. The performance index for the scheduling is normally multi-
objective, and some of the objectives are difficult to express quantitatively. It is also difficult to
derive a schedule while considering all types of constraints. For these reasons, the functions
that are used to modify the schedule manually (such as drawing a Gantt chart on a CRT and
 63

moving part of it by using a mouse), are regarded as the main functions of a scheduling
system. However, it is clear that functions to derive a good schedule or to improve the
schedule automatically are required when the scheduling system is connected with many other
systems, as mentioned above.
In addition to a good man-machine interface, two kinds of flexibility are required for the
system. One is ease in schedule modification, and the other is ease in modification of the
scheduling system itself.
A generated schedule is regularly modified by considering new production requirements.
Furthermore, the schedule would also be modified each time a customer asks for a change in
the due date of a scheduled material, an urgent order arrives, or an unexpected delay occurs
while the current schedule is being executed. Therefore, a scheduling system must be
developed so that the scheduling result can be modified easily.
In a batch plant, it is often the case that a new production line is installed or a part of the
plant is rebuilt according to variations in the kinds of products and/or their production rates. As
a result, a batch plant undergoes constant modifications, such as installation of recycle flow,
splitting of a batch, replacement of a batch unit by a continuous unit, etc. A new storage policy
between operations is sometimes introduced, and the operations which can be carried out at
night or over the weekend may be changed. It is important that the scheduling algorithm has a
structure which can easily be modified so as to cope with changes in the various restrictions
imposed on the plant.

Flexible Scheduling System

By taking these facts into account, a flexible scheduling system for multiproduct and
multipurpose processes is developed. Figure 13 shows an outline of the proposed scheduling
system. In this system, a plausible schedule is derived by the following steps:
First, an initial schedule is generated by using a module-based scheduiing algorithm. Each
gi in the figure shows one of the possible processing orders of jobs at every unit, and is called a
"production sequence."
Then, a set of production sequences is generated by changing the production orders of
some jobs for the production sequence prescribed by go. Here, two reordering operations, the
insertion of a job and the exchange of two jobs, are used to generate a set of production
sequences [8]. For each production sequence gi, the starting moments of jobs and the
performance index are calculated by using the simulation program. The most preferable
sequence of the generated production sequences is regarded as the initial sequence of the
recalculation, and modification of the production sequence is continued as far as the sequence
can be improVed.
 64

production sequence

starting lime and P.1.

generation of new production

sequences, g 1 • g 2 ••••• 9 N simulator
from go calculation of the starting
time of each job and the
performance index

Figure 13. Structure of a Scheduling Algorithm

One feature of this system is that the generation of the initial schedule, the improvement of
the schedule, and the calculation of the starting moments of jobs are completely separated.
Therefore, we can develop each of these subsystems independently without taking into account
the contents of others. The concept of the module-based scheduling algorithm and the
constraints which must be considered in the simulation program are explained using the rest of
this chapter.

Module-Based Scheduling Algorithm

In order to make the modification of the algorithm easier, the algorithm must be developed so
as to be easily understood. That is, a scheduling program should not be developed as a black
box. The algorithm explained here is similar to an algorithm that the operators of the plant have
adopted to make a schedule manually. The idea of developing a scheduling algorithm is
explained by using an example.
Let us suppose the problem of determining the order of processing ten jobs at a batch unit.
It is assumed that the changeover cost depends on a pair of successively processed jobs. Even
for such a small problem, the number of possible processing orders becomes 1O! ( = 3.6
million). How do the skilled operators make the schedule of this plant?
 65

They detennine the schedule step by step using the characteristics of the jobs and the plant.
If there are some jobs with early due dates, they will determine the production order of these
jobs first. If there are some similar products, they will try to process these products
consecutively, because the changeover costs and set-up time between similar products are
usually less than those between different products. By using these heuristic rules, they reduce
the number of processing orders to be searched.

The manual scheduling algorithm explained above consists of the following steps:
(1) A set of all jobs (set A in Fig. 14) is divided into two subsets of jobs (set B and set C). Set
B consists of jobs with urgent orders.
(2) The processing order of jobs in set B is determined first.
(3) Remaining jobs (jobs in set C) are also classified into two groups (set D and set E). Set D
consists of jobs producing similar products.
(4) The processing order of jobs in set D is determined.
(5) Products in set D are regarded as one aggregated job.
(6) The aggregated job (jobs in set D) is combined with jobs in set E. Then, set F is generated.
(7) The processing order of jobs in set F is determined.
(8) The aggregated job in set F is dissolved and its components are again treated as separate
jobs.
(9) Finally, by combining set B and set F, a sequence of all jobs can be obtained. In other
words, a processing order of ten jobs is determined.

In this case, the problem is divided into nine subproblems. The algorithm is graphically
shown in Fig. 14. Each ellipse and circle in the figure corresponds to a set of jobs and a job,
respectively. An arrow between the ellipses denotes an operation to solve a subproblem.
Here, it should be noted that the same kinds of operations are used several times to solve
subproblems. For example, steps (1) and (3) can be regarded as a division of a set of jobs, and
steps (2), (4), and (7) are ordering of jobs in a set.
Ideas used here are summarized as follows: First, by taking into account the
characteristics of the problem, the scheduling problem is divided into many subproblems.
Since the same technique can be used in solving some of the subproblems, these subproblems
can be grouped together. In order to solve each group of subproblems, a generalized algorithm
is prepared in advance. A scheduling algorithm of the process is generated by combining these
generalized algorithms. The production schedule is derived by executing these generalized
algorithms sequentially.
One feature of the proposed algorithm is that each subproblem can be regarded as the
problem of obtaining one or several new subsets of jobs from a set of jobs. As the problem is
divided into many subproblems and the role of each subproblem in the algorithm is clear, we
 66

8
c812)
( 0-0--0)

o : A product
c::J: A set of products

G
~)

Figure 14. Scheduling Algorithm Using the Characteristics

production
line 1

production
line 2

OJ : batch unit i
Figure 15. Process Consisting of Parallel Production Lines
 67

can easily identify the part which must be modified in order to adapt the change of restrictions.
As the number of jobs treated in each subproblem becomes small, it becomes possible to apply
a mathematical programming method to solve each subproblem. Since 1989, a system
developed by applying this method has been successfully implemented in a batch resin plant
with parallel production lines shown in Fig. 15 [10].

Simulation algorithm
One of the predominant characteristics of batch processes is that the material leaving a batch unit
is fluid, and it is sometimes chemically unstable. Therefore, the starting moments of operations
must be calculated by taking into account the storage policy between two operations.
Furthermore, the operations that can be carried out at night or over the weekend are limited and
the simultaneous execution of some operations may be prohibited. So, even if the processing
order of jobs at each batch unit is fixed, it is very difficult to determine the optimal starting
moments of jobs at each unit that satisfies these constraints. Here we will try to classify the
constraints that must be considered in order to determine the starting moments of jobs [6].
Production at a batch unit consists of operations such as filling the unit, processing
materials, discharging, and cleaning for the next batch. Each of these operations is hereafter
called a "basic operation." In many cases, it is possible to insert a waiting period between two
basic operations being successively executed. Therefore, in order to calculate the completion
time of each job, the relationship among the starting moments of basic operations must be
considered.
A variety of constraints are classified into four groups:

(1) Contraints on Waiting Period

Four types of interstage storage policies have been discussed in the available literature [3],[17],
[19],[23]:
(a) An unlimited number of batches can be held in storage between two stages (UIS).
(b) Only a fmite number of batches can be held in storage between two stages (FIS).
(c) There is no storage between stages, but a job can be held in a batch unit after processing is
completed (NIS).
(d) Material must be transferred to the downstream unit as soon as processing is completed
(ZW).
It is possible to express the UIS, NIS, and ZW storage policies by assigning proper values
to hij and h'ij in the following inequalities:
t·1 + h··IJ < - t·1 + h··IJ + h'··IJ
- t·J < (1)
where
ti : starting moment of basic operation i,
hij , h'ij : time determined as a function of basic operations i and j.
 68

Eq. (1) can express not only the UIS, NIS, and ZW storage policies but also some other
storage policies, such as the possibility of holding material in a batch unit before the processing
operation. When the FIS storage policy is employed between two batch units, it is very
difficult to express the relationship between the starting moments of two basic Operations by
using simple inequality constraints. Therefore, the FIS storage policy should be dealt with
separately as a different type of constraint

(2) Contraint on Working Patterns

The second type of constraint is the restriction with respect to the processing of particular basic
operations during a fixed time period. In order to make this type of constraint clearer, we show
several examples.
(a) The discharging operation cannot be executed during the night
(b) No operation can be executed during the night It is possible to interrupt processing
temporarily, and the remaining part of the processing can be executed the next morning.
(c) No operation can be executed during the night. We cannot interrupt processing already in
progress as in (b), but it is possible to hold the unprocessed material in a batch unit until the
next moming.
(d) A batch unit cannot be used during the night because of an overhaul.
Figure 16 shows schedules for each of the above constraints. In this figure, the possible
starting moments for the filling operation are identical, but the scheduling results are completely
different.

(3) Utility Constraints

The third type of constraint is the restriction on simultaneous processing of several basic
operations. If the maximum level of utilization of any utility or manpower is limited, basic
operations that use large amounts of utilities cannot be executed simultaneously. The
distinctive feature of this constraint is that the restricted period is not fixed but depends on the
starting moments of basic operations which are processed simultaneously.

I=l : filling
H: processing
(a) ••.•.•~I-I : discharging
(b) ••••.•.••••. tw'i Ivw/: cleaning
(c) 1==1............. ........•... H
(d) ~ ............................. 1==1 I-H
restricted
period I
Figure 16. Schedules for Various Types of Working Patterns
 69

(4) Storage Constraint

In an actual batch process, the capacity of each storage tank is finite. If the FlS storage policy
is employed between two batch units, we must adjust the starting moments of some basic
operations so that the storage tank does not overflow. Holdup at a tank depends not only on
the basic operations being executed at that time but also on the operations executed before that
time. Therefore, there are many ways to resolve the constraint when the FlS constraint is not
satisfied.

By increasing the constraint groups to be considered, calculation time is also increased. It

is possible to develop a simulation program that satisfies the constraints on each group
independently. Therefore, by selecting the constraints to be considered, schedules can be
generated with suitable degrees of speed and precision. Figure 17 shQws an example of part of
a schedule for the process shown in Fig. 15. In Fig. 17, all operations are prohibited between
the period 172 hr to 185 hr, but it is assumed that the holding of material in each batch unit is
permitted. The broken line in the lower figure shows the upper bound of a utility, and the
hatched area shows the amount of utility used.
16
17
IS
14
13 .p,..""",• .1..----1"""'.....
12
11
10

6
o 1
:z:
2

5
6

100 120 140 160 180 200 220 240 260 280 300
Time (hr)

....,... 25 ,-'I
::; 20 I I
I I
I I
~ 15

10

\00 120 140 160 180 200 220 240 260 280 300

Figure 17. Schedule that Satisfies Every Type of Constraint

 70

5. Multi-Purpose Pipeless Batch Chemical Plant

In a multi-purpose batch plant, many pipes are attached to each batch unit for flexible plant
operation. The number of such pipes increases as the number of products increases, and it
eventually becomes difficult even for a skilled operator to grasp the operational status of the
plant. Meticulous cleaning operations of pipelines as well as of the batch units are required to
produce high-quality and high-value-added products. The frequency and the cost of cleaning
operations increase when the number of products is increased. Moreover, costs of the
peripheral facilities for feeding, discharging, and cleaning operations increase when an
automatic operation system is introduced to the plant.
In order to reduce these costs, sharing of these facilities among many batch units and the
reduction of the number and the length of pipelines become necessary. With this recognition,
much attention has been riveted on a new type of plant called a "pipeless batch plant." In this
chapter, characteristics of pipeless batch plants and their present status are explained, and a
design method and future problems are discussed.
There are many types of pipeless batch plants proposed by Japanese engineering
companies. The most common type involves the replacement of one or more processing stages
with a pipeless batch plant [12],[21],[22]. In this type, the pipeless batch plant consists of a
number of movable vessels; many types of stations where feeding, processing, discharging,
and cleaning operations are executed; and automated guided vehicles (AGV) to carry vessels
from one station to another. Waiting stations are sometimes installed in order to use the other
stations more efficiently.
A movable vessel on an AGV is transferred from one station to another to execute the
appropriate operations as shown in Fig. 18. Figure 19 shows an example of the layout of a
pipe less batch plant. This plant consists of six movable vessels, three AGVs, and eight
stations for feeding, reacting, distilling, discharging, cleaning, and waiting. About ten
commercial plants of this type have been constructed during the last five years. Various kinds
of paints, resins, inks, adhesives, and lubrication oils are produced in these plants. It is
possible to use movable vessels instead of pipelines. In this case, batch units are fixed, and the
processed material is stored in a movable vessel and then fed to the next batch unit.
Figure 20 shows a different type of pipeless plant [5]. In this plant, reactors are rotated
around the tower in order to change the coupling between each reactor and the pipes for feeding
and discharging.

Characteristics of pipeless batch plants

A pipeless plant has a structure different from that of an ordinary batch plant. Here the
characteristics of the pipeless plants are explained from the following three points:
 71

m\ 2:"1".~~"
Distilling station
ee i ng station

" !iam ea mg s a Ion

Movable vessel
~ r'T.~
eng ,:team

!it\....
water

Figure 18. Conceptual Diagram of Pipeless Batch Plant

.-- -.,
L ___'
Distilling station

.- - --,
L __ _
,--- ...,
Waiting station 2 Discharging station

© : Vessel

D : Automated guided
Vehicle

g : Vessel on AGV
Storage yard of vessels

Figure 19. Layout of a Pipeless Batch Plant

 72

Feed tank

Coupling
Ventilation pipe --:----rr~'!1.,.,¥~t-.....--Valve

4 - - - - . Reactor

Product tank

Product
Product

Figure 20. Rotary Type Pipeless Batch Plant

Feed tanks Premixing tanks

for additives

Conventional Batch Plant

Discharging equipment

Feeding stations

Pipeless Batch Plant

Discharging stations

Figure 21. Configuration of Conventional and Pipeless Piants

 73

(1) Reduction of the number of pipelines and control valves

For ordinary batch plants, the number of pipes and control valves increases along with
increases of the kinds of raw materials and products. If some raw materials share pipes, it is
possible to reduce the number of pipes even for an ordinary batch plant. However, meticulous
cleaning of pipelines as well as that of the batch unit is then required. Figure 21 shows
examples of plant configurations of a conventional batch plant and a pipeless batch plant [13].
It is clear from this figure that the number of pipelines is drastically decreased, and the plant is
greatly simplified by adopting the pipeless scheme.

(2) Effective use of components

The working ratio of a unit is defined as the ratio of the working period of the unit to the whole
operating period of the plant. A high working ratio means that the process is operated
effectively. Therefore, the working ratio has been used as an index of the suitability of the
design and operating policy of a batch plant.
A batch unit consists of many components for feeding, processing, discharging, and
cleaning. Not all of them are used at all of the operations to produce a batch of product. The
working periods of these components are shown in Table 2. As is clear from Table 2, the
vessel is the only component used at every operation to produce a batch of product. In other
words, the working ratios of some components in a batch unit are not so high.
The working ratios of the components have not been discussed, because these components
have been regarded as being inseparable. The costs of these peripheral components have
increased with the introduction of sophisticated control systems for automatic operation of the
plant. In the pipeless batch plant, each of these components is assigned to a station or to a
movable vessel. Therefore, it is possible to use these facilities efficiently and to reduce the
capital costs for these facilities by determining the number of stations and movable vessels
appropriatel y.

Table 2. Working Period of Components

Feeding Processing Discharging Cleaning Distilling

Vessel 0 0 o o o
Jacket. Heating and 0 o
Cooling facilities
Agitator 0
Measuring tank and 0
Feeding facility
Discharging facility o
Cleaning facility o
Distilling facility o
 74

(3) Increase in flexibility

i) Expansibility of the plant
In a pipeless batch plant, the size of each type of station and that of movable vessels can be
standardized. Therefore, new stations and/or vessels can be independently added when the
production demand is increased.
ii) Flexibility for the production path
As the stations are not connected to each other by pipelines, the production path of each product
is not restricted by the pipeline connections. That is, a pipeless plant can produce many types
of products with different production paths. The production path of each product can be
determined flexibly so that the stations are used effectively.
iii) Flexibility of the production schedule
For many types of stations, the cleaning operation is not required when the product type is
changed. Therefore, the production schedule at each station can be determined by taking into
account only the production demand of each product. It is possible to develop a lIT system
and reduce the inventory drastically.

Design of a Pipeless Batch Plant

The process at which a pipeless plant is introduced must satisfy the mechanical requirement that
the vessel be movable by AGV and the safety requirement that the material in the vessel be
stable during the transfer from one station to another. The decision as to whether a pipeless
plant may be introduced takes into account the above conditions and the cleaning costs of
vessels and pipelines. There are many possible combinations for the assignment of the
components shown in Table 2 to stations and vessels. For example, the motor of an agitator
can be assigned either to the processing station or to the movable vessel. The assignment of the
components to the stations and vessels must be decided by taking into account the working
ratios of these components and the expansibility of the plant.
When all of the above decisions are made, the design problem of a pipeless batch plant is
formulated as follows:
"Determine the number of each type of station, the number and the size of movable vessels, and
the number of AGVs so as to satisfy the given production requirement and to optimize the
performance index."
When the number of products becomes large and the amount of inventory of each product
becomes small, a production schedule must be decided by taking into account the due date of
each production requirement. Inventory decreases inhibit the generation of a good schedule.
This problem cannot be ignored, because the production capacity of the plant depends on the
production schedule as well as on the number and the size of stations and vessels
By taking into account these two factors regarding production capacity, a design algorithm
for a pipeless batch plant was proposed [7]. In the proposed algorithm, the upper and the
 75

lower bounds of the number of vessels, AGV s,and each type of station are fIrst calculated for
each available vessel size. Then, iterative calculations including simulation are used to
determine the optimal values of design variables.
Let us try to qualitatively compare the capital costs of the pipeless batch plant and the
ordinary batch plant. Here, it is assumed that each batch unit has the functions of feeding,
processing, discharging, and cleaning, and that the cost of an ordinary batch unit is equal to the
sum of the costs of a vessel and four types of stations.
When the vessel volume is the same for both the ordinary and pipeless plants, the required
number of batch units in the ordinary plant is larger than the number of stations of any given
type in the pipeless plant. Especially for feeding, discharging, and cleaning operations, the
number of each type of stations is very small because the feeding, discharging, and cleaning
periods are very short. Feeding, discharging, and cleaning equipment has become expensive in
order to automatically execute these functions. Therefore, there is a large possibility that the
pipeless plant is more desirable than an ordinary batch plant
Many mechanical and safety problems must be resolved when a pipeless plant is installed
in place of an ordinary batch plant. For example, the material in the vessel must be kept stable
during the transfer from one station to another, and vessels and pipes must be coupled and
uncoupled without spilling. Therefore, many technical problems must be addressed and
resolved in order to increase the range of application of the pipeless plant.
Expansibility and flexibility for the production of different products are very important
characteristics of the future multipurpose plant. Methods to measure these characteristics
quantitatively must be studied. By assessing these characteristics appropriately, pipeless plants
may be widely used as highly sophisticated multipurpose plants.
By installing various sensors and computers to each vessel, it may be possible for each
vessel to judge the present condition and then decide autonomously to move from one station to
another in order to produce the required product. In such a plant, the information for
production management can be distributed to stations and vessels, and the malfunction of one
station, vessel, or AGV will not affect the others. It is an autonomous decentralized production
system, and is regarded as one of the production systems for the next generation.

6. Conclusion

Due to the increase in product types, inventory costs have become considerably large. In order
to cope with this situation, elM is being promoted actively in batch plants. One way to reduce
inventory is to generate a sophisticated schedule and modify it frequently by taking into account
 76

the changes in plant condition and demand. In order to execute frequent modification of the
schedule, integration of production planning, scheduling, inventory control and production
control systems, and the full computerization of scheduling are indispensable. Development of
a fully computerized and flexible scheduling system is still an important research area in
process systems engineering.
The other way to reduce inventory is to increase the frequency of changeovers. In order to
avoid increases of changeover cost and changeover time,. improvement of plant hardware and
more advanced automation are needed. However, in batch plants, difficulty in handling fine
panicles and the need for meticulous cleaning obstruct the introduction of automation. In many
Japanese companies, introduction of pipeless batch plant is regarded as one of the methods to
cope with this dilemma. In pipeless plants, new equipment can be added independently,
without having to take other units into consideration, and the production path of each product
can be determined flexibly. Expansibility and flexibility for the production of different
products are very important characteristics of the future plant. Design methods of multipurpose
batch plants considering these characteristics quantitatively must be developed.

References

I. Arima, M.: Multipurpose Chemical Batch Plants, Kagaku Souchi (Plant and Process), vol. 28, no. I I, pp.
43-49 (1986) (in Japanese).
2. Doi, 0.: New Production System of Foodstuffs, Seminar on Multi-Product and Small-Quantity Production
in Foodstuff Industries, Society of Chemical Engineers, Japan, pp. 25-29 (1988) (in Japanese).
3. Egri, U. M. and D. W. T. Rippin: Short-tenn Scheduling for Multiproduct Batch Chemical Plants, Compo &
Chern. Eng., 10, pp. 303-325, (1986).
4. Eguchi, K.: New Production Systems in Chemical Industry, MOL, vol. 28, no. 9, pp. 21-28 (1990) (in
Japanese).
5. Funamoto, O. : Multipurpose Reaction and Mixing Unit MULTIMIX, Seminar on multi-product and small-
quantity production systems, Society of Chemical Engineers, Japan, pp. 21-31 (1989) (in Japanese).
6. Hasebe, S. and I. Hashimoto: A General Simulation Programme for Scheduling of Batch Processes, Preprints
of the IFAC Workshop on Production Control in the Process Industry, pp. PSI-7 - PSI-12, Osaka and
Kariya, Japan, (1989).
7. Hasebe, S. and I. Hashimoto: Optimal Design a Multi-Purpose Pipeless Batch Chemical Plant, Proceedings
of PSE'91, Montebello Canada, vol. I, pp. 11.1-11.12, (1991).
8. Hasebe, S., I. Hashimoto and A. Ishikawa: General Reordering Algorithm for Scheduling of Batch Processes,
J. of Chemical Engineering of Japan, 24, pp. 483-489, (1991)
9. Honda, T., H. Koshimizu and T. Watanabe: Intelligent Batch Plants and Crucial Problems in Scheduling,
Kagaku Souchi (plant and Process), vol. 33, no. 9, pp. 52-56 (1991) (in Japanese).
10. Ishikawa, A., S. Hasebe and I. Hashimoto: Module-Based Scheduling Algorithm for a Batch Resin Process,
Proceedings of ISA'90, New Orleans Louisiana, pp.827-838, (1990).
II. Kondili, E., C. C. Pantelides and R. W. H. Sargent: A General Algorithm for Scheduling Batch Operations,
Proceedings ofPSE'88, Sydney, pp. 62-75, (1988).
12. Niwa, T.: Transferable Vessel-Type Multi-Purpose Batch Process, Proceedings of PSE'91, Montebello
Canada, vol. IV, pp. 2.1-2.15, (1991).
13. Niwa, T. : Chemical Plants of Next Generation and New Production System, Kagaku Souchi (plant and
Process), vol. 34, no. I, pp. 40-45, (1992) (in Japanese).
14. Plant Operation Research Group of the Society of Chemical Engineers, Japan: The Current Status of and
Future Trend in Batch Plants, Kagaku Kogaku (Chemical Engineering), 45, pp. 775-780 (1981) (in
Japanese).
 77

15. Plant Operation Research Group of the Society of Chemical Engineers, Japan: Repon on the Present Status
of Plant Operation, Kagaku Kogaku symposium series, 19, pp. 57-124 (1988) (in Japanese).
16. Plant Operation Research Group of the Society of Chemical Engineers, Japan: Repon on the Present Status
of Plant Operation (No.2), unpublished (in Japanese).
17. Rajagopalan, D. and I. A. Karimi: Completion Times in Serial Mixed-Storage Multiproduct Processes with
Transfer and Set-up Times, Compo & Chern. Eng., 13. pp. 175-186, (1989).
18. Sueyoshi, K.: Scheduling System of Batch Plants, Automation, vol. 37, no. 2, pp. 79-84, (1992).
19. Suhami, I. and R. S. H. Mah: An Implicit Enumeration Scheme for the Flowshop Problem with No
Intermediate Storage: Compo & Chern. Eng., 5, pp. 83-91,(1981).
20. Suzuki, K., K. Niida and T. Umeda: Computer-Aided Process Design and Production Scheduling with
Knowledge Base, Proceedings of FORCAPD 89, Elsevier, (1990).
21. Takahashi, K. and H. Fujii: New Concept for Batchwise" Speciality Chemicals Production Plant,
Instrumentation and Control Engineering, vol. 1, no. 2, pp. 19-22, (1991).
22. Takahashi, N.: Moving Tank Type Batch Plant Operation and Evaluation, Instrumentation and Control
Engineering, vol. 1, no. 2, pp. 11-13 (1991).
23. Wiede Jr, W. and G. V. Reklaitis: Determination of Completion Times for Serial Multiproduct Processes-3.
Mixed Intermediate Storage Systems, Compo & Chern. Eng., 11, pp. 357-368, (1987).
Batch Processing Systems Engineering in Hungary

Gyula K5rtvelyessy

Szeviki, R&D Institute, POB 41, Budapest, H-1428, Hungary

Abstract: The research work in batch processing systems engineering takes place at universities
in Hungary. Besides the system purchased from known foreign companies, the Hungarian drug
industry has developed their own solution: the Chernitlex reactor in Chinoin Co. Ltd. has been

distributed in many places because of its simple programming and low price.

Keywords: Batch processing, pharmaceuticals

Introduction

More than 20 years ago, there were postgraduate courses in the Technical University, Budapest
on continuous processing systems. G. A. Almasy, G. Veress and I. M. Pallai [1, 2] were at that
time the persons working in this field. Only the mathematical basis of evaluating a computer aided
design algorithm from data and the mathematical model of the process could be studied. At that
time the main problem in Hungary was that there were not any control devices available which
could work in plant conditions. Today, process control engineering can be studied in all of the
Hungarian Universities. Some of them can be seen in Table 1.

Table 1. Universities in Hungary

Technical University Budapest Scientific University of Szeged

University ofVeszprem University ofMiskolc

Eotvos Lorand Scientific University Budapest

 79

General Overview of Batch Processing Systems Engineering in Hungary

The development in Hungary has moved into two directions: Some batch processing systems were
purchased completely from abroad together with plants. They originated e.g. from Honeywell,
Asea Brown Boveri, Siemens and Eckardt. Naturally, the main user of these systems is the drug
industry, therefore the second direction of our development took place in this part of the industry.
The office of the author, the Research Institute for the Organic Chemical Industry Ltd. is one of
the subsidiary companies of the six Hungarian pharmaceutical firms which can be seen in Table
2. In this review, a survey of independent Hungarian developments made in drug industry is given.

Table 2. Six Hungarian PhannaceuticaJ Companies That Support the Research Institute for Organic Chemical
Industry Ltd.

ALKALOIDA Ltd., Tiszavasvari BIOGAL Ltd., Debrecen

CHINOIN Ltd., Budapest EGIS Ltd., Budapest
Gedeon Richter Ltd., Budapest REANAL Fine Chemical Works, Budapest

Hungarian Research and Developments in Batch Automation

The research work takes place mainly in the universities. In the Cybernetics Faculty of the
University ofVeszprem, there are projects to develop algorithms for controlling the heating of
autoclaves. The other project involves a computer aided simulation based on using PROLOG as
a computer language.

Gedeon Richter Pharmaceutical Works Ltd

Here, they work on fermentation process automation. Figure 1 shows the fermentor and the
parameters to be measured and controlled. They are the temperature, the air flow, the pressure,
the RPM of the mixer, the pH, oxygen content in solution, the level of foam in the reactor, power
 80

consumption in the mixer, the weight of the reaction mass, and the oxygen and CO 2 contents in
the effluent air. The volume of the fermentor is 25 liter. The equipment is for development
purposes and works quite well. It has been used for optimizing some steroid microbiological
oxidation technologies.

Cip
I
I
I
I

Waste

RG-100 FERt1ENTOR

Figure 1: Fermentation Process Automation in Gedeon Richter Ltd.

EGIS Pharmaceuticals Ltd

The other Hungarian batch automation engineering work was done at the Factory EGIS
Pharmaceuticals. They use Programmable Logic Controllers ofFESTO for solving problems of
specific batch processing system engineering. Some examples ofthis are: Automatic feeding of
aluminum into boiling isopropyl alcohol to produce aluminum isopropylate. The feeding of
aluminum is controlled by the temperature and the rate of hydrogen evolution. The problem is that
the space, from where aluminum is fed has to be automatically inertized to avoid the mixing of air
with hydrogen.
Another operation solved by automatic control at EGIS is a crystallization process of a very
 81

corrosive hydrochloride salt, with clarifying. The first washing liquid of the activated carbon has
to be used as a solvent in a next crystallization and then the spent carbon has to be backwashed
into the waste to empty the filtering device. The main problem here was to find and use the
measuring devices which enable a long term work without corrosion.
There is a central control unit room, where these PLC-s are situated and one can follow up
the stages of the process. However there is a possibility of manual control in the plant in case of
any malfunction.

CHINOIN Pharmaceutical Works Co. Ltd

In case of CIDNOIN, quite a different approach was realized. A few years ago, they developed
the CHEMIFLEX Direct system for controlling the temperature of an autoclave for batch
production. The short description of this system can be read in a brochure. Now, CIDNOIN have
developed the CHEMIFLEX Reactor system. The general idea ofCIDNOIN's approach is that
the development of the system has to be made by the specialists at Chinoin, and the client should
work only on the actual operational problems. The well-known steps of batch processing control
engineering can be seen in Table 3. Yet, CIDNOIN uses the so-called multi-phase engineering
method. The clients programming work ofthe system is simple, since there is an in-built default

Table 3. Steps of the Batch Processing Control Engineering

Process Control
Plant Plant Management
Production Line Batch Management
Unit (Reactor) Recipe
Basic Operation Phases and Steps
Devices Regulatory, Sequence (Element) and Discrete Control; Safety Interlocks

control and parameter system. The engineering work can start with the basic operation instead of
steps. That is why the time of installation ofthe system is 2 weeks only and the price is only 10%
of the price of the equipment compared with the usual, simple-phase method, where the cost of
engineering is the same as the price ofthe equipment. The measured and controlled values of the
 82

Table 4. Measurements and Controls in the CHEMIFLEX System

Pressure in autoclave Pressure drop in vapor pipe Pressure in receivers

Mass of reactor and filling Rate of circulation in jacket Liquid level in jacket
Rpm of stirrer Rate of flow of feed Level in supply tanks
pH in autoclave Liquid level in receivers Conductivity in reactor
Temperature in vapor phase Temperature in separator Permittivity in separator
Temperature in autoclave Jacket temperatures (inlout) Pressure in jacket

process can be seen in Table 4. The drawing of the whole system is in the Figure 2.
The heating and cooling system in jacket can use water, hot water, cooling media and steam;
Chemit1ex can change automatically from one system to another. There is a built-in protection in
the system to avoid changing e.g. from steam heating to cooling with cooling media. In case it is
needed, first filling the jacket with water takes place and then changing to cooling media.
Figure 3 and Table 5. show the whole arrangement opportunities and the operations one can
realize with the Cherniflex system, respectively.

Table 5. Operation of the Chemiflex Reactors

Temperature manipulations Boiling under reflux Distillation, atmospheric

Distillation, vacuum Steam-distillation Evaporation, atmospheric
Distillation, water separation Evaporation, vacuum Feeding, time controlled
Feeding, temperature controlled Inertization Feeding, pH controlled
Emptying autoclave by pressure Cleaning autoclave Fi1ling autoclave by suction

The advantage of the multiphase programming can be realized from the Table 6.
The programming of the Chemit1ex system is very simple and the steps to follow can be seen
in the Table 7. There are possibilities to the upgraded programmers to use the second
programming cycle and change the basic parameters of the system.
 ----1':1 ~N""

__
=--=: vtooLZ
1r*1~~r::~l_ I~w,,,.

NIHIOGEI'! y V(XHt.

'HACTANT I IN __ vf[fOi
--f-)---V~-- -S';VF([O
-~-~
[Vm," ~ 8
I --t;:-".j
VREV V[XHl
ST[AH 1M - ---Ci<J--
EXHAUST
00
W
I
[ } __ .I __ { }
VWAltN
(DOLING INl£l ---t;?::}-------- VR[el

(DOLING HEDIUH INLET

---i?~£TIN

OUTlET
[Y.HAUST

VAC.

VMErDUT

(Ol/D[NSAT[ OUTlET

wAHR OUlt[T

Figure 2: The Chemiflex Reactor

 84

Table 6. Comparison of Simple Phase and Multiphase Batch Engineering

Item Simple Phase Multiphase

Programming: Complicated Simple

So!. of Control: In Development ----

Start of Engineering at: Step Basic Operation

Time of Installation: 6-8 0.5 months

Price Compared to Equipment: 100 10%

WATIlR RETURN

COOLING MEDIA
RETURN

VCO

Condensate

Figure 3: The Different Heating Systems in Chemiflex Reactors

 85

Table 7. Recipe Building

1. Name Filling 2. Name of Phase 3. Select Operation 4. Select Parameters

for Operation
5. Fill Basic 6. Extend Building? 7. Fill Extended 8. Select Extended
Parameters YIN (Added) Parameters Parameters

References

I. Almasy, A., Veress, G., Pallai, M.: Optimization of an ammonia plant by means of dynamic programming.
Chemical Engineering Science, 24, 1387-1388 (1969)
2. Veress, G., Czulck, A.: Algebraic Investigation of the Connection of Batch and Continuous Operational Units.
Hungarian Journal oflndustrial Chemistry, Veszprem 4, Sup!. 149-154 (1976)
Design of Batch Plants

L. Puigjaner, A. Espuiia, G. Santos and M. Graells

Department of Chemical Engineering, Universitat Politecnica de Catalunya

ETSEIB, Avda. Diagonal, 647, E-08028, SPAIN

Abstract: In this paper, a tutorial on batch process design is presented. A brief review of the
present status in batch process design is first introduced. The single and multiproduct plant design
problems are considered next and alternate methods of solution are compared and discussed. The
role of intermediate storage is then analyzed and integrated into the design strategy. Scheduling
considerations are also taken into account to avoid equipment oversizing. The complexity of
decision making when incorporating production planning at the design stage is brought out
through several examples. The paper concludes with a summary of present and future
developments in batch plant design.

Keywords: Batch plant, process design, intermediate storage, multiproduct plant, multipurpose
plant, flexible manufacturing

Introduction

Chemical Plants are commonly designed for fixed nominal specifications such as capacity of the
plant, type and quality of raw materials and products. Also, they are designed with a fixed set of
predicted values for the parameters that specify the performance of the system, such as transfer
coefficients or efficiencies and physical properties of the materials in the processing. However,
chemical plants often operate under conditions quite different from those considered in the design.
If a plant has to operate and meet specifications at various levels of capacity, process different
feeds, or produce several products, or alternatively when there is significant uncertainty in the
parameter values, it is essential to take all these facts into account in the design. That is, a plant
must be designed to be flexible enough to meet the specifications even when subjected to various
operating conditions. This is even more true for multiproduct and multipurpose plants, where
 87

alternate recipes and production routes must be contemplated.

In practice, empirical overdesign factors are widely used to size equipment, in the hope that
these factors will compensate for all of the effects of uncertainty in the design. However, this is
clearly not a very rational approach to the problem, since there is no quantitative justification for
the use of such factors. For instance, with empirical overdesign it is not clear what range of
specifications the overdesigned plant can tolerate. Also, it is not likely that the economic
performance of the overdesigned plant will be optimum, especially if the design of the plant has
been optimized only for the nominal conditions.
In the context of the theory of chemical process design, the need for a rational method of
designing flexible chemical plants stems from the fact that there are still substantial gaps between
the designs that are obtained with currently available computer- aids and the designs that are
actually implemented in practice. One of these gaps is precisely the question of introducing
flexibility in the design of a plant. It must be realized that this is a very important stage in the
design procedure, since its main concern is to ensure that the plant will economically be able to
meet the specifications for a given range of operating conditions.
The design of this kind of process can be viewed at three different levels: selection of the
overall structure of the processing network; preliminary sizing of the processing units and
intermediate storage tanks, and detailed mechanical design of the individual equipment items. In
the present work, we focus on the first two levels because the bulk of the equipment used in this
type of processing tends to consist of general purpose, standard items rather than unique,
specialized designs. Therefore, synthesis and sizing strategies for single and
multiproduct/multipurpose batch plant configurations are reviewed. Present trends and future
developments are also dealt with, concluding with a summary of research and development issues
in batch plant design.

Literature review: The problem of determining the number of units in parallel and the sizes of the
units in each stage such as to minimize the capital cost of the equipment, given the annual product
requirements and assuming that all stages operate under ZW, can be posed as a mixed integer
nonlinear prograrruning problem (33,14,18]. Such optimization problems can be solved using the
branch and bound strategy providing that the continuous nonlinear subproblems arising at each
node are convex [1]. Because of the severe limitations of this requirement and the generally large
computing times needed to solve the MINLP problem, approximate solution procedures have been
 88

proposed by Sparrow et al. [33], for the pure batch case and extended to include semi-continuous
units by Wiede et al., [40). Also worth of mention is the work by Flatz [7] who presented a hand
calculation procedure to determine a set of equipment sizes and distribute the overall operating
time among different products based on a selected unit shared by all the products.
More recently, Yeh and Reklaitis [41] have developed an approximate approach for single
product plants which takes into account task merging and splitting. Birewar and Grossmann [3]
further demonstrated that task merging may lead to lower total equipment cost. Better
performance is achieved by the heuristic design method of multiproduct plants developed by
Espuna and Puigjaner [4], which is superior to these earlier methods, typically obtaining designs
within at most 3% of the optimal solution in a few seconds of computer time.
Consideration of intermediate storage in the design has been reported by Takamatsu et al.,
[33] who proposed a combined dynamic programming-direct search procedure which
accommodates location and sizing of intermediate storage in the single product case. The results
ofKarirni and Reklaitis [16, 17] readily show that the storage size is a discontinuous function of
the cycle times of the adjacent units and that storage size can be quite sensitive to batch size and
cycle time. Thus, the approach ofTakamatsu et al. [35], can only yield one of a great number of
local minima in the single product case and is computationally impractical because of the "curse
of dimensionality" in the multiproduct case. Further work in the design of multiproduct plants
with intermediate storage has been reported by Modi and Karimi and Espuna et al. [20, 5).
Simulated annealing has also been used to locate intermediate storage and select operating
modes [22). It has also been demonstrated that a better design can be obtained by incorporating
production scheduling considerations at the design stage [3, 4, 6, 10). Very recently, a
comprehensive methodology has been developed that incorporates all the above elements
(intermediate storage, prediction scheduling) in an interactive strategy that also allows for non-
identical parallel units, in-phase, out-of-phase, mixed operating modes, task merging, task splitting
and equipment reuse [12, 28).
The design of multipurpose plants requires detailed prediction planning and scheduling of the
individual operations for multiple production routes for each product. Two forms of operation can
be considered for these type of plants. In the cyclic multipurpose category, the plant runs in
campaign mode, while in the non-cyclic category non-campaign mode is considered, excluding
general recipe structure and generating aperiodic schedules [26). Early works were limited to
plants with single prediction route for each product [34, 36, 8). Faqir and Karimi [9] extended the
 89

previous work to allow for multiple prediction routes for each product. More recently, a more
general problem formulation was presented [21, 8, 9] which considers flexible unit-to-task
allocations and non-identical parallel units. The time horizon is divided into a number of
campaigns of varying length within which products may be manufactured in parallel. In these
works, semicontinuous units are accommodated but intermediate storage is excluded from design
considerations. The only work to present a comprehensive formulation that includes intermediate
storage and full consideration of scheduling problem has been recently proposed by Puigjaner and
coworkers [23, 12,29].
Many of the models proposed rely on MINLP formulations, but recently there is some
increase on generating MILP models. Voudouris and Grossmann [37] considered the more
realistic case of discrete sizes and reformulated the most classical non-linear models on batch plant
design as MILP problems. Shah and Pantelides [31] presented a MILP model which also considers
the problem of unit-to-task allocation as well as the limited availability of intermediate storage and
utilities. Uncertainty in production requirements has been also considered in the design of
multipurpose batch plants [32]. The scheduling and prediction planning of multipurpose batch
chemical plants has been also addressed [38, 39], considering the formation of single-product and
multiple-product campaigns.
To account for the features usually found in real batch processes like batch mixing and
splitting, intermediates, raw materials, flexible unit-to-task allocation, Kondili et al [19] introduced
the State Task Network (STN) representation that models both the individual batch operations
("tasks") and the feedstocks, intermediate and final products ("states"), which are explicitly
included as network nodes. Using the STN representation, Barbosa-Povoa and Macchietto [2]
developed an MILP formulation of batch plant design and detailed scheduling, determining the
optimal selection of both equipment units and network of connections, considering the horizon
time as a set of discrete quanta of time.
Chemical plant layout is also a problem of high interest, since an efficient plant layout can be
the biggest cost saver after process design and equipment design. This problem has been recently
addressed [15, 27], and can be a growing field in the area of flexible batch chemical processing.
In the following study, we will concentrate on the multiproduct type of prediction networks
and indicate the latest developments in this area. The design of multipurpose plants will be
enunciated for the cyclic mode of operation and present solution trends will be indicated. Current
developments in this field will be discussed in later papers.
 90

The Design Problem

The design problem consists basically in determining the sizing and configuration of equipment
items for given production requirements so that capital and/or operating costs.are minimized.
Prior to the predesign of this kind of processes, the following information is required:
• list of products and the amount of each to be manufactured and the available production
time,
• the individual recipes for each product,
• the size/duty factors for each task,
• the material flow balance for each task of the manufacturing process and flow
characterization,
• the equipment available for performing each task, including:
- the cost/size ratio
- the processing time of each task to be performed on that unit
• a suitable performance function involving capital and/or operating cost components
to determine:
(a) the number of equipment stages and the task allocations
(b) the intermediate storage requirements
(c) the parallel equipment items in each stage
(d) the size capacities of all equipment items
Thereafter, the objective of the predesign calculation is to optimize the sizing of the
processing units by minimizing the selected performance function subjected to specific plant
operating conditions.
The following assumptions are made at the predesign stage, and will subsequently be modified
when additional information coming from the production planning is obtained:
1. Only single product campaigns are considered. When storage costs are substantial, the
demand pattern will determine the proper ordering of production campaigns.
2. Each equipment unit is utilized only once per batch.
3. Parallel equipment units are assigned to the same task and out-of-phase mode is also
permitted.
4. Only the overlapping mode of operation is assumed.
5. A continuous range of equipment sizes is assumed to be available.
6. Multiple equipment items of a given type are identical.
 91

7. Instantaneous batch transfer mode (ZW transfer rule).

The first three categories of variable -(a) through (c)-, as indicated above, define the structure
of the process network and constitute the synthesis problem, while the last category (d) refers to
the sizing problem.

The Single Product Case

It is assumed that the product is to be produced in a process which consists ofM different batch
equipment types and K types of semicontinuous units. Only parallel batch units operating out-of-
phase which reduce the cycle time will be allowed. The size Vj of batch equipment oftype j can
be calculated -once the size factor Sj for the same equipment and the batch size of the product are
known- as follows:

V'-BS'
rJ j = 1..... M
(1)
For the overlapping mode of operation, it has been shown [41] that the optimal sizing problem
can be formulated as a nonlinear programming problem (NLP):

Minimize f(V} Rk)

subject to:

j = l •...• M (2)

k = 1.....K (3)

j = 1,....M (4)

(5)
those constraints indicate that the sizing of batch units must be done in such a way that processing
the product will be met (2), that filling and emptying times are limited by the maximum
semi continuous time involved in semicontinuous equipment k with ~ processing rate and duty
factor Dk (3) and, that the cycle time for a given batch size cannot be less than that required by
any ofthe semicontinuous operations which are involved (5). The cycle time is calculated via the
general expression (4) with m j: parallel units operating out-of-phase.
 92

Additionally, the total production time should be less than or equal to the available production
timeH.

QT<H (6)
B -

where Q is the amount of product required over production time horizon H.

Finally, the available ranges of~quipment sizes require that

(7)

R min ..... R < R max

k .::. k- k (8)

The function to be minimized is of the form

f(Vj RJJ= tnj[CJ1+Cj2vj']+ Lm

. . k[Ck1+Ck2~]
L
(9)
Jol

which reduces to

(10)

in the pure batch case. If"J is fixed, at the optimum

B=QT/H

Then, the only variable B is restricted to satisfY

(11)

The result is a single-variable optimization for the limiting batch size B.

The Multiproduct Case

The general multiproduct problem can be described in the same terms indicated for the single
product case, although the computation time and complexity of the solution increases significantly.
Therefore, reasonable simplifications must be introduced. The introduction of appropriate heuristic
rules helps to simplifY the solution to this problem. In a recent publication [4], the objective
function selected only considers the influence of the equipment sizes on the plant investment cost.
Thus, the objective function becomes the sum of the individual process equipment costs:
 93

(12)
The plant and process specifications and problem assumptions remain the same as before. The
minimization of expression (12) is subject to the reformulated constraints, for all i and j:

(13)

(14)

(15)

(3

p ..
IJ
=p~~)+p~2)
IJ IJ'
[~lPij'
P
m. (16)
J

(Fij+Pij+Ei~
m~ (17)
J

(18)

(19)

Again, these constraints ensure that the sizing obtained will meet the intended production
requirements (13), that the filling and emptying times are limited by the maximum semicontinuous
processing time involved (14, 15), that the limiting cycle time for product Lcannot be less than that
required for any of the batch operations involved (18), and that the total production time cannot
exceed the available time (19). The upper and lower bounds on batch and semi continuous unit
sizes remain as before (7 and 8).
The proposed simplified strategy [4] comprises all of the constraints enumerated above into
a single decision parameter which considers that the makespan or completion time required to
 94

meet the specified production cannot be greater than the available time (19). Then, the only
variables to be adjusted are the equipment sizes. Logically, at the optimum plant production levels
the batch size for each product is
v·
B i .... = min j (mj, S .J) (20)
1J

and the processing time for batch and semi continuous units can be obtained from (14) and (16)
respectively. Then, the limiting cycle time for a given limiting batch size B can be obtained from
(17) and (18).
It follows that the total batch processing time for each product will be the maximum of all
times calculated above (overlapping mode). Then, the time required to meet the specified
production levels can be determined and, consequently, the remaining time for the specific sizing
used will be known. Optimum sizing will be obtained by minimizing the objective function keeping
the remaining time positive.
The optimization procedure is based on the calculation of partial derivatives of the objective
function and associated constraints with respect to the unit sizes. These values will be used to
modify the size of unit L(batch or semicontinuous) depending on the step size hI' The unit 1 to
be modified is selected according to the feasibility of the current sizing point. The unit which most
improves the objective function without excessive loss of marginal time will be selected. When a
non-feasible point is reached, a unit is selected that gives the highest increase in the marginal time
at the lowest penalty cost (objective function). Whenever completion time or boundary restrictions
are violated, the step length hi is decreased accordingly. The optimization procedure ends when
hI values become insignificant. Convergence is accelerated using additional rules which take into
account the size reduction for all units not actually involved in processing time calculations and
by keeping the step lengths within the same order of magnitude for several computation cycles.

Improving the Design: Solution Strategies

Previous formulations of the design problem may prove to be oversimplified, thus giving rise to
an unjustifiable oversizing. Present developments in hardware, together with the use of appropriate
optimization techniques, allow introduction of further refinements in the design procedure that will
eventually result in improvement of the final design. The following items are also considered at
the design stage:
• task allocation of individual equipment and production planning.
• alternative task transfer policies that may incorporate limited storage for intermediates.
 95

Unit-to-task Allocation

Flexible unit-ta-task allocation can be achieved by using the binary variable "ijm that identifies all
possible assignments oftasks j = I, ... , J j to equipment m = I, ... , M for every product i = I, ... , I

1 if j-th task is assigned to m-th unit to produce i-th product

o otherwise
(21)

subject to

M Ii I
LXj;n
m=l
= 1 LXj;n
j=l
= 1 Lx~
i=l
= 1
(22)
Thus, each stage j could have associated with it a set of cycle times given by:

Xijm [ ijm
BL\P~~l
P + P MfJ
(1) (2) (
ijm

tijm - ---':~---D~~--­
Mjm (23)

where M tm and M!m indicate the number of equipment out of phase and in phase at stage j
respectively. For simplicity of exposition, only batch equipment has been considered. The
extension to include semicontinuous units is straightforward.
The limiting cycle time and the batch size equipment constraints become:

(24)

(25)

B·
Vm~ Xijm Sijm-t-
Mjm (26)

and the objective function is

 96

(27)

with

(28)

Ji
P ~ P
Mm = £.. Xijm M jm
(29)
j-I

subject to constraints (19,22, 26), and taking into account (23,24,25,28,29).

Intermediate Storage

The use of intermediate storage (IS) in batch plants increases production flexibility by
de-bottlenecking conflicting process stages, reducing the equipment sizes and alleviating the
effects of process parameter variations.
It has also been noted [10] that intermediate storage can serve to decouple upstream and
downstream trains. This decoupling can occur in two different forms. If the amount stored is of
the order of an entire production campaign, then the trains can operate as two independent
processes, but the storage capacity can be selected just to be large enough to decouple the cycle
times but not the batch sizes.
In general, the maximum number of available locations, S, is calculated by

M+L M+L
S= L L SIIIII (30)
m=l !l=m+l

where Sml1 is given by

(31)
 97

where emil indicates the equipment connectivity, E imll is a binary variable that represents the
stability of intermediates and J\n", which is also binary, indicates the availability of data for IS.
The problem of sizing and location of intermediate storage and its influence on the overall
equipment cost has been studied by Karimi and Reldaitis [161. In multiproduct plants, two
possibilities may occur:
• The same storage location is set for all products
• Different storage locations are allowed for each product
Assuming that the storage cost is negligible compared to the equipment cost, as given by
equation (12), the insertion of intermediate storage has the following overall consequences [24]:
(a) Minimizing the sum of the individual subtrain costs is equivalent to minimizing the plant
cost
(b) The minimum cost with N storage locations s the minimum costs with N-1 storage
locations
(c) The minimum cost with different storage locations for each product s the minimum cost
where all locations are the same
But as the intermediate storage costs become relevant, the objective function to be minimized
takes the form [29]

(32)

The right hand side of expression (32) takes into account the cost of batch Fde (Vm) and
semi continuous equipment Fsc (R1) and the contribution of storage units F is (Zs) to the plant
investment cost. Each term can be calculated as in (27).
As before, the minimization of (32) is subject to the batch size constraint (13), the limiting
cycle time requirement (18), upper and lower bounds on equipment sizes (7,8), including
intermediate storage, and the "surplus· time constraint (19). Additionally, the minimum
productivity constraint for each product in every Atrain generated by the allocation ofIS also has
to be taken into account.

IAj E {O,l}
(33)

where the binary variable IAj indicates the presence ofIS in train A.
 98

The sizing of storage tanks depends on the batch size and cycle time in each train associated
with it. Calculation of the optimum size for each storage unit, taking into account the requirements
for each product, leads to the following general expression, which considers the actual values of
the variables (batch size , p, cycle time SC and storage capacity 9) of the up u and down d trains
associated with it [28].

(34)

Production Scheduling

Typically, in the design of multiproduct plants, the dependency on scheduling considerations is

eliminated by the assumption of long production runs for each product and limiting cycle times
for a fixed set of similar products.
If the products are dissimilar and/or require the use of different equipment stages, the
processing facility may eventually be shared by different products at the same time, giving rise to
simultaneous production of several products and, possibly, a lower design cost. This is the
situation in multipurpose configurations, where multiple production routes may be allowed for
each of the products and even for successive batches of the same product, making it necessary to
introduce scheduling considerations at the design stage [11, 13].
When scheduling considerations are incorporated at the design stage, time and demand
constraints are needed. These constraints cannot be expressed in terms of cycle time and batch
size, as they are not fixed in the general multipurpose case. Instead, time and demand constraints
are described as functions of the scheduling variable

X ijkmn E {O,l} 'd i,j,k,m,n (35)

where

1 if task j of batch n and product i is assigned to the k-th use of unit m

Xijkmn= {
o otherwise

the amount processed in batch n for product i and its bounds

M K.. V
<B~
O -<B·m- In
B~ax=
In {I I
m=l k=!
XUkmn s.~ }
1Jm
(36)
 99

the utilization factor llkm of unit m the k-th time it is used (used capacity / nominal capacity)

(37)

and the overall constraint for batch n and product i

MK..

I,I,Xijkmn ll km ;~ = Bin
m=!k=1 1Jm (38)

Hence, operation times and initial and final times (tkm, TIkm , TFkm) can be calculated:

(I) (2) V m p.(3)

. )
(
tijkmn = X ijkmn P ijm + P ijm ( 11 km ~) 1Jm
n=l ;=1 j=1 1Jm (39)

(40)

The waiting time 1W eventually required is constrained by the "Finite Wait" time (FW) policy
[25, 11, 13].
N I I.

o~ TW km ~ I.I.I. Xijkmn TWij:

n=l ;=1 j=l (41)
Finally, the global time constraint for the problem is:

Vk,m
(42)

Vk,m
(43)

Scheduling variables are also used to determine production for all products:

Vi, j ~ J i
(44)

which is subject to the overall demand Dj constraints

(45)
 100

It is straightforward to show that this general fonnulation reduces to the restricted case of the
multiproduct plant seen before by assuming on a-priori schedule [6].

LXijm
m
= 1 ; 11m =1 (46)

Overall Solution Strategy for the Design Problem

The solution to the design problem in its full complexity as fonnulated above can be obtained
through a multilevel optimization approach that includes (Fig 1):
1. Predesign without intennediate storage.
2. Optimum sizing resulting after intennediate storage allocation.
3. Campaign preparation (scheduling under constraints) and selection.
4. Predesign of general utilities.
In summary, the solution procedure [10, 28, 30] uses two main computational modules
which may be used independently:
• Design Module
• Scheduling and Production Planning Module
The two modules may interact with each other through an optimization algorithm, thus
preliminary plant design. Additional design variables are: following an iterative procedure until a
valid solution common to both modules is eventually reached.
From start to finish, the optimization procedure always remains under the designer's control
in an interactive manner. Thus, specific decisions to be made during the optimization steps, which
are difficult to automate, and valuable personal experience with specific processes can be
integrated into the optimization procedure in a reasonable way. For instance, the designer may
decide to introduce new options which are not included in the set of basic production sequences
already generated, or else he may want to eliminate some of the alternatives which seem
inadequate or incompatible with personal experience in a specific type of processes.
 101

Campaign elaboration
(scheduling with restrictions)

Campaign selection
(under user's control)

Fig. 1. Simplified flowchart of the optimization algorithm

 102

Design Module

This module produces the optimum plant design regardless of its actual operating conditions. The
calculation procedure is based upon the computation of the Surplus Time (SP) in ideal single
product campaigns as described before [4]. It also incorporates finite intermediate storage (FIS)
analysis which uses the following strategy:
a) Plant Design without FIS
b) intermediate storage location
c) initial sizing with intermediate storage
d) final design with FIS

Scheduling and Production Planning Module

The general formulation of the problem described before requires large MINLP problems.
Substantial reduction in computational effort can be obtained by reducing the number of different
campaigns to be analyzed and subsequently selected according to an adequate performance
criterion. Campaign selection is always under the user's control, who can input appropriate
decisions based on his own expertise and know-how.
An optimization algorithm forces the interaction between the two modules through a common
optimization variable, again the "surplus time". Thus, the design module will produce optimum
equipment sizes and production capacities for every input data set including intermediate storage,
assuming that the solution reached offers the maximum profit for all input data. Product recipes
and specific requirements of process equipment are taken into consideration to obtain the
1. The overall completion time (makespan)
2. Product market profile
3. Penalty costs for unfilled demand
Then, the production planning module will try to produce the best schedule -for a given plan
configuration with specific equipment sizes- in the production campaigns that best suit the market
demand profile. Results obtained over long-term periods are described in Table 1.
The alternative solutions given by the production module are subjected to evaluation. The use
of appropriate cost-oriented heuristics under the supervision of an experienced user leads to
suitable modifications of times and production policies. The design module will incorporate such
modifications to obtain a new design which is more suited to actual production requirements, thus
increasing the overall production benefits [6].
 103

Table 1. Alternative solutions given by the Production Planning Module

Production Production FuW Oa:.'oaUon Maximumrequired Maximum

Policy Benefits Production Time Storage Delay

Cover specified demand (NISfZW) Bene f I (PI.I.···.Pl.n) Stock 1 Delay I

Cover specified demand (UISfZW) Bene f2 (P2.1.···.P2.n) Stock 2 Delay 2
LProd.LTasI2tock < Value I Bene f3 (P3.1····· P3.n) Stock 3< Valuea Delay 3

Use all available time (NIS/ZW) Bene fm-l (Pm.I.I •...•Pm.l.n) Time horizon Stock m-I Delay m-I
Use all available time (UlSIZW) Bene fm (Pm.I •...• Pm •n ) Timehorizon Stockm Delaym

(a) To be specified by the user

Retrofitting Applications

When the objective is to increase the set of products manufactured in the plant by using existing
units and adding new ones for specific treatment of these products, the problem of identification
and size selection ofthe new equipment that makes the best use of the existing ones can be solved
by the optimization procedure indicated above, by taking into account the new demand pattern
and plant flow-sheet However, we must remove from the set of available equipment units to be
optimized those already existing in the actual plant. Thus, the problem of producing additional
products can be solved by an optimal integration of specific "new" units into the existing "general
purpose units".
When the objective is to increase the overall production of the plant, the optimal (minimum
cost) solution will usually consist in adding new identical parallel equipment units. Consequently,
it will only be necessary to identifY the places where they should be added and the operating mode
of these new units [4].
In the last case, the final solution could result in oversizing if identical in-phase parallel
equipment units have been introduced to enlarge batch sizes. To avoid this eventual oversizing,
the assumptions made that all parallel equipment units are identical has been relaxed in the design
module. The criterion is that all parallel sets of equipment units operating out-of-phase should
have the same production capacity, and they should be comprised of Mj~ identical units, except
for the existing ones plus one (to adjust production capacity). Since this could produce different
processing times for parallel in-phase units, the greatest processing time must be selected [29].
 104

A Comparative Test Case Study

A sample problem (example 3 of [3]) is used to illustrate the design strategy described before.
Table 2 summarizes the data for this comparative study.

Table 2. Data used for comparative study (from [3], expo 3)

Production requirements: QA = 40,OOOkg. QB = 20.000kg

Cost coefficients: aj = $250. 13j = 0.6
Horizon time: H=6000h.
Size factors (lkg- 1) Batch processing times (h)

Product Product Product Product

stage A B stage A B
1 2 4 1 8 16
2 3 6 2 20 4
3 4 3 3 8 4

Final designs obtained under different test conditions are shown in Table 3.

Table 3. Design and Production Planning results

A B C D E F
ZW ZW B Including C Including DIncluding E Including
(ref. 2) (this work) holdups NIS seasonality stoCk limits

VI 429 429 501 446 445 534

V2 643 643 752 668 667 801
V3 857 857 1003 891 890 1068
BA 214 215 251 223 222 267
BB 107 107 125 111 111 133
Cost 35.973.50 35,977.50 39,513.16 36,819.75 36,790.73 41.042.58
Iterations' 3 5 2 4 10
Surplus time (1) 8+0 140'+420 100 + 180 100 + 180 830+ 96
Max. Stock A (2) 179 224 213 367 441
Max. Stock B (2) 90 112 106 4958 2947
CPU time (3) 3.59 (4) 1.60 5.90 2.43 4.51 10.92

(1) Surplus time is expressed in hours and as the sum of idle time due to lack of demand and useless time remaining until
next plant hold-up.
(2) Initial stock of half a batch is asswned. (3) On a SUN workstation SPARCI. (4) On an mM 3086.
 105

Cases A and B shown similar results for the solution of the design problem when the ZW
policy is considered.
In case C, the design with ZW scheduling is faced with a discrete time horizon. A
multiproduct campaign AB is considered and the time horizon of 6000h contains short and
medium-term periods so that foreseeable shut-downs can be also considered. The time horizon
has 10 medium-term periods with an associated demand estimation for each product of one tenth
the global demand. Every medium-term period comprises 4 short-term periods of 150h each one
including the final shut-down.
Two opposite effects lead to the final design: time savings due to operation overlapping in the
multiproduct campaign tend to reduce the plant and its cost. On the other hand, time wasted
because of holdups tends to increase it. The final result shows the second effect to be the most
important in this case.
Comparison of cases B and C shows that only considering the first effect leads case C to a
cheaper plant, but it will not satisfy the demand if a continuous period of 6000h is not available.
In case D, where NIS policy is contemplated, it is shown that production is increased at
the same time that wasted time is reduced. As a result we have a lower design cost.
In the following case E, demand seasonality is also considered. Global demand for product
B remains the same but demand estimation throughout the whole time horizon is defined for every
medium term period as follows:

Medium term 2 3 4 5 6 7 8 9 10
Demand estimation 1500 500 500 1500 1000 2000 2000 3000 4000 4000

As no restrictions are imposed, the final design is almost the same as in case D. Also, the
production plan is similar because idle time is minimized in the same way. The difference is in the
large stock that a constant production plan causes when faced with the given demand profile.
The last case F introduces limitation of stock capacity into the previous case. A maximum of
3000 kg of stock for product B is allowed in the production plan. In this case, a larger plant will
be needed and, consequently, a larger surplus time, which is minimized under the stock restriction.
Note that the time horizon provided as input for the design module is chosen as the
independent variable. The lowest cost plant satisfying the given overall demand over this
continuous time horizon is next sized using the design module. The output provided by this first
 106

module is the corresponding capital cost and equipment sizing, the latter being automatically used
as the input for the production planning module.
The actual discrete time horizon is used in this second module to accommodate the specified
demand and the extra or surplus time obtained after satisfying demand is the resulting output.
The design strategy used is illustrated in Figures 2 to 5. Figure 2 shows the quasi-linear
relationship between capital cost and the independent time variable. Obviously, at larger horizon
times, smaller and cheaper plants are obtained.

44000

42000

-
,-..
~
'-'

'"0 40000

-
c.I

-;
06. 38000
C'I
U
36000

34000
4500 5500 6500 7500
Design time horizon (h)

Fig. 2. Capital cost vs. design time horizon

Figures 3, 4 and 5 refer to case D. They reveal the discontinuous behavior of surplus and extra
time. Figure 3 shows that extra time is needed if plant design is performed using an input time
horizon greater than 6500 h. The resulting plants will not be able to satisfy the demand.
Figure 4 illustrates similar but opposite behavior for surplus time. A design time horizon of
less than 6500 h. produces overdesign. Plants sized using larger horizon values cannot satisfy the
demand although stilI remains surplus time, that being the useless time due to shutdowns.
The discontinuous behavior of surplus and extra time observed is due to degeneracy. At a
given time horizon, alternate designs with different capital cost are obtained although all they lead
to a production plan with the same number of batches and thus the same extra time.
Therefore, the function to be minimized is the sum of extra and surplus time that is shown in
Figure 5 at the same time scale as Figure 2. The optimum design time horizon obtained this way
will lead to the optimum sizing when introduced in the design module.
 107

800

.-.
--
.c 600
......

--
.§
----
--
400-
cr..u

--
----
~
I";o;l
200

o ~
6000 6200 6400 6600 6800 7000
Design time horizon (h)

Fig. 3. Extra time vs. design time horizon for case D

800,----------------------------

g 600---_
--......
- ----
CII
.5
f Il 400-
:I
C.
r..
:I
fIl 200·

O+-~--r_~~--~~~--~~~
6000 6200 6400 6600 6800 7000
Design time horizon (h)

Fig. 4. Surplus time vs. design time horizon for case D

 108

2000 -r--------------,
4)

~ 1500
'":I
E-
:I
'"
'":I
1000
-......
"'.., •...-.
C. 500 ~...-.-. .-~
.
eo:
i<
W O+--~-_,--T--_.-~-~
4500 5500 6500 7500
Design time horizon (b)

Fig. 5. Extra plus surplus time vs. design time horizon for case D

The Future in Batch Plant Design

In this paper, we have addressed the problems of optimal design of batch plants. Two subproblems
ave been identified: equipment sizing and network synthesis. It has been shown how the
complexity of problem increases significantly from single to multiproduct and then to multipurpose
batch plant design.
Although present formulations of the design problem consider the detailed representation of
the batch system constituents even at the subtask level [25], much still remains to be unveiled in
these main directions:
• Development of a realistic and more integrated framework for the design and retrofitting
of flexible production networks that includes the feedback from planning and scheduling
evaluations of train design performance.
• Adequate treatment of more general (i.e. concurrent) recipe structures.
• Design under uncertainty
• Further development of efficient optimization algorithms capable to efficiently solve large-
scale industrial problems.
• Energy integration and waste minimization.
• Integrated control strategies
• Development of more intuitive interfaces to overcome the difficulties associated with the
use of complex modeling.
 109

Acknowledgments

Support by the European Communities (JOUE-CT90-0043 and JOU2-CT93-0435) and the

Comissio Interdepartamental de Recerca i Tecnologia (QFN89-4006 and QFN93-4301) is
gratefully appreciated.

Nomenclature

Anu Binary variable that indicates the availability of data for the location of storage
between unit m and unit J.1
Bi Batch production capacity for product i
BiJ.. Batch production capacity for product i in the subplant A
Bin Batch production capacity for product i in batch n
Cjl Independent constant for cost calculation (discontinuous equipment)
cj2 Linear constant for cost calculation (discontinuous equipment)
Cj3 Exponential constant for cost calculation (discontinuous equipment)
ckl Independent constant for cost calculation (semicontinuous equipment)
ck2 Linear constant for cost calculation (semicontinuous equipment)
ck3 Exponential constant for cost calculation (semicontinuous equipment)
(1)
cm Independent constant for cost calculation (discontinuous equipment)
(2)
cm Linear constant for cost calculation (discontinuous equipment)
(3)
c Exponential constant for cost calculation (discontinuous equipment)
C:I! Binary variable that indicates ifthere is physical connection between unit m and
unitJ.1
Dj Present market demand of product
Dik Duty factor of semicontinuous equipment k for product i
E-. Emptying time of the discontinuous stage j for product i
IJ
Eiml! Binary variable that indicates the stability of intermediate after unit m and
before unit J.1 for product i
Objective function to be optimized
Filling time of the discontinuous stage j for product i
fractional change (step size) of unit 1 (batch or semicontinuous) during
optimization procedure
H Time horizon
Total number of products
IAi Binary variable that indicates if product i is produced in subplant A
J Total number of tasks in the receipe
K Total number of available semicontinuous equipment
L Total number of available semicontinuous equipment
II) Number of parallel out-of-phase units for the discontinuous stage j
Number of parallel out-of-phase units for the discontinuous stage j
Number of parallel in-phase units for the discontinuous stage j
Number of parallel in-phase units for the semicontinuous stage k
 110

M Total number of available discontinuous equipment

M0 Number of parallel units for the units for the unit m operating out-of-phase
m
MP Number of parallel units for the unit m operating in-phase
m
M.o Number of parallel units for the unit m in task j operating out -of-phase
JDl
M.P Number of parallel units for the unit m operating in-phase
Jm
N Total number of batches to be produced in the time horizon
Pij Processing time ofthe discontinuous stage j for product i
p.~l) Independent constant for processing time calculation (discontinuous stage)
IJ
p.~2) Linear constant for processing time calculation (discontinuous stage)
IJ
p.~3) Exponential constant for processing time calculation (discontinuous stage)
IJ
p.~l) Independent constant for processing time calculation (unit m)
IJm
p.~2) Linear constant for processing time caluculation (unit m)
IJ
p.~3)
IJm
Exponential constant for processing time calculation (unit m)

Qi Production of procuct i
Rtc Processing rate of the semicontinuous stage k
S Total number of possible locations for a intermediate storage unit
SP Surplus time
S ij Size factor of discontinuous stage j for product i
Sis Size factor of storage s for product i
Sijm Size factor of task j for product i using equip m
Smfl Binary variable that indicates if a storage unit can be located between unit m
and unit f1
Ti Limiting cycle time for product i
TiA Limiting cycle time for product i in the subplant ).
TFjn Final time of task j and batch n
~n Initial time oftaskj and batch n
TWjn Waiting time oftaskj and batch n
tij Operation time of the discontinuous j for product i
tkm Operation time of the k-th use of unit m
Vj Sizing of the discontinuous stage j
 111

Sizing of the unit m

Binary variable that indicates if task j of product i is carried out in equipment m
Binary variable that indicates if task j of product i in batch n is carried out in
unit m for the k-th time
Sizing of the intennediate storage s for product i
Sizing of the intennediate storage s

Greek Letters

Pisu batch size of product i in the upstream subplant of the storage s

llkm Utilization factor of unit m the k-th time it is used
9¥ time needed to fill the storage 5 for product i
at time needed to empty the storage s for product i
OJ Transfer time of the discontinuous stage j
.¥ limiting cycle time for the subplant located before the storage 5 for product i
.t limiting cycle time for the subplant located after the storage s for product i

Subscripts

Product number
j Task number
k Sernicontinuous equipment
m Discontinuous equipment
n Job number
s Storage number

Superscripts

(1) Independent parameter for cost or time calculations

(2) Linear parameter for cost or time calculations
(3) Exponential parameter for cost or time calculations
o out-of-phase
P in-phase
u upstream
d downstream
 112

References

I. Balas, E.: Branch and Bound Implicit Enumeration. Annals of Discrete Mathematics, 5, pp. 185, North-Holland,
Amsterdam, 1979.
2. Barbosa-Povoa AP., Macchietto S.: Optimal Design of Multipurpose Batch Plants.!. Problem Formulation.
Computers & Chemical Engineering, 175, pp. S33-38, 1992.
3. Birewar, D.B., Grossmann, I.E.: Simultaneous Synthesis, Sizing and Scheduling of Multiproduct Batch Plants.
Paper presented at AIChE Annual Meeting, San Francisco, 1989.
4. Espuna, A, Puigjaner, L.: Design of Multiproduct Batch Chemical Plants. Computers and Chemical Engng., 13,
pp. 163-174, 1989.
5. Espuna, A, Palou, I., Santos, 0., Puigjaner, L.: Adding Intermediate Storage to Noncontinuous Processes.
Computer Applications in Chem. Eng. (edited by H. Th. Bussmaker and p.o. Iedema), pp. 145-152, Elsevier,
Amsterdam, 1990.
6. Espuna, A, Puigjaner, L.: Incorporating Production Planning in to Batch Plant Design. Paper 82f AIChE Annual
Meeting, Washington D.C., November, 1988.
7. Flatz, W.: Equipment sizing for Multiproduct Plants. Chemical Engineering, 87, pp. 71-80.
8. Faqir, N.M Karimi, !A: Optimal Design ofBatch Plants with Single Production Routes. Ind. & Eng. Chern. Res.,
28, pp 1191, 1989a.
9. Faqir, N.M. Karimi, I.A.: Design of Multipurpose Batch Plants with Multiple Production Routes. Conference on
the Foundation of Computer Aided Process Design, Snowmass, CO., 1989b.
10. Graells, M, Espuna, A, Santos, G., Puigjaner, L.: Improved Strategy in Optimal Design of Multiproduct Batch
Plants, Computer Oriented Process Engineering (edited by L. Puigjaner and A Espuna), pp. 67-74, 1991.
II. GracUs, M, Espuna, A, Santos, G., Puigjaner, L.: Improved Strategy in the Optimal Design of Multiproduct Batch
Plants. Computer-Oriented Process Engineering (Ed.: L. Puigjaner and A Espuna). Elsevier Science Publishers
B.V., Amsterdam, pp. 67-73,1991.
12. Graells, M.: Working Paper. Universitat Politecnica de Catalunya, Barcelona, 1992.
13. Graells, M, Espuna, A., Puigjaner, L.: Modeling Framework for Scheduling and Planning of Batch Operations.
The II th International Congress of Chemical Engineering. Chemical Equipment Design and Automation.
CHISA'93, ref. 973, Praha, Czech Republic, 1993.
14. Grossmann, I., Sargent, R.W.H: Optimum Design of Multipurpose Chemical Plants. Ind. Eng. Chern. Process Des.
Dev., 18, pp. 343-348, 1979.
15. Jayakumar S., Reldaitis G. V.: Chemical Plant Layout via Graph Partitioning. I. Single level. Computers & Chemical
Engineering, 18, N. 5,pp. 441-458,1994.
16. Karimi, lA, Reldaitis, G.V.: Variability Analysis for Intermediate Storage in Noncontinuous Processes: Stochastic
Case. 1. Chern. Eng. Sym. Series, 92, pp. 79, 1985.
17. Karimi, LA., Reldaitis, G. V.: Deterministic Variability Analysis for Intermediate Storage in Noncontinuous
Processes. AIChE J., 31, pp. 1516, 1985
18. Knopf, F.C., Okos, MR., Reldaitis, G.v.: Optimal Design of Batch/Semicontinuous Processes. Ind. Eng. Chern.
Process Des. Dev., vi, pp. 76-86, 1982.
19. Kondili E., Pantelides C.C., Sargent R. W.H.: A General Algorithm for Short-Term Scheduling of Batch
Operations-I. MILP Formulation, Computers & Chemical Engineering, 17, N.2, pp. 211227,1993.
20. Modi, AX, Karimi, 1.A.: Design of Multiproduct Batch Processes with Finite Intermediate Storage. Computer
Chern. Eng., J3,pp.I27-J39,1989.
21. Papageorgaki, S., Reldaitis, G.v.: 1990.: Optimal Design of Multipurpose Batch Plants. Ind. Eng. Chern. Res., 29,
pp.2054-2062,1990.
22. Patel, AN., Mah, R.S.H, Karimi, IA: Pre1im.inary Design of Multiproduct Noncontinuous Plants Using Simulated
Annealing, AIChE meeting, Chicago, 1990.
23. Puigjaner, L.: Advances in Process Logistics and Design of Flexible Manufacturing. CSChE Conference, Toronto,
1992.
24. Puigjaner, L., Reldaitis, G. V.: Disseny I Operacio de Processos Discontinus. COMMET - CAPE. Course Notes,
Vol. I, UPC Barcelona, 1990.
25. Puigjaner, L., Espuna, A. Santos, G., Graells, M.: Batch Processing in Textile and Leather Industry In: Reldaitis,
G.V., Sunol, A.K., Rippin, D.W.T., Hortacsu o. (cds.) Batch Process Systems Engineering, NATO AS! Series
F, Springer Verlag, Berlin. This volume, p. 808.
26. Reklaitis G.V.: Design of Batch Chemical Plants under Market Uncertainty. Seminar given at Universitat
Politecnica de Catalunya, June 1994.
 113

27. Reklaitis G.V.: Chemical Plant Layout via Graph Partitioning. Seminar given at Universitat Politecnica de
Catalunya, June 1994.
28. Santos, G., EspIlll8, A, Graells, M, Puigjaner, L.: Improving the Design Strategy of Multiproduct Batch Plants with
Intermediate Storage. AIChE Annual Meeting, Florida, 1992.
29. Santos, G.: Working Paper. Universitat Politecnica de Catalunya, Barcelona, 1993.
30. Santos, G., Espuna, A, Puigjaner, L.: Recent Developments on Batch Plant Design. The 11th International
Congress of Chemical Engineering. Chemical Equipment Design and Automation. CmSA'93, ref 973 Praha, Czech
Republic, 1993.
31. Shah N., Pantelides C.C.: Optimal Long-term Campaign Planning and Design of Batch Operations. Ind. Eng. Chern.
Res., 30, pp. 2308-2321,1991.
32. Shah N., Pantelides C.C.: Design ofMultipwpose Batch Plants with Uncertain Production Requirements. Ind. Eng.
Chern. Res., 31, pp. 1325-1337, 1992.
33. Sparrow, RE., Forder, GJ., Rippin D.w.T.: The Choice of Equipment Sizes for Multiproduct Batch Plants. Ind.
Eng. Chern. Process Des. Dev., 14, pp. 197-203,1975.
34. Suhami, I., Mah, RSH: Optimal Design ofMultipwpose Batch Plants. Ind. & Eng. Chern. Proc. Des. Dev., 21 (I),
pp. 94-100, 1982.
35. Takamatsu, T., Hashimoto, I., Hasebe, S.: Optimal Design and Operation of a Batch Process with Intermediate
Storage Tanks. Ind. Eng. Chern. Process Des. Dev. 21,pp. 431,1982.
36. Vaselenak, l.A., Grossmann, I.E., Westerberg, AW.: An Embedding Formulation for Optimal Scheduling and
Design of Multipurpose Batch Plants. Ind. Eng. Chern. Res., 26, pp. 139-148, 1987.
37. Voudouris V.T., Grossmann I.E.: Mixed-Integer Linear Programming Reformulations for Batch Process Design
with Discrete Equipment Sizes. Ind Eng. Chern. Res., 31, pp. 1315-1325, 1992.
38. Wellons MC., Reldaitis G.V.: Scheduling of Multipurpose Batch Chemical Plants. I. Formation of Single-Product
Campaigns. Ind. Eng. Chern. Res., 30, pp. 671-688, 1991.
39. Wellons MC., Reklaitis G. V.: Scheduling of Multipurpose Batch Chemical Plants. 2. Multiple Product Campaign
Formation and Production Planning. Ind. Eng. Chern. Res., 30, pp. 688-705, 1991.
40. Wiede, W., Jr., Yeh, N.C., Reklaitis, G. V.: Discrete Variable Optimization Strategies for the Design of
Multi-product Processes. AIChE Annual Meeting, New Orleans, LA, 1981.
41. Yeh, N.C., Reklaitis, G. V.: Synthesis and Sizing of Batch Semicontinuous Processes. Computers and Chemical
Engng., 6, pp. 639-654, 1987.
 Predesigning a Multiproduct Batch Plant
by Mathematical Programming
D.E. Ravemark and D. W.T. Rippin

Eidgenossische Technische Hochschule, CH-8092 ZUrich, Switzerland

Abstract: This paper contains a number of MINLP formulations for the preliminary
design of a multiproduct batch plant. The inherent flexibility of a batch plant leads to
different formulations depending on which aspects we take into account. The formulati-
ons include parallel equipment in different configurations, intermediate storage, variable
production requirements, multiplant production, discrete equipment sizes and allowing
the processing time to. be a function of batch size. A task structure synthesis formu-
lation is also presented. The examples are solved with DICOPT ++ and the different
formulations are coded in GAMS. The resulting solutions (plants) have different objec-
tive functions (Costs) and structure depending on the formulation used. Solution times
vary significantly in the different formulations.

Keywords: Batch plant design, MINLP, DICOPT++, Intermediate Storage, Multiplant,

Synthesis

1 Introduction
Batch processing is becoming increasingly important in the chemical industry. Trends
away from bulk commodity products to higher added-value speciality chemicals, result
from both increased customer sophistication (market pull) and high technology chemi-
cal engineering (technology push). Batch operations are particularly suitable for spe-
ciality chemicals and similar types of complex materials because these operations can
be more readily scaled up from benchscale experimental data, designed from relatively
modest engineering information and structured to handle multiple products whose indivi-
dual production requirements are not large enough to justify construction of a dedicated
plant. If two or more products require similar processing steps and are to be produced
in low-volume, it is economical to use the same set of units to manufacture them all.
Noncontinuous plants are very attractive in such situations. Two limiting production
configurations are common, multiproduct and multipurpose. In a multipurpose plant or
jobshop, the different products follow different routes through the plant. In a multipro-
duct plant or flowshop, all products follow essentially the same path through the plant
and use the same equipment. In this work only the multiproduct plant is modelled.

2 Previous work
The design problem without scheduling considerations using a minimum capital cost de-
sign criterion was formulated by Loonkar and Robinson [15J and Robinson and Loonkar
 115

[21]. They solved it as a direct search problem to obtain the minimal capital cost. Semi-
continuous equipment was included but they did not include parallel equipment in their
formulation, and they used nonoverlapping production of batches (at anytime there is
only one batch in the plant).
Sparrow et. al. [23] assumed that the cost of semicontinuous equipment was negligible
and only considered batch equipment. They developed a heuristic method and a branch
and bound method to solve the MINLP problem. They considered the problem with
discrete sizes. The heuristic was to size the plant for ahypothetical product (a weighted
average of the products) and then sequentially add units in parallel until no improvement
was found. The heuristic obtained a continuous solution of the unit sizes and this was
rounded up to the nearest discrete size. A comparison between the two methods showed
that the branch and bound method produced better solutions (average 1%, max 12%)
than the heuristic but the heuristic was order(s) of magnitude faster in computing time.
Grossman and Sargent [9] permitted processing time to be a function of batch size
and semicontinuous equipment was not included in their formulation. They relaxed the
MINLP problem and solved it as a nonlinear programming problem. Then they rounded
the relaxed integer variables to the nearest integer and solved it again. If the gap between
the relaxed MINLP and the MINLP with fixed integer variables was small the integer
point was assumed optimal. If the gap was large they proposed that one could use a
branch and bound method and at every node solve a NLP problem. They also proposed
a reformulation of the original relaxed MINLP to a geometric program.
Takamatsu et.al. [25J dealt with the optimal design of a single product batch process
with intermediate storage tanks. The major limitation of their work was that it dealt
with a single product process only.
Suhami and Mah [24J formulated the optimal design problem of a multipurpose batch
plant as a MINLP. The production scheduling is done by heuristics and a strategy to
generate the nonredundant horizon constraint was presented.
Knopf et.al. [12J solved the problem given by Robinson and Loonkar in nonoverlapping
and overlapping operation as a NLP. They proposed a logarithmic transformation and
showed that this convex NLP reduced the CPU time by 300-400% compared with the
original NLP.
Rippin [20J reviewed the general structure of batch processing problems and did a
classification of batch processing problems.
Vaselenak et.al. [26J formulated the problem of a retrofit design of multiproduct batch
plants as an MINLP and solved it with an outer approximation algorithm. In order to
circumvent a non convex objective function they replaced the function with a piecewise
linear underapproximator.
Yeh and Reklatis [28J proposed the partitioning of the design problem into two parts:
the synthesis heuristic problem and the equipment sizing subproblem. Their heuristic
procedure for sizing yielded near optimal solutions but was applicable only to the single
product problem. Their sizing procedure did not include sizing and costing of interme-
diate storage since they assumed that the cost of storage was negligible. Their synthesis
heuristic include splitting and merging of tasks and adding parallel equipment and sto-
rage tanks. The synthesis heuristic is partitioned in a train design heuristic and storage
heuristic. The train design heuristic is a sequential search which starts from maximum
 116

merging with no parallel equipment and tries to improve the train by adding parallel
equipment and splitting of tasks. The storage heuristic is to solve the train for no storage
and for max storage and add storage to the no storage train until the difference between
max storage was small.
Espuna et.al. (7) formulated the MINLP problem with parallel units in and out of
phase. They solved the NLP sizing subproblem with a gradient search which included
heuristic rules for faster convergence. The required production time for a point is evaluated
and the difference between this time and the horizon time is called surplus time. Positive
surplus time means that the plant is oversized (but feasible) and some unit sizes should
be reduced. Negative surplus time means an infeasible plant and the unit sizes have to be
increased. The discrete optimization is done by sequentially adding parallel equipment
out of phase at the most time limiting stage. If a unit is close to the upper or lower
bound, the option exists to add or delete a unit in phase.
Modi and Karimi [16) developed a heuristic procedure for the preliminary design of
batch processes with and without intermediate storage. In their procedure a sequence of
single variable line searches was carried out, yielding good results with small computatio-
nal effort. However, in their work the location of storage units was fixed.
Birewar and Grossman, [3) considered synthesis, sizing and scheduling of a multi-
product plant together. Their formulation contained a number of logical constraints to
control the selection of units for tasks. They solved their nonconvex MINLP formulation
with DICOPT ++ but for some of their examples they did not obtain the global optimum.
Patel et.al. [18) used simulated annealing to solve the MINLP. They formulated the
problem with intermediate storage tanks and parallel equipment in phase, allowing parallel
units in phase to be of unequal size. They allow products to be produced in parallel paths
that are operated out of phase.
Salomone and Iribarren [22) present a formalized procedure to obtain size factors and
processing time.

3 The design problem with parallel units out-of-

phase
The design problem has been formulated with a nonlinear objective function involving
capital cost of the batch equipment by a number of authors, some also included semi con-
tinuous equipment. We will only consider batch equipment in our formulations.
The problem is to minimize the objective function by choice of M j parallel units, unit
size ltj.

J
Cost min LMjaj(ltjt' (1 )
1
ltj ~ BjS jJ (2)
T·
LCT; > ~ (3)
Mj
 117

H EQ;LCTi
~ (4)
i=1 B;
V!- ~ v.·<V u (5)
J J - J

The goal of predesign of multiple product plants is to optimize the sizing of manufacturing
units by minimizing the total capital cost of the units (equation 1). The capital cost of
a unit is a simple power function of its size a;(Vj)Qj, where aj is the cost factor, Vj is
the size of unit j and OJ is the cost exponent (less th~ unity). Equation (2) is the unit
size constraint. Bi is the final amount of product i in a batch. Si,j is the size factor, it
is the relation between the actual size of a batch of product i in stage j and the final
batch size B i • The unit is sized to accommodate the largest batch processed in that unit.
Equation (3) is the limiting cycle time constraint. The processing time for product i in
stage j is Ti,j. The number of parallel equipment items out of phase Mj increases the
frequency with which a stage can perform a task and this reduces the stage cycle time.
The limiting cycle time LCTi for product i is the longest of the stage processing times.
Equation (4) is the horizon constraint. Qi is the production requirement of product i.
Q;/ Bi is the number of batches of product i and LCTi is the time between batches. The
time to produce all the batches of all the products has to be smaller than or equal to the
time horizon H. The size of units in stage j is usually bounded (equation 5) by an upper
bound V;u and lower bound V;L

4 Logarithmic transformation
The design problem formulation in chapter 3 is non convex as noted by Kocis and Gross-
man [13]. The horizon constraint and objective function are nonlinear and nonconvex and
using the Outer Approximation/ Equality Relaxation algorithm [6, 14] global optimality
cannot be guaranteed. Through logarithmic transformation, the formulation can be mo-
delled as a convex MINLP problem. We define the new (natural) logarithmic transformed
variables, InVj = In(Vj), InB; = In(Bi ), InLCT; = In(LCT;) and InMj = In(Mj ), thus
"In" in front of a variable name sh~ws that the variable expresses the (natural) logarithmic
value of the variable. Using these new variables the formulation is now:

J
Cost =min Laj exp(lnMj + ojInVj) (6)
j=1
InVj ~ InBi + InSi,j (7)
InLCT; > InT;,j -InMj (8)
N,.
H > LQiexp(lnLCTi -inBi) (9)
i=1
InVL
J ~ InVj ~ InV;u (10)
Now all the nonconvexities have been eliminated and the nonlinearities in this model
appear only in the objective function (6) and the horizon time constraint (9), and in both
equations the exponential term is convex. Hence when we use the OA/ER algorithm we
are guaranteed to find the global optimum.
 118

o Integer variables by binary expansion

We use DICOPT++ based on the OA/ER described by [27] to solve the problem formu-
lations, and as DICOPT++ cannot accept integer variables, the number of parallel units
Mj has to be formulated in terms of binary variables [8].

MU
J
InMj L In(k) Yk,j (11)
k=1
MU
J
1 = LYkJ (12)
k=1
If binary variable YkJ is equal to 1 then the logarithmic number of parallel equipment
InMj is equal to In(k). Equation (12) ensures that a stage is only assigned one of the
possible number (1,2, ... ,k) of parallel units.

6 MINLP-Formulations
6.1 Parallel units out-of-phase
The problem formulation presented by Kocis and Grossman [13] consists of equations
(6-12).

6.2 Parallel units in and out of phase

We add parallel units out of phase if the stage is time limiting and if the stage is capacity
limiting we can add parallel units to operate in phase. This increases the largest batch
size that can be processed on a stage. The batch from the previous stage is split and
assigned to all the units in phase on that stage. Upon completion they are recombined
and transferred to the next stage. This does not affect the limiting cycle time but the
batch size of a stage is multiplied by the number of in phase units in parallel since we
always add equipment of the' same size. The formulation is now.

J
Cost min Lajexp(lnMj + InNj + G:jlnYj) (13)
j=1

InYj + lnNj > InB; + InS;,; (14)

NU
J
Lln(c)Zc,j (15)
c=l
NU
J

LZcJ (16)
c=l
and equations (8 12)
InMj is the number of parallel units operating out of phase and InNj is the number
of parallel units operating in phase. The binary variable YkJ is equal to one if stage j
 119

has k= My
1,2, ... , parallel units operating out of phase. My
is the upper bound on
parallel units operating out of phase. The binary variable ZeJ is equal to one if stage j
Ny
has c = 1,2, ... , parallel units operating in phase. Ny
is the upper bound on parallel
units operating in phase. In constraint (14) the number of parallel units in phase is
included. This reduces the size of units needed in stage j. The formulation is still convex
but it contains twice as many binary variables as formulation 6.1, which will increase the
solution time. Many of the possible configurations may not be advantageous and they can
be discarded in advance. For example, 4 units in phase and 4 units out of phase means
16 units at that stage. A constraint allowing, for example, only four units in a stage can
be included.
MjNj ::; 4 (17)
constraint (17) expressed in logarithmic variables

(18)

Constraint (18) is linear in the logarithmic variables and can be included in the formulation
above without increasing the number of nonlinear constraints.

6.3 Unequal sizes of parallel equipment in phase

Parallel equipment is usually assumed to be of equal size. If we only allow equal sizes of
parallel equipment operating out of phase we have only one batch size B; for each product.
If equipment out of phase were allowed to be nonidentical we would have several different
batch sizes for the same product. This would complicate the formulation and it is not
obvious that this would lead to an improvement. Nonidentical equipment in phase may
lead to a improvement of the objective function. Batches are split into the two units in
phase and then recombined when finished on that stage and we do not have to split the
batch 50/50. Due to the economy of scale, it is cheaper to have one large and one small
unit than two equal size units with the same total capacity.
The formulation below is for a maximum of two units in phase. It can be expanded
to more units but for clarity we restrict it to two units.

J
Cost min I)j exp( InMj + Cl:jln V;1)
;=1
J
+ IJIi )"ia; exp(lnMj + Cl:jln V;u) (19)
;=1
InV;U + InNj > InB; + InS;,; (20)
I; > exp(lnN;) - 1 (21 )
InVI
J ~ InN; + In V;U - U(1 - Z!j} (22)
InV;1 > InVJu - U(1 - Z2 ,J.) (23)
2
1 = L:ZeJ (24)
=1
(InV;L -lnV;U) ::; InNj ::; In(2) (25)
 120

o $ 1; $ Z2,i (26)
and equations (8 12)

First it must be said that this formulation is only convex for a given vector Mj of
parallel units in phase. If the number of units out of phase is included in the optimization
the objective function (19) contains bilinear terms. In the size constraint (20) the number
of parallel equipment items in phase InNj is a continuous variable bounded by (25) and
the actual capacity of the parallel units operating in _phase on a stage is InVF + InNj .
InVi is the volume of the first unit in parallel. If we only have one unit in phase, equation
(22) assigns to the unit the size In Viu + InNj . Here InNj is the logarithmic fraction of
the upper bound unit e.g. the volume of the first unit is expressed as a fraction of Viu .
If we have two units in parallel the equation (23) ensures that the first unit assumes the
value InViu. This is from the simple heuristic (economy of scale) it is cheaper to have
one large (at upper bound size) and one smaller unit if we have two units in phase. Ij is
the fraction of the size of the second unit in phase compared to the upper bound in size.
1; is equal to zero if we have no parallel units in phase and 0 $ 1; $ 1 if we have two
units in parallel (26). In the objective function the cost of the second unit is (fj)a; times
the cost of a unit at the upper bound. The formulation may not give an optimal solution
if we have a lower bound on size on the second unit since we always assume that if the
second unit exists the first unit is at an upper bound. The size of the second unit Vi2 can
be calculated by: Vi2 = 1;Viu.

6.4 Flexible use of parallel units in and out of phase

We are producing a number of different products in a plant, all with different batch
sizes and requirements for size and processing time. A stage can be size limiting for one
product and time limiting for another and this leads to the possibility of changing the
configuration of equipment for different products. We have equipment working in phase
for some products, if the stage is size limiting for these products, and out of phase for
others,-for which the stage is time limiting. We call this flexible use of parallel equipment.
This leads to a formulation with a large number of binary variables.
J
Cost min I)jexp(lnTotMj + Qj/nVj) (27)
;=1
In Vj + InNi,; > InBi + InSi,; (28)
InLCT; > InTiJ - InMiJ (29)
Np
H > L Qi exp(lnLCT; - lnBi) (30)
1
MU
J
InMi,j L ln( k) l'I.,i,i
= 1:=1 (31)
NU
J
InNi,; = 2: In( C)Zc,i,i (32)
=1
 121

MU
I:Y
J

1 ..
k ,1..7 (33)
k=1
NU
J

I:Zc,iJ (34)
=1
In Tot M j = lnM;J + lnN;J (35)
lnMj + lnNj ~ In(4) (36)
lnV.L
)
lnV; ~ ln~'F
~ (37)
and equations (8-12) and (14-16)
M;,j is the number of units out of phase at stage j for product i and N;J is the number
of units in phase at stage j for product i. The product of these two is equal to the total
number of units at a stage TotMj and in logarithmic variables it is lnTotMj (equation 35).
Constraint (35) ensures that the total number of parallel equipment at a stage is equal for
all products. The binary variable Yk,iJ is equal to one if for product i stage j has k units
in parallel out of phase. The binary variable Yc,iJ is equal to one if for product i stage j
has c units in parallel in phase. Constraint (36) sets the upper bound on the number of
units in and out of phase on a stage. This formulation contains a large number of binary
variables.

6.5 Intermediate storage

6.5.1 The role of intermediate storage
The equipment utilization in multiproduct batch plants is in many cases relatively low. By
increasing equipment utilization, it is possible to increase the efficiency and profitability
of a batch plant. In a perfect plant the stage cycle time of each product for all stages
should be equal to its LeT;. This cannot be achieved in practice due to the difference in
processing time at each stage, the best that can be done is to try and make the difference
in stage processing time as small as possible. One way of decreasing stage idle time is to
add parallel units, another is to insert intermediate storage between stages. The insertion
of a storage tank causes the process to be divided into two subprocesses and decouples
the operations upstream and downstream of the storage tank. This in turn allows the
LeT and the batch sizes on either side to be chosen independently of each other. The
installation of a tank would increase the cost of the plant, but due to the decoupling,
subprocesses with smaller LeT or larger limiting batch sizes are created, either of which
can lead to smaller equipment sizes. It is possible to install storage at appropriate locations
in the train which may result in a net reduction of the overall plant cost. Apart from
the financial advantage outlined above, the insertion of storage tanks at suitable positions
in a batch process yields a number of other benefits as stated by Karimi and Reklaitis
[10J. These include an increase in plant availability, a dampening of the effects of process
fluctuations and increased flexibility in sequencing and scheduling. They also pointed out
that there are a number of drawbacks that are difficult to quantify including the inventory
cost of the material stored, maintenance and clean out costs, spare parts costs, labour and
supervision costs. The disadvantage of including intermediate storage tanks include the
increased likelihood of material contamination, safety hazards, operator errors, processing
 122

delay or the requirement for an expensive holding operation such as refrigeration. For
a given plant there will be multiple possible locations for an intermediate storage tank.
This tank is inserted only to separate the up and down stream batchsizes and limiting
cycle times.

6.5.2 Sizing of storage tank

In order to be able to properly assess the cost effects due to storage tanks, we should also
include the cost of storage in the objective function. This is made possible by the work of
Karimi and Reklaitis [10). Karimi and Reklaitis developed useful analytical expressions
for the calculation of the minimum storage size required when decoupling two stages of
operation for a single product process. The major difficulty is that the exact expression for
the storage size is a discontinuous function of process parameters. This makes it impossible
to use the exact expression in an optimization formulation as functional discontinuities
can create problems for most of the optimization algorithms. Karimi and Reklaitis also
developed a simple continuous expression which gives a very good approximation to the
actual size. They show that for a 2-stage system with identical, parallel units operating
out of phase in each stage, as in our case, the following equation gives a very close upper
bound for the required storage size for the decoupling of subtrains:

_
VS, - mF {. ( [ _ O;,up ] [ _ O;,down ] )
S", Bup 1 LCT;,up + Bdown 1 LCT;,down
} (38)

where the storage size is determined by the requirement of the largest product. The
O;,up and O;,down refer to the up and down stream batch transfer times but since it
is not part of our model to model the sernicontinuous equipment we use the following
simplification (39) for evaluating the size of a storage tank. Our alternative equation (39)
is linear in normal variables.

VS. 2: S;,.(B;,up + B;,down) (39)

Modi and Karimi [16) also used these equations (38), (39) for storage size. With logarith-
mic variables and using the binary variable to show if a storage tank is located in site j
(for storage) between stage j and stage j + 1 we obtain

The terms U(1 - Xj,q) and U(1 - Xj+l,q+l) are to ensure that VSj (=volume storage
between stage j and j + 1) is given a size only if both Xj,q and X j+1,q+1 are equal to 1
e.g. that unit j belongs to subtrain q and the next unit j + 1 belongs to the next subtrain
q + 1. This change of subtrain is caused by a storage tank used to decouple trains. SFSi,j
is the size factor for storage, for product i and location j (which is between unit j and
unit j + 1). This equation is nonlinear but convex so we can use it in our model and
the resulting optimum is guaranteed to be the global optimum. This equation will add
additional nonlinear equations so we can try and simplify the problem. Since the storage
tank, when it is inserted, will probably need to be bigger than the smallest possible, we
 123

can use twice the larger of the batches instead of the sum of up and down stream batches.
The equations are now linear in the logarithmic variables:

U(1-Xj ,q)+U(1-Xj +l,q+l)+lnVSj ~ In(SFSiJ) +lnBi,q +In(2) (41)

U(1 - Xj,q) + U(1- X j+l,q+l) + InVSj ~ In(SFSiJ ) + InBi,q+l + In(2) (42)

(41) provides for storage at least double the size of the downstream batch, (42) provides
for storage at least double the size of the upstream batch and the storage will of course
assume the larger value of these two.

6.5.3 Cost of storage

The cost function for the storage tank is the standard exponential cost function.
M-l
E bj(VSj)"YJ (43)
j=1

Where bj is the cost factor, /j is the cost exponent and the costs of the tanks are
summed over the M-1 possible locations. If we are using equations (41) and (42) for
the sizing of storage we get the logarithmic size of the storage and use equation (44) for
costing.
M-l
E bj eXPbjlnVSj ) (44)
j=1

Alternatively it may be argued that when storage is inserted it will have a size large
enough to accommodate all possible batch sizes and may also absorb a little drift in the
productivity of the two trains. Then this sizing equation is not necessary at all. We just
add a fixed penalty for every storage tank inserted. For example in the form below where
XM,q is the binary variable identifying the subtrain to which the last unit ("unit M")
belongs, the subtrain index q minus one is equal to the number of storage tanks in the
optimal solution. We can just add the total cost of storage to the cost of the equipment
items.
M-l
Total cost of storage = Cost * E (q -1)XM ,q (45)
q=1

6.5.4 MINLP storage formulation

The formulation uses equation (38) for the sizing of the storage tank and equation (43)
is added to the objective function.

J M-l
Cost = min Laj exp(lnMj + ojlnV;) + L bj(VSj)'"'i (46)
j=1 j=1
InV; > InBi,q + InSi,j - U(1 - Xj,q) (47)
InLCT;,q > InT;J - InMj - U(1 - Xj,q) (48)
InPRO; lnLCT;,q - InB;,q (49)
 124

Np
H ~ L Qi exp(lnP ROi) (50)
;=1
U(1 - Xj,q) + U(l - X i +1.q+1) + VS. ~ SiJ(exp(Bi,q) + exp(B;,q+1)) (51)
J
1 = LXi,q
q=1
(52)

Xi,q ~ X;+1,q q=l (53)

Xi,q ~ . Xj+l,q+1 q=j (54)
Xi,q + Xj,q+1 ~ Xi+1,q+1 (55)
and equations (10 -12)
If Ii. storage tank is inserted the different subtrains can operate with different batch sizes
and limiting cycle times but the productivity of both up and downstream trains must be
the same. This is ensured by constraint (49). We use this productivity in the horizon
constraint (50). Units are sized (47) only to accommodate the batch size of the sub train
that the unit belongs to. The limiting cycle time (48) of a train is the longest stage
processing time of the stages that belong to the subtrain. Constraint (52) ensures that
every unit is assigned to one and only one subtrain. Constraint (53) ensures that if a
unit belongs to the first subtrain, the previous unit belongs to the first subtrain too.
Constraint (54) ensures that if unit j + 1 belongs to subtrain q + 1 (j = q) then the
previous unit j belongs to subtrain q (j = q). Constraint (55) ensures that if unit j + 1
belongs to subtrain q + 1 the previous unit j either belongs to the same subtrain q + 1
or the previous subtrain q. Constraint (55) ensures that infeasible train structures are
prevented.
If in the problem a location for storage is not allowed (for example between unit j'
and j' + 1) we just have to add a constraint
Xi',q = X;'+1,q (56)
This forces unit j' + 1 to belong to the same sub train as j' and the location of a storage
tank by the solution algorithm is inhibited at this location. We can also force the location
of a storage tank between unit j' and j' + 1 by adding a constraint
(57)
This forces unit j' + 1 to belong to the next subtrain and the location of a storage tank
by the solution algorithm is forced at this location.

6.6 Variable production requirement

In the preliminary design of a plant the production requirement Qi is likely to be an
estimate which might be allowed to move between bounds Qf 5 Qi 5 Qf. Therefore,
the production requirement is a possible variable for optimization purposes. It will help
us to direct our production facilities towards the more profitable products. We add the
following equation in the objective function:
I
Ec;(Qfef - Q;) (58)
;=1
 125

(note: large production gives a negative profit but since we are comparing the profit
with the capital cost of the plant this is correct. We just have to add the capital cost and
the profit from equation (58).)
The change in profit on sales associated with over or under production compared with
the nominal value may be offset by changes in capital cost charges incurred for a larger
or smaller plant. This equation is linear, for an increase in production the profit is always
the same. There are two reMons why this linear function may not be good. First, if the
range of product requirements is large the economy of scale leads to the trivial solution
that the optimal plant is the largest possible (cost of plant increases exponentially with
a factor of '" 0.6 and profit increases linearly). Second, for the special chemicals that are
produced in a multiproduct plant it is probably not true that the marginal value of an
extra kg product stays the same. Increasing the production may lead to a decrease in the
marginal value of an extra kg product.
Instead we use the function:

1
~e; Qref}
({ -'-, -1 ) (59)
,=1 Q,
(note: parameter e; is not the same as in equation (58»
This function is sensitive to large changes in production requirement. It adds a incre-
asing penalty cost if Qi < Qfef and a decreasing profit as Qi increases above Qfef. The
marginal value of an extra kg around Qfef is almost constant but if we produce twice as
much as Qfef the marginal value of an extra kg is is only half of that. Likewise, if we
produce only half of Qref the marginal value of an extra kg of product is twice as large.
We do not want the plant to produce less than we can sell.
The function is also convex when it is included in a formulation with logarithmic
variables. With logarithmic variables, including equation (59) in the objective function
M 1
Cost = min La; exp(1nM; + o;lnV;) + L e;(exp(1nQfef -lnQ;) - 1) (60)
;=1 ;=1
We have to change the horizon constraint and include the variable InQi This constraint
is also convex.
1
H ~ L exp(1nQi + InLCT; - InB;) (61)
i=l
Now we can replace the equations in the formulation for parallel units out of phase
(chapter 6.1). We get:

M
Cost = min La; exp(1nM; + 0; In V;)
;=1
1
+ Le;(exp(lnQfef -lnQ;) -1) (62)
;=1
1
H ~ Lexp(lnQ; + InLCT; -lnB;) (63)
;=1
and equations(7 - 8) and (10 -12)
 126

6.7 MUltiplant production

Dividing the multiproduct plant into several multiproduct plants, each producing a fixed
subset of products, can be of advantage in increasing equipment utilization, thus decrea-
sing the needed sizes of equipment and also reducing the long term storage costs because
we produce each product for a longer period of time.

6.7.1 Cost of long term storage

We assume a constant demand over the time horizon. Ti i " is the time that we produce
product i on plant p. The fraction of the total time that we do not produce product i is
(1 - Tii.,/ H). We assume that we have a constant demand over the time horizon and we
have to store material to satisfy demand for this time. The average stored amount is a
function of the production requirement and the time that we do not produce a product.
The cost of long term storage is expressed in equation (64) and is included in the objective
function.

(64)

where
Ti. _ Q.LCT;.,
I., - I B." (65)

This equation has been proposed by Klossner and Rippin [11). PCi is the production
cost, taken as the product value for purpose of inventory, and Ii is a weight factor for
production cost which may include a discount factor for charges on inventory. Qi is the
production requirement. Ti i ., is the time plant p is dedicated to producing product i.

6.7.2 MINLP multiplant formulation

We define a binary variable Xi., = 1 if product i is produced in plant p, otherwise O.
Parallel equipment (binary variable Y"J,,) can be excluded in the interests of a simpler
problem. Binary variable Y",j" = 1 if there are k parallel units at stage j at plant p. The
formulation without cost of long term storage:

P J
Cost min EEaj exp(lnMj" + QjlnV;,,) (66)
,,=lj=1
InV;,,, ~ InB·',p + InS·,,). - U(l - X-ItI' ) (67)
InLCTi,,, > InT.·',1. -lnM·J,P - U(l - X-t,p ) (68)
Ti i,,, = Qi exp(1nLCT;,,, - InBi,,,) (69)
PT;,,, ~ Tii,,, - H(Xi,,,) (70)
Np
H ~ E{PTi,,} (71)
i=1
MY,p
InMj,,, = E In(k)Y",j,,, (72)
"=1
 127

My'p
1 L:YkJ
k=l
,1' (73)
p
1 = L:Xi ,1' (74)
1'=1
lnV.L
J
::; lnV; ::; lnVF (75)

Constraint (74) ensures that every product is produced in one and only one plant. YkJ,p
are the binary variables for parallel units out of phase, and Xi,p are the binary variables
for the plant allocation. YkJ,l' = 1 means that unit j in plant p has k number of parallel
units. Constraint (70) ensures that if a product is not processed in a plant the total
processing time for the product in this plant is equal to the time horizon H in order to
make the storage cost equal to zero in this plant. The horizon constraint (71) is that the
production time for all units produced in a plant have to be less than the time horizon
and if a product is not produced in a plant the Ti i ,1' is equal to H and this has to be
subtracted.
The above formulation is a set partitioning problem. A set covering problem would
allow each product to be produced in more than one plant. This can be realized by
removing constraint (74) or replacing it by:

(76)

This constraint allows any product to be produced in, at most, two plants. Constraint
(74) is not sufficient to avoid redundant solutions. A more precise constraint is needed
to ensure that the solution of the MILP master problem does not produce redundant
solutions. For example if we have two products, the solutions
• product 1 in plant 1 and product 2 in plant 2

• product 1 in plant 2 and product 2 in plant 1

are equal since a plant is defined by the products produced in it. To date, the proper
formulation of such a constraint has not been found.

6.B Discrete equipment sizes - With parallel equipment

In the original problem formulation the optimal sizes of the equipment items in our plant
are chosen from a continuous range, subject to upper and lower limits reflecting technical
feasibility. Often the sizes of equipment items are not available in a continuous range but
rather in a set of known standard sizes available from a manufacturer at a known price.
A standard unit larger than required is likely to cost less than a unit specially built to the
exact size. Producing equipment in special sizes will probably not be economical. The
choice of the most appropriate equipment size is now a discrete decision. This forces us
to add more binary decision variables to choose a defined size, but the solution of the
problem is more accurate than in the continuous case as we know the cost of the item
and we do not have to use some approximate cost function. Allowing parallel units out
of phase and several products gives us the formulation
 128

J
Cost = min E exp(lnCostj) (77)
j=1
G
lnV; :::; E In V sizej,gXj,g (78)
9=1
G
InCost j ~ lnMj + E In V costj,gXj,g (79)
g=1
G
= EX j,9 (80)
g=1
and equations (7-12)

InCostj is the logarithmic cost of stage j with InMj parallel units of size In V sizej out
of phase. In V sizej,g is a set of logarithmic standard sizes for units capable of performing
tasks in stage j. InCostj,9 is the logarithmic cost of standard units InVsizej,g.

6.9 Discrete equipment sizes - without parallel units

6.9.1 Single product
If we only produce one product the horizon constraint is

Q LCT = H or B = Q LCT (81)

B H
If the production requirements, the time horizon and limiting cycletime (no or fixed
number of parallel units) are known and constant the required batch size can be calculated
directly. If we know the batch size we get the required volume in stage j by V; = BSj
and we just have to round up to the nearest discrete size V sizej,9

6.9.2 Multiproduct plant, long single product campaigns, MINLP model

When no parallel equipment is allowed or the structure of parallel units is fixed, we can
formulate the problem in normal (not logarithmic) variables, then the limiting cycle time
(LCT;) is known as it is simply the largest processing times in all stages. With normal
variables.
J G
Cost = min EEV costj,9XjolJ (82)
j=lg=1
G
V; :::; E V sizej,gXj,g (83)
g=1
V; > BiSiJ (84)
H > EQi LCTi (85)
i=1 Bi
G
EXj,g (86)
9=1
 129

Without parallel units the formulation is only nonlinear in the horizon constraint (85),
and we can reformulate the MINLP problem as a MILP problem.

6.9.3 Multiproduct plant, long single product campaigns, MILP model

For more than one product, in long single product campaigns with no parallel units (LCT,
is known), we reformulate the problem by transformation of the variables: lIB, invB, =
to give:
J G
Cost minEEVcostj,gXj,g (87)
;=lg=1
invBi ~
Si,j
t
.1/=1
x.j,g
V SlZt!j,g
(88)

I
H :::; E QiLCTiinvBi (89)
i=1
G
1 = EXj,g (90)
.1/=1

With no parallel equipment or fixed equipment structure and the inverse transforma-
tion, the formulation is a MILP.

6.9.4 Multiproduct plant-Multi product campaigns

For all our previous problems we have assumed that products are produced in long single
product campaigns, but the incorporation of scheduling in the design can be advantageous
as shown by Birewar and Grossman [3J.
Birewar and Grossman have in a number of papers [lJ,[4J and [2J developed linear
constraints for the scheduling in design problem. We use their formulation and get:

J G
Cost = min EEV costj,.l/Xj,9 (91)
j=I.1/=1
ni QiinvBi (92)
Np
ni ENPRSi,k (93)
k=1
Np
nk = ENPRSi,k (94)
i=l
Np Np Np
H ~ E {niTi,j} + E E NPRSi,kSLi,k,j (95)
i=1 i=1 k=1
ni = nk (96)
NPRSi,k ~ 0 (97)
and equations (88), (90)
NPRSi,k is the number of times that a batch of product i is followed by a batch of
product k (i = 1,2, ... , Np; k = 1,2, ... , Np). Np is the number of products. SLi,k,j
 130

is the minimum idle time between product i and product k on unit j, in [lJ there is a
systematic procedure to calculate the slack times. ni is the number of batches of product
i. nk is the number of of batches of product k. The NPRSi,k should have integer values
but since we have the number of batches ni as a continuous value NPRSi,k is probably
not integer. Birewar and Grossman report that this problem has a zero integer gap and
therefore a branch and bound search would be effective but, a simple rounding scheme
would probably produce sufficiently good solutions.

6.9.5 Continuous sizes - Multiproduct campaigns

If the discrete size constraint is relaxed we get a NLP problem with linear constraints:
J
Cost min I>i(invVit"i (98)
i=l
invBi
invV; < (99)
SiJ
and equations(92) and (93-97)

To expand this convex NLP problem to a convex MINLP problem which includes
parallel units is not straight forward.

6.10 Processing time as a function of batchsize

In previous problems the processing time is always assumed constant but this is frequently
not true. The processing time is normally dependent on the batch size. We can assume
that scaling a batch of product until it becomes twice as large. will certainly lengthen the
processing time. The change in processing time is dependent on which tasks are executed
in the stage. This has previously been noted by other authors, Grossman and Sargent
[9] and Yeh and Reklaits [28], and they usually make processing time a function of the
batchsize in the form of equation (100) where R;,i' Pi,i and Ai are constants.

(100)

This will increase the number of nonlinear equations if we model with logarithmic
variables but we could take another model that is linear in logarithmic variables.

(101)
Ri,i and Ai are constants. With logarithmic variables this equation is linear and can
be added to any formulation, thus allowing for variable processing times without adding
nonlinear equations and discrete variables.

InT;,i = InR;,i + AilnBi (102)

6.11 U nderutilization of equipment

Equipment is designed for a fixed capacity to handle the largest product batch size. As a
result the smallest product batch size using that unit can be far below its design capacity,
 131

as noted by Coulman [5]. For example a jacketed stirred tank would have greatly reduced
mixing and heat transfer capabilities when it is less than half full. Coulman proposed the
constraint
(103)
to avoid solutions with large underutilization's for some products. <pj represents the lowest
permitted level of a unit 0 < <pj :::; 1. <pj = 1 means that we do not allow the unit to
process batches smaller than.J.he unit size.
This constraint is sufficient if <pj :::; 0.5 (e.g. we do not allow units to be less than half
full). If 0.5 < <pj :::; 0.75 and we have two units on a stage a second constraint must be
added
(104)
with the binary variable Z2,; equal to one if on stage j we have two units in parallel
in phase. Constraint (104) assumes that we always use both parallel units in phase to
process batches. But, we can always use just one unit when we process small batches.
Then if we use just one unit the batch size has to be smaller than the unit size.

(105)

This constraint is infeasible for the products that use two units, so we define a binary
variable X;,;,VI that is equal to one if product i in stage j uses w units for processing. The
constraints are now with 0.0 < <pj ~ 0.875 and a minimum of three units in phase on a.
stage.

<pj Vi :::; B;S;,; (106)

v: 2:
] B;S;,j - U(l - X;,j,l) (107)
<pj2 Vi :::; B;S;,; + U(l - X;,j,2) (108)
2Vi > B;S;,j - U(l - X i ,i,2) (109)
<pj3Vi :::; B;S;.j + U(l - Xi,j,J) (110)
1 LX- 1 ,1,W (111)

Z3,j + Z2,i 2: X;.i.t (112)

Z3,i + Z2,j 2: X i ,i,2 (113)
ZJ,i 2: X i ,j,3 (114)

In figure 1 we will try to explain the underutilization constraint. On the y·axis we have
the batch size in a unit and on the x-axis we have the total batch size of all units in phase.
Equation (106) ensures that batches that only use one unit fulfil our underutilization
formulation and this equation will always hold irrespective of how many units we actually
have on the stage. Only equation (106) is needed when 0.0 < V'i :::; 0.5 and we do not
need any binary variables. This can be seen in the upper graph in figure 1 where the one
constraint is shown. Equation (107) ensures that if product i on stage j uses only one
unit, the batch size on that stage is smaller than the unit size. Equation (108) ensures
 132

'Pj ~ 0.5
No discrete variables Xi,;,w nee-
ded

0.5 < 'Pi ~ 0.75 and N j ~ 2

We have two production ~ win-
dows" and we need the discrete
variables Xi,;,l and X i ,i.2

0.75 < 'Pj ~ 0.875 and N j ~ 3

We have three production "win-
dows" and we need the discrete
variables X'J,h X'J,2 and X i ,j,3

Figure 1: Underutilization constraints.

that if product i on stage j uses two units, the batch size on that stage is larger than the
capacity of two units times the underutilization factor 'Pj. These two constraints together
with constraint (106) are shown in the middle graph.
Equations (106-108) are needed only for stages j where 0.5 < 'Pj ~ 0.75 and we have
the possibility to have two units in phase.
Equation (109) ensures that if product i on stage j uses two units, the batch size on
that stage is smaller than twice the unit size. Equation (110) ensures that if product i
on stage j uses three units, the batch size on that stage is larger than the capacity of
three units times the underutiliz'ation factor 'Pj. All constraints (106-110) are shown in
the lower graph.
Logical constraint (Ill) ensures that a batch at a stage uses only one of the options:
one, two or three units for processing. Logical constraints (112), (113) and (114) ensure
that only if parallel units exist is it possible to process in them. Zo,i is the binary variable
for parallel units out of phase and Zo,j = 1 if stage j has c parallel units in phase. The
binary variables Xi,i,w add a number of discrete variables to any formulation and we should
try to minimize the number of discrete variables. For stages j where 'Pj ~ 0.5 all binary
=
variables Xi';,,,, 0 and for stages j where 0:5 < 'Pj ~ 0.75 binary variable X i ,i,3 o.=
The constraints (106-110) in logarithmic variables (the logical constraints (111-114)
stays the same).

In'Pi + InVj < InB, + InS,,; (115)

InVj > InB; + InS,.; - U(l - X'.;.d (116)
In'Pj + In(2) + In Vj ~ InBi + InS,.; + U(l - X,.;,2) (117)
In(2) + InVj ~ InB, + InS,.; - U(l - X'';.2) (118)
 133

In<pj + In(3) + In Vi ~ InBi + InSi'; + U(1 - Xi';.3) (119)

If the underutilization constraints differ from product to product on a stage we can
replace <pj with <Pi'; in the constraint above.
All these constraints are linear in the binary variables.

7 DICOPT++ Example solutions

7.1 Infeasible NLP subproblem - Slack variable
DICOPT ++ will first solve a relaxed NLP problem and from this solution relax equality
constraints and linearize the nonlinear functions by a first order Taylor approximation to
obtain a MILP master problem. The master problem is then solved to obtain a integer
point wbre a NLP subproblem is solved. The relaxed starting point may not give good
linear approximations of nonlinear equations for problems with a poor continuous rela-
xation. The nonlinear equations are linearized far from the integer point that the master
problem is trying to predict and the result may be that the master problem predicts
integer points that are infeasible.
In a DICOPT++ iteration, no linearizations of nonlinear equations are made if the
NLP subproblem is infeasible and only a integer cut is added to the master problem. The
master problem has now, from the same bad linearization point, to predict a new integer
point and this point is probably also infeasible. These iterations continue until the integer
cuts forces the master problem into a feasible area where the problem can be relinearized.
In order to speed up the iteration and avoid solving the same master problem several
times we add a slack variable. Having an infeasible integer point means that the proposed
plant structure cannot fulfil the production requirement even with units at the upper
bound. We can make sure that the problem is feasible for all possible integer points by
adding a positive slack variable to the horizon constraint and including this slack in the
objective function with a large penalty U.
The horizon constraint
Np
H + SLICK 2: LQiexp(lnLCTi -lnBi ) (120)
1

and the objective function

J
Cost = min Laj exp(lnMj + a-;ln Vi) + U * SVCK (121)
j=1

Now a feasible integer point will be optimized as usual and an infeasible integer point
will optimize the plant to minimize the SLICK, which is the extra time over the time
horizon needed for fulfilling the production requirements. This is a valid point at which
to linearize the nonlinear equations since the problem being convex any point satisfying
the constraints will be a valid linearization point.
One disadvantage with this slack variable lies in the termination criterion used in
DICOPT ++. It terminates the iterations when a subproblem has a larger objective value
than the previous one. This means that if from a feasible (but bad) point the master
 134

problem projects an infeasible point the iteration is terminated since the objective value
of the subproblem will be larger due to the high penalty on the slack variable. The
algorithm will present the feasible (but highly sub-optimal) point as the optimal solution.

8 Example data
Processing times Size factors Production requirement

n, ~ [l 3
2
3
nS., ~ [! l] 9
10
3
Qi = [260000
260000
]
260000
Cost exponent Cost factor Time horizon Bounds on size

OJ = [0.6 0.6 0.6] aj = [250 250 250] H = 6000h 0 :::; Vi :::; 3000
Data for intermediate storage

SFSi,j =[ 9 1990
~O -=] 'Yj =[
0.5]
0~5 hj =
[ 350 ]
3~0

Data for flexible production requirement

260 000 ] [ 78 * 106 96.2 * 106

Qfef = [ 260 000 c: = 78 * 106
]
c; = [
96.2 * 106
]

260 000 78 * 106 96.2 * 106

c! represents a marginal value of 300 for an extra unit of product around Qfef
ct represents a marginal value of 370 for an extra unit of product around Qfef
Discrete equipment sizes
Vsizej = [500 1000 2000 3000]

Costj = [ 10407 15774 23909 30494]

For problems with no parallel equipment allowed the discrete sizes are:

Vsizej [400 630 1000 1600 2500 4000 6200 10000 15500J
Costj [9103 11955 15774 20913 27334 36239 47138 62787 81684J

9 Results
We have solved the problems presented in chapter 8 with different formulations and we
present the results in the next chapter where we give a schematic picture of the structure
of the plant, total cost and unit sizes. The computational requirements (on a VAX 9000)
and the number of major iterations required by the algorithm are also given. The example
problem is only a small one and to show the variations of solution time we give a table
with the solution time for randomly generated problems with the different features.
 135

9.1 Summary of results

Formulation 6.1 Parallel units out of phase

Cost=208 983 CPU=0.98 s

f~
VI 1096 Major iterations=2
V2 3000 Binary variables 12
V3 2466

Formulation 6.2 Parallel units in and out of phase

Cost=180 472 CPU=1.73 s
VI 1614 Major iterations=2
~ 2633 Binary variables 24
V3 3000

Formulation 6.3 Parallel units in and out of phase, nonidentical units

Cost=180 318 CPU=2.84 s
VI 1650 Major iterations=3
It;l 3000 Binary variables 12
It;2 2143
V3 3000

Formulation 6.4 Flexible use of parallel units in and out of phase

This is the plant configuration for products 1 and 2
Cost=157 864 CPU=3.26 3
VI 2052 Major iterations=2
V2 2692 Binary variables 72
V3 2309

This is the plant configuration for product 3

Formulation 6.5 Intermediate storage and parallel units out of phase

Cost=183613 CPU=20.78 s
VI 2370 Major iterations=3

-e0u-0-o-
V2 3000 Binary variables 21
V3 2666
S S VSj 8871
VSj 12283
Comments: Size of storage, sum of up and
down stream batch size
 136

Formulation 6.5 Intermediate storage and parallel units out of phase

Cost=194 285 CPU=7.01 s
VI 2370 Major iterations=4
\12 3000 Binary variables 21
lt3 1333
VSj 12283
VSj 9553
Comments: Si~ of storage, twice the larger
of up and down stream batch size

Formulation 6.6 Variable production requirement and units out of phase

Cost=208 157 CPU=1.58 s
Plant=201 062 Major iterations=2
Margin= 7 096 Binary variables 12
VI 1000 Ql 251 171
\12 3000 Q2 254 734
lt3 2250 Q3 251 171
Comments:This is the solution in w~ich the marginal value of an extra kg is 300
"money units" at Q[ef

Formulation 6.6 Variable production requirement and units out of phase

Cost=206 020 CPU=1.l1 s
Plant=226 722 Major iterations=2
Margin= -20 702 Binary variables 12
VI 1000 QI 305 555
\12 3000 Q2 273 297
lt3 2250 Q3 264 618
Comments:This is the solution in which the marginal value of an extra kg is 370
"money units" at Qfef

Formulation 6.7 Multiplant production

Total Cost=l77 731 CPU=37.7 s
Plant 1 product 1 Cost=63216 Major iterations=3
VI 520 Binary variables 36
V2 1560
V3 520

Cost=58169 Cost=56346

H
Plant 2 product 2
VI 202 VI 693 =:ant 3 proc_':: 3
\12 1011 V2 502
lt3 202 V3 1560
 137

Formulation 6.8 Discrete unit sizes and parallel units out of phase
Cost=219 718 CPU=2.3 s
v,. 1000 Major iterations=2
Vi 3000 Binary variables 24
"'J 2000

Formulation 6.9.3 Discrete unit sizes - MILP

Cost=175 961 CPU=0.25 s
v,. 6200 Binary variables 27
V2 15500
~ 6200

Formulation 6.9.4 Discrete unit sizes and multiproduct campaigns

Cost=165 061 CPU=0.31 s
v,. 4000 Binary variables 27
V2 15500
~ 6200
Comments: The campaigns used are: 18 batches of product 1, 178 batches of
product 1 and 2 alternately, and 378 batches of product 3.

Formulation 6.9.5 Continuous unit sizes and multiproduct campaigns

Cost=155 036 CPU=0.13 s
\tI. 3597 Binary variables 0
Vi 10790
"'J 8092
Comments: The campaigns used are: 217 batches of product 1 and 2 alternately,
24 batches of product 2 and 3 alternately, and 265 batches of product 3.

Formulation NLP-3 Continuous unit sizes and long single product campaigns
Cost=158371 CPU=0.15 s
VI 3726 Binary variables 0
-0-0-0- V2 11180
"'J 8385

Formulation 6.11, Underutilization 50%

Cost=201 628 CPU=1.72 s
VI 1545 Major iterations=2
V2 2318 Binary variables 12
V3 1738
 138

Formulation 6.11, Underutilization 65%

Cost=235 842 CPU=6.56 s
Vi 1355 Major iterations=2
~ V2 2937 Binary variables 36
V3 1909

Comments: Product 1 uses windows {2 2 2} e.g. on all stages two or more units
are used. Product 2 uses windows {2 2 I} e.g. on stage. three only one of the units
are used. Product 3 uses windows {2 1 2}

Formulation 6.11, Underutilization 77%

Cost=251 013 CPU=7.00 s
"Vi 933 Major iterations=2
V2 2766 Binary variables 48
~ 1909

Comments: Product 1 uses windows {3 2 2} e.g. on stage 1 all three units on the
other stages only two units are used. Product 2 uses windows {2 3 I} e.g. on stage
1 two units, stage 2 three units and stage3 one unit. Product 3 uses windows {3 1
3}

9.2 Discussion of results

It is difficult to compare all the different formulations with each other but as the super-
structure gets larger from one formulation to an other the solution improves (or stays the
same). Formulations 6.1 to 6.4 include only parallel units and it can be seen that when
the flexibility increases the cost of the plant decreases. The solution time also increases
with the flexibility of the formulation. The decrease of cost for the nonidentical units
formulation is marginal but for other examples the savings can probably be greater. In
the flexible use of parallel units we have to reconfigure the plant, in stage three, between
products 1 and 2 and product 3.
The storage formulation (6.5) shows that the cost is reduced by inserting tanks. When
we use the linear sizing constraint (storage twice the largest batch size), the third stage
will have two units out of phase to reduce the batch size of the last subtrain. The linear
storage constraints increase the cost, since the needed storage tank is larger, but reduces
the solution time.
The variable production requirement formulation (6.6) is solved for two marginal values
of the product. With a marginal value of 300 for all products it is advantageous to produce
less than the reference requirement and take the penalty for under production. With a
marginal value of 370 for all products we get a larger plant and the increased cost is
reduced by the profit of the extra products.
The solution of the multiplant formulation (6.7) is that it is more economical to pro-
duce each product in a separate dedicated plant. This problem has the longest solution
 139

time of all problem formulations.

When parallel equipment is not allowed the constraint that units are only available in
discrete sizes increases the cost. Using multiproduct campaigns reduces the cost for both
continuous and discrete sizes.
The underutilization (6.11) constraints increase the cost. The cost increases as the
underutilization demand increases. To avoid batches in the forbidden regions the solution
has to have many parallel units in phase. The solution time increases as we have to add
more binary variables.

9.3 Results on randomly generated problems

In .the table below the solution time and the number of binary variables for the different
features for different problem sizes are given. BV stands for binary variables. CPUs is the
solution time, in seconds, in a VAX 9000. ave. is the average solution time of 5 problems
and max is the largest solution time of the 5 problems. Size factors and processing times
are randomly generated numbers in the range [1, 10J. All problems have three products
(exept 6.7 (multiplant) that has two products) and production requirements Qi = 260 000.
The time horizon is H = 6000 and other data for the features are the same as in Chapter
8.

Problem type Problem size

3 Stages 4 Stages 5 Stages 6 Stages
BV CPUs BV CPUs BV CPUs BV CPUs
ave. max ave. max ave. max
6.1 MINLP 12 0.98 16 1.71 2.40 20 2.01 2.21 24 3.52 6.23
6.2 MINLP 24 1.73 32 5.53 9.97 40 9.23 16.3 48 18.1 47.0
6.4 MINLP 72 3.26 96 13.1 23.0 120 86.2 166 144 334 789
6.5.4 MINLP 18 7.01 26 7.91 12.0 35 18.9 30.7 45 30.9 40.7
6.6 MINLP 12 1.58 16 2.42 4.42 20 2.96 3.70 24 6.65 8.97
6.7 MINLP 36 37.7 35 9.59 16.9 43 75.17 239 51 109 385
6.8 MINLP 24 2.3 32 9.11 11.6 40 15.7 33.2 48 47.5 101
6.9.3 MILP 27 0.25 36 0.24 0.28 45 0.39 0.46 54 1.31 1.59
6.9.4 MILP 27 0.31 36 0.58 0.71 45 0.76 1.09 54 1.35 1.66
6.9.5 NLP 0 0.13 0 0.15 0.18 0 0.16 0.20 0 0.19 0.24
6.11 MINLP 12 1.72 16 3.92 5.58 20 12.0 14.9 24 44.8 131
Table 1. The solution time on randomly generated test problems.
The number of binary variables is a measure of the size of the problem. But the
number of binary variables is not directly proportional to solution time for the different
formulations. The binary variables for parallel units have better continuous relaxations
than do the binary variables for storage and multiplant. The branch and bound procedure
in the master problem can then reduce the search tree by better bounds and this affects
the solution time.
Parallel units out of phase (6.1) and parallel units in and out of phase (6.2) have a
moderate increase in solution time as the size of the problem increase. Flexible use of
 140

parallel equipment (6.4) as a dramatic increase in solution time as the problem gets larger.
The storage formulation with linear sizing constraints (6.5.4) has a moderate increase in
solution time but the solution found by DICOPT++ is usually not the global optimum
due to the termination criterion used. Variable production requirement (6.6) has a slight
increase in solution time compared with (6.1). The multplant formulation (6.7) takes
much more time to solve than solving three problems (two products) with parallel units.
For the formulation with discrete equipment sizes and parallel units (6.8) the solution
time increases more than for (6.1). Without parallel equipment (6.9.3) the formulation
reduces to a MILP and this can be solved easily even with additional constraints for
multiproduct campaigns (6.9.4). Without discrete equipment sizes the problem reduces
to a NLP (6.9.5). With underutilization constraints (6.11) not allowing units to be less
that half full the problem becomes harder to solve than when we drop the constraints
(6.1).

10 Splitting and merging - Synthesis

In all the previous formulations the task structure is fixed and there is no way to optimize
this structure. The search for a good or optimal task structure was previously done in
a sequential manner by splitting tasks at the time limiting stage and merging tasks that
are not time limiting. Yeh and Reklaitis [28] gave a splitting MINLP formulation with
many binary variables for splitting tasks in a single product plant. They formulate the
problem with the binary variable Xr,k,j = 1 if the k th task of stage j is assigned to the
r th unit, 0 otherwise. They come to the conclusion that the formulation is too time
consuming for a preliminary synthesis procedure. Birewar and Grossman [3] give a very
interesting synthesis formulation with a binary variable YtJ = 1 if task t is executed in
unit j, 0 otherwise.
Their formulation contains a number of non convex functions and it is not surprising
that the authors have not found the global optimum to some of their examples. Some
of the example solutions given have even an infeasible structure. One drawback of this
formulation is the elaborate logical constraints on the binary variables and the "semi
binary (Y~;, YFt,j) variables".
Using logarithmic variables to make their problem formulation convex does not work.

10.1 Our synthesis formulation

We presented [19] a similar formulation but with logarithmic variables, This results in a
problem with only one non-convex equation (126). We use the binary variable XtJ = 1
if task t is executed in unit j, 0 otherwise. If two or more tasks are executed in the
same unit we use the largest cost factor of the tasks (equation 128) and the cost factor is
expressed in logarithmic form to avoid a non-convex objective function (122). We define
a set of units J t that can perform tasks t and a set of tasks Tj that can be executed in
unit j.
The logical constraints are much simpler. There are only two logical constraints in
this formulation. Equation (131) ensures that each task is performed in one and only one
unit and equation (132) allows a task to be performed in a unit only if the previous task
 141

is performed in that unit. The first task in a group of mergeable tasks is always fixed to
a unit.
In this formulation we do not have to generate "super units" since the cost (and type)
of a unit is a function of the tasks performed in it. We can therefore save some binary
variables.

J
Cost min L exp(lnaj + InMj + ajlnVj) (122)
j=1
Np
H ;::: LQiexp(lnLCTi -lnBi ) (123)
,=1
lnVj > lnB, + lnS'j - U(1 - X 1j) (124)
T
T.-I,J- LTi.tl';j (125)
1=1
exp(ln'Ii,j) TiJ (126)
InLC'Ii InTiJ -lnMj (127)
Inaj ;::: In(atJ) - U(1 - X 1J ) (128)
MU
L
J

InMj In(k)Yk,j (129)

k=O
MU
J

LYk,J- (130)
k=O
T
LXtJ (131)
t=1
Xt+l,j ~ X tJ (132)
XtJ 0 t ~ 1j (133)
Inv.L
J ~- InVj ~ InVt (134)

In the binary variable Yk,j for parallel equipment k goes from 0 to My but k cannot be
equal to 0 and instead we use k = 1O-~, 1, ... ,MY.
The equation (126) is non-convex. We have to use logarithmic variables in the for-
mulation in order to avoid bilinear horizon constraints. For each product we sum the
processing times of the tasks that are executed at a stage to give the stage cycle time
(125). Constraint (131) ensures that all tasks are executed in one and only one stage.
Constraint (132) ensures that only consecutive tasks are performed at a stage. If any task
is performed on a unit it is either the first task on that unit or its immediate predecessor
task is also allocated to the same unit. A task can only be performed in units that can
execute the task (133).

10.1.1 The noncovex processing time constraint

The equality relaxation in DICOPT++ will relax equation (126) to

exp( InTiJ ) ~ 'IiJ (135)

 142

this equation is non-convex and linearizations obtained by DICOPT++ possibly cut away
the global optimum. We can also replace the non convex equations with a piecewise linear
function as described by Vaselenak et. al. [26] and implement the outer approximation
with piecewise linear approximator with APROS [17] in GAMS. This introduces a large
number of binary variables.

10.1.2 The linear processing time constraint

We can also replace equation (126) with a number of linear equations. For example when
we have three tasks in a group that can be merged we get the linear constraints:

InTiJ > In(Ti,I) - U(1 - X IJ ) (136)

InTiJ > In(Ti,1 + Ti,t+t) - U(2 - XIJ - XI+l J ) (137)
InTiJ ~ In(Ti,1 + Ti,t+l + Ti ,I+2) - U(3 - XIJ - XI+l J - X I+2J) (138)
InTi,HI > In(Ti,I+l) - U(1 - XI+l,i+I) (139)
InTi,i+l ~ In(Ti,I+I + Ti,t+2) - U(2 - XI+1J+I - X I+2J+l) (140)
InTi,i+2 ~ In(Ti,t+2) - U(1 - Xt+2J+2) (141)
Task t, t + 1 and t + 2 can be merged into one unit j or performed separately in units
j,j + 1 and j + 2. Four different configurations are possible.

• Unit j performs tasks t, t + 1 and t + 2

• Unit j performs tasks t and t + 1, and unit j + 2 performs task t + 2
• Unit j performs task t and unit j + 1 performs tasks t + 1 and t + 2

• Unit j performs task t and unit j + 1 performs task t + 1 and unit j + 2 performs
task t + 2
For the first configuration equation (138) gives the processing time. For the second con-
figuration the equation (137) and equation (141) give the processing time. For the third
configuration equation (136) and equation (140) give the processing time. For the fourth
configuration the equation (136), (139) and equation (141) give the processing time.
These constraints replace the nonconvex constraint (127) but they have to be formu-
lated depending on the problem. We will show how this is done on our example problem.

10.1.3 Example data

We use the problem 2 stated in Birewar and Grossman [3]. We have to formulate the
problem parameters somewhat differently.
Processing time TiAh)
products mixl rxnl distln mix2 rxn2 crystl Production Requirement (kg)
A 2 8 4 1 6 9 600000
B 2 4 3 1 5 4 600000
C 1 6 5 3 7 4 700000
D 3 5 6 2 9 3 700000
E 3 7 5 2 8 1.5 200000
F 2.5 4 4 3 4 2 100000
 143

Size factors Si.t

products mixl rxnl distln mix2 rxn2 crystl
A 3 1 3 4 2 1
B 5 4 4 5 3 4
C 4 2 2 3 2 3
D 3 2 2 3 1 3
E 3 2 2 4 1 4
F 4 4 4 4 4 5
Horizon time (H) = 6000 h. Cost factor OJ (for all units) = 0.6
atJ is the cost factor if task t is performed in unit j

Task Unit
1 2 3 4 5 6
mixl 200
rxnl 300 300
distln 450
mix2 200
rxn2 300 300
crystl 550 550 250
It.i is the fixed cost factor if task t is performed in unit j

Task Unit
1 2 3 4 5 6
mixl 45000
rxnl 55000 55000
·distln 70000
mix2 45000
rxn2 60 000 60 000
crystl 95000 95000 50000
The relation between Tasks and unit types

Task Unit
1 2 3
4 5 6
mixl CI, NJ, A
rxnl CI, J, A CI, J, A
dis tIn Des. Col
mix2 CI, NJ, A
rxn2 SS, NJ, ASS, NJ, A
crystl SS, J, ASS, J, A CI, J
CI cast iron, SS stainless steel, NJ nonjacketed, J jacketed, A agitator and Des. Col.
destilation column.
We can see that if task" crystl" is performed in unit 4, this unit is a "super unit" as
Birewar and Grossman call it. Tasks "mix2" and "rxn2" have to be performed by the
same unit.
 144

10.1.4 Convex MINLP formulation for example

N
Cost = min L: exp(In,j + InMj )
j=1
N
+ L exp(Inaj + OJ InYj) (142)
j=1
Np
H > L Qi exp(InLC1'; - 1nBi) (143)
i=1
InYj ~ InBi + InSi" - U(I- X,J) (144)
InLCTi ~ InTiJ -lnMj (145)
Inaj In(atj) - U(1 - X tj ) (146)
In,j In( ,t,j) - In,m=(1 - Xtj) (147)
M!l

L: In(k)Ys:,j
J

InMj = (148)
k=O
MU
J
1 LYk ,J" (149)
k=O
T
LX"j = 1 (150)
1=1
X ' +lJ :S X"j t E 1'; (151)
X,J = 0 (152)
InVLJ
< In Yj :S In V;U (153)
and the problem specific constraint on the processing time

InTi,l ~In(Ti,i) - U(1 - XI,d (154)

InTi,1 ~ In(T;,1 + T;,2) - U(2 - XI,l - X2,d (155)
InTi,2 ~ In(Ti,2) - U(1 - X2,2) (15p)
InTi,3 ~ In(T;,3) - U(1 - X 3,3) (157)
InTi,4 > In(T;,4) - U(1 - X 4,4) (158)
InTi,4 ~ In(T;,4 + T;,s) - U(2 - X 4,4 - XS,4) (159)
InTi,4 ~ In(T;,4 + T;,s + T;,6) - U(3 - X 4,4 - XS,4 - X 6,4) (160)
InTi,S > In(T;,s) - U(1 - Xs,s) (161)
InTi,S > In(T;,s + T;,6) - U(2 - Xs,s - X6,s) (162)
InTi,6 > In(Ti,6) - U(1 - X6,6) (163)
These constraint have very poor continuous relaxation.

10.1.5 Results
In [3J solutions are presented for four different cases (a) to (d):
 145

• (a) Merging not allowed, Single product campaigns.

• (b) Merging allowed, Single product campaigns, Zero wait scheduling.

• (c) Merging allowed, Multi product campaigns, Zero wait scheduling.

• (d) Merging allowed, Unlimited intermediate Storage.

Problem
(a) (b) (c) (d)
Unit size size size size
mixl 13592
rxnl 9057 15250 19310 18 129
distIn 9091 11 268 12000 11 250
mix2 12500 15000
rxn2 6897 8511 15000 15000
crystl 12500 15000 15000 15000

CPUs for solving the problems (Microvax II)

2898s 1092s 577s
Cost presented in paper
775 840 713 276 649 146 640 201
We have recalculated the costs from the unit sizes given in [3] with the following results:
752685 711 074 649 146 640201

With the unit sizes given in [3] we have calculated the mimimum total processing time
reqired to meet the demand with the following results:
6805 h 6283 h 6266 h 6266 h

All solutions presented are infeasible (availiable time is 6000 h). The results of (c) and
(d) even have a infeasible structures. Even with the four units in (d) (rxnl, distIn, rx2,
crystl) at the upper bound on size the production requrements cannot be fullfilled within
the time horizon. If we cannot solve the problem with Unlimited intermediate Storage
the multiproduct campaign is also infeasible.
We have not tried to solve the Multiproduct campaign problem (c) since formulating
the campaign constraints will introduce a number of non convex (bilinear) constraints and
therefore the solution cannot be guaranteed to be globally optimal.

Proposed solutions Problem

(a) (b) (d)
Unit size size size
mixl 15000
rxnl 10 000 17800 16893
distIn 10727 11 867 11 262
mix2 14298 15000 15000
rxn2 7500 8900 8446
crystI 14298 15000 15000
 146

Proposed cost and solution time (seconds on Vax9000)

786 356 726 328 716 995
CPUs 1.19 s 26.04 s 50.78 s

The large savings in cost reported in the paper (16.3% for MPC and 17.5% for UIS)
are mainly due to the fact that the reported solution moves into a region where one unit
can be deleted. This structure is in fact infeasible. With a feasible plant the cost savings
for UIS is 8.8% and probably less for MPC.

11 Discussion
In this paper we have developed a number of different formulations, which can all be
solved to optimality, for the same general problem. The question now arises which of
these are the best or how do we get a "total global optimum". There are as we see it
two ways. One is to combine all formulations to a super formulation which would then
contain all discrete choices but probably be too large to solve even for small problems.
The other way is to solve a small part and at the same time generate information about
what additional features could with advantage be included. this could provide the basis
for an iterative improvement of the solution.

11.1 How the different formulations are related

The model formulations of Section 6, shown in Fig. 2, are also graphically depicted in
section 9.1.

Figure 2. The relation between models

The formulations on the left side of the vertical line in Figure 2 do not include parallel
equipment. The arrows pointing upwards indicate that the new formulation will have
a larger or equal objective function value but the relative position of a formulation to
all others in Figure 2 is meaningless. We start from NLP-3 which is the formulation
presented in chapter 3 but without paral\el equipment and without bounds on unit size
and assuming that the products are produced in long single product campaigns. From
this formulation we can:
 147

• impose an upper bound on size and allow parallel units out of phase which leads to
formulation 6.1

• add a constraint that units are only available in discrete sizes which leads to formu-
lation 6.9.3

• allow multiproduct campaigns which leads to formulation 6.9.5

From formulation 6.9.3 we get formulation 6.9.4 by allowing multiproduct campaigns.

Formulation 6.1 is the MINLP sizing problem with parallel units out of phase and from
this we can get a number of other formulations:

• allowing parallel units to operate in phase leads to formulation 6.2

• allowing intermediate storage leads to formulation 6.5

• allowing the products to be produced in different plants leads to formulation 6.7

• allowing variable production requirement leads to formulation 6.6

• imposing the constraint that units are only available in discrete sizes leads to for-
mulation 6.8

from formulation 6.2, with parallel units in and out of phase, we can:

• allow parallel equipment with flexible operation, different operation for different
products, which leads to formulation 6.4

• allowing units in phase to have unequal sizes leads to formulation 6.3

• add the underutilization constraints in chapter 6.11

The storage formulation can be formulated as in chapter 6.5.4 or with the linear storage
sizing constraints in chapter 6.5.2.

12 Conclusions
More than a dozen examples have been presented of the basic problem of equipment
sizing for a multiproduct batch plant. Almost all of the examples require integer choices.
They are presented in a common format and results are given for the solution of a basic
problem in which the various extensions are incorporated in turn. Some of the extensions
or formulations for particular extensions have been presented by previous authors, but
several are new. These include the flexible use of parallel items, both in and out of phase,
with different configurations for different products, the choice of intermediate storage
location, the trade off between marginal product value and increased equipment cost,
constraints preventing underutilization of equipment, transformation of the discrete sizing
problem to an integer linear program, the multiplant problem with parallel equipment
items and a new MINLP formulation of the process synthesis problem of task splitting
and merging.
 148

The simple demonstrative examples are used to show the result of the different formu-
lations. The problem size is given in terms of the number of binary variables in the result
summary (chapter 9.3) and the problem difficulty is indicated by the solution times.
The MINLP formulation seems to be quite well suited for some problems i.e. paral-
lel equipment items, variable production requirements and underutilization. For others
including storage location and multiplant selection, unless better MINLP formulations
can be found it seems preferably to use enumeration, or for larger problems a stochastic
method such as simulated annealing (Patel et.al. (1991)) which effectively makes a partial
enumeration of promising alternatives.

13 Nomenclature
Cost factors Zcj, Zc,ij Binary variables
atj, amar Cost factors P ij Time constant
bj Cost factor, storage PTi,p Production time
B i , Bi,q, Bij Batch size U Large skalar
B up , Bdown Batch size Greek letters
Ci Marginal value of product aj, Cost exponent
Costj Cost of stage 'Pi> Underutilization
Ii Production cost factor Aj, Time exponent
J Number of stages Ij, Cost exponent for storage
LCTi, LCTij Limiting cycle time
nk, ni Number of batches Subscript
Np Number of products Product
Mj , Mij , M U Parallel units out of phase j Stage
N j , Nij , N U Parallel units in phase k,c For parallel units
N P RSi,k Multi product campaigns q Subtrain
PCi Production cost Discrete equipment
9
P RO i Inverse productivity p Plant
Qi, Qref Production requirements t Task
R;,j Time factor Superscript
Si,j, Si,t Size factor max Maximum
SF Si,j Size factor storage U Upper (bound)
S Li,k,j Slack time L Lower (bound)
Tij , n,t Processing time Transformation of variables
Tii,p . Total processing time lnX Variable expressing the log.
TotMj Total number of parallel units value of variable X
Vj, Vj,p Unit size (Volume) invX Variable expressing the
VSj , VS, Storage volume inverse of variable X
V sizej,g Set of discrete sizes Logarithms and exponential's
V costj,g Set of discrete costs In(X) Taking the logarithm of
Xj,q, Xi,p Binary variables parameter X
Xij,w, X j,9 Binary variables exp(X) The exponential of
Ykj , Yk,ij, Yt,j Binary variables parameter or variable X
 149

REFERENCES

1. D.B. Birewar and I.E. Grossmann. Efficient optimization Algorithms for Zero-Wait Scheduling of
Multiproduct Batch Plants. Ind. Eng. Chem. Rer., 28: 1333-1345, 1989.
2. D.B. Birewar and I.E. Grossmann. Simultaneous Production Planning and Scheduling in
Multiproduct Batch Plants. Ind. Eng. Chem. Rer., 29: 570-580, 1990.
3. D.B. Birewar and I.E. Grossmann. Simultaneous Synthesis, Sizing and Scheduling of Multiproduct
Batch Plants. Ind. Eng. Chem. Rer., 29: 22A2-2251, 1990.
4. D.B. Birewar and I.E. Grossmann. Incorpaating Scheduling in the Optimal Design of Multiproduct
Batch Plants. Complllerr and Chem. Eng., 13(1fl): 141-161, 1989.
5. G.A. Coulman. Algorithm for Optimal Scheduling and Revised Formulation of Batch Plant Design.
Ind. Eng. Chem. Res., 28: 553, 1989.
6. . M.A. Duran and I.E. Grossmann. A Mixed-Integer Nonlinear Programming Algorithm for Process
System Synthesis. AlChE J., 32(4); 592-606, 1986.
7. A. Espuna, M. Lazaro, I.M. Martinez, and L. Puigjaner. Efficient and Simplified Solution to the
Predesign Problem of Multiproduct Plants. Computers and Chem. Eng., 13: 163-174, 1989.
8. R.S. Garfmkel and GL. Nemhauser. Integer Programming. Wiley: New York, 1972.
9. I.E. Grossmann and R.W.H. Sargent Optimal Design of Multipurpose Chemical Plants. Ind. Eng.
Chem. Procerr. Der. DIN., 18(2), 1979.
10. LA. Karimi and G.V. Reldaitis. Intermediate Storage in Noncontinuous Processes Involving Stages
of Parallel Units. AlChE J., 31: 44, 1985.
11. 1. Klossner and D.W.T. Rippin. Combinatorial Problems in the Design of Multiproduct Batch
Plant - Extension to Multiplant and Partly Parallel Operation. Prerented at the AlChE Annual
Muting, San Francirco, Nov. 1984.
12. F.C. Knopf, M.R. Okos, and G.V. Reldaitis. Optimal Design of Batch/Semicontinuous Processes.
Ind. Eng. Chem. Procers. Des. Dev., 21: 79-86, 1982.
13. G.R. Kocis and I.E. Grossmann. Global Optimization of Nonconvex Mixed-Integer Nonlinear
Programming (MINLP) Problems in Process Synthesis. Ind. Eng. Chem. Res., 27: 1407-1421,
1988.
14. G.R. Kocis and I.E. Grossmann. Relaxation Strategy for the StruCtural Optimization of Process
F1owsheets. Ind. Eng. Chem. Res., 26: 1869-1880, 1987.
15. Y.R. Loonkar and I.D. Robinson. Minimization of Capital Invesunents for Batch Processes. Ind.
Eng. Chem. Process. Des. Dev., 9(4), 1970.
16. A.K. Modi and I.A. Karimi. Design of Multiproduct Batch Processes with Finite Intennediate
Storage. Complllers and Chem. Eng., 13(1/2): 127-139, 1989.
17. G.E. Pales and C.A. F1oudas. APROS: Algorithmic Development Methodology for Discrete-
Continuous Optimization Problems. O~rations Research, 37(6): 902-915, 1989.
18. A.N. Patel, R.S.H. Mah, and LA. Karimi. Preliminary Design of Multiproduct Non-Continuous
Plants Using Simulated Annealing. Complllers and Chem. Eng., 1991.
19. D.E. Ravemark and D.W. T. Rippin. StruCture and Equipment for Multiproduct Batch Production.
Presenled til AlChE 1991 Annual Meeting, Nov. 1991.
20. D.W. T. Rippin. Design and Operation of Multiproduct and Multipurpose Batch Chemical Plants -
An Analysis of Problem Structure. Complllers and Chem. Eng., 7(4): 463-481, 1983.
21. I.D. Robinson and Y.R. Loonkar. Minimizing Capital Investments for Multiproduct Barchplants.
Process Technol. fill., 17(11), 1972.
22. H.E. Salomone and O.A. Iribarren. Posynomial Modeling of Batch Plants: A Procedure to Include
Process Decision Variables. Complllers andChem. Eng., 16(3): 173-184, 1992.
23. R.E. Sparrow, G.I. Forder, and D.W.T. Rippin. The Choise of Equipment Sizes for Multiproduct
Batch Plants. Heuristics vs. Branch and Bound. Ind. Eng. Chem. Process. Des. Dev., 14(3), 1975.
24. I. Suhami and R.S.H. Mah. Optimal Design of Multipurpose Barch Plants. Ind. Eng. Chem.
Process. Des. Dev., 21:94-100,1982.
25. T. Talaunatsu, I. Hashimoto, and S. Hasebe. Optimal Design and Operation of a Batch Process with
Intennediate Sunge Tanks. Ind. Eng. Chem. Process. Des. Dev., 21: 431-440,1982.
26. 1. Vaselenak, I.E. Grossmann, and A.W. Westerbez"g. Optimal Retrofit Design of Multiproduct Batch
Plants. Ind. Eng. Chem. Res., 26: 718-726, 1987.
27. 1. Wiswanathan and I.E. Grossmann. A Combined Penalty Function and Outer-Approximation
Method for MA1NLY Optimization. Complllers and Chem. Eng., 14(7): 769-782, 1990.
28. N.C. Yeh and G.V. Reklaitis. Synthesis and Sizing of Barch/Semicontinuous Processes: Single
Product Plants. Complllers and Chem. Eng., 11(6): 639-654, 1987.
The Influence of Resource Constraints on the Retrofit
Design of Multipurpose Batch Chemical Plants

Savoula Papageorgaki1, Athanasios G. Tsirukis 2, and Gintaras V. Reklaitis 1

1. School of Chemical Engineering, Purdue University, W. Lafayette, IN 47907, USA

2. Department of Chemical Engineering, California Institute of Technology, Pasadena, CA 91125,
USA

Abstract: The objective of this paper is to study the effects of resource availability on the
retrofit design of mUltipurpose batch chemical plants. A mixed integer nonlinear model will
be developed to address retrofit design arising from changes in the product demands and
prices and revisions in the product slate (addition of new products, removal of old products,
modifications in. the product recipes). Resource constraints on the availability of utilities such
as electricity, water and steam, manpower, etc. will be incorporated into the formulation. In
addition, the option of resource expansion to accommodate the needs of the new plant will be
explored.
A decomposition solution strategy will be presented to allow solution of the proposed
MINLP optimization model in reasonable computation time. The effectiveness of the
proposed model and solution strategy will be illustrated with a number of test examples.

Keywords: Batch design, retrofit, resource, mathematical programming.

Introduction

Retrofit design is defined as the redesign of an existing facility to accommodate

revisions in the product slate and/or readjustments in product demands and feedstock
availability, as well as to improve the operability of the process by means of increases in the
process flexibility and reduction of the operating costs and the energy consumption. This
problem is an important one to process operations because of the need to respond to
variations in the availability and prices of feedstocks and energy, the short life cycles of many
specialty chemicals and the continuing pressure to develop and accommodate new products.
The availability of resources can significantly influence the feasibility and quality of the
design during retrofit as resource limits may impose constraints on the extent of retrofit
modifications.
 151

The objective of this paper is to study the effects of resource availability on the retrofit
design of multipurpose batch chemical plants. Retrofit in batch processes has been only
sparingly investigated with most of the attention directed at the retrofit of multiproduct batch
plants. A complete survey on the existing approaches can be found in [6]. The authors of this
paper developed a MINLP optimization model and a subsequent solution strategy for the
retrofit design of multipurpose batch plants in view of changes in the product demands/prices
and/or revisions in the product slate. The issue of resource availability during retrofit,
however, has not been addressed to date, although resource restrictions may impose the need
for extensive modifications during retrofit.

Problem Statement

The deterministic retrofit design problem for a general multipurpose batch facility can
be defined as follows [6]:
Given:
1. A set of products, the current production requirements for each product and its selling price
and an initial configuration of equipment items used for the manufacture of these products.
2. A set of changes in the demands and prices of the existing products over a prespecified
time period (horizon) and/or a set of modifications in the product slate in the form of addition
of new products, elimination of old products or modification of the product recipe of existing
products.
3. A set of available equipment items classified according to their function into equipment
families, the items of a particular family differing in size or processing rate. Items that are
members of the same equipment family and have the same size belong to the same equipment
type.
4. Recipe information for each product (new and old) which includes the task precedence
relationship, a set of processing times I rates and a corresponding set of size I duty factors,
both associated with every feasible task- equipment pair. In general, the processing time may
be specified as a function of the equipment capacity.
5. The set of feasible equipment items for each product task.
6. The status (stable or unstable) and the transfer rules for the intermediates produced
between tasks.
7. Resource utilization levels or rates and change-over times between products with their
associated costs.
8. Inventory availability and costs.
9. A suitable performance function involving capital and/or operating costs, sales revenue and
inventory costs.
 152

Determine:
(a) A feasible equipment configuration which will be used for the manufacture of each
product in the plant (new and old),
(b) The sizes of new processing units and intermediate storage vessels and the number of
units required for each equipment type (new and old),
so as to optimize the given performance function.

Model Formulation

The mixed integer nonlinear programming (MINLP) formulation developed by [6], to

address the optimal retrofit design of general multipurpose batch chemical plants with no
resource considerations, can be extended to incorporate resource restrictions. The key
structural choice variable required by this formulation is the binary variable X imegk which is
defined as follows:

I if task m of product i is performed in unit type e in equipment group g

X imegk = { and in campaign k
o otherwise
The variable set also includes variables denoting number of units, number of groups,
batch sizes, cycle times, number of batches, campaign lengths and production amounts. The
constraint set consists of seven principal subsets:

1. Assignment and Connectivity constraints

2. Production Demand constraints
3. Cycle Time and Horizon constraints
4. Equipment Size and Equipment Number constraints
5. Batch Size and Batch Number constraints
6. Direct and Derived Variable Bounds
7. Degeneracy Reduction Constraints

Finally, the objective function is a maximization of net profit (or minimization of -net
profit) including the capital cost of the new equipment, the operating cost associated with the
new and old equipment and the sales revenue resulting from the increased production of the
old products and the production levels of the new products over a given time period.
Since resource availability will be considered in this work, an additional constraint set
will be introduced into the formulation, namely
 153

8. Resource Utilization Constraints

along with the appropriate set of variables [7]. If resource expansion is also considered, then
an additional linear tenn will be introduced in the objective function as will be shown below.
The set of resources includes all the plant utilities that supplement the operation of equipment
items in a batch plant. For example, the reaction task of some product has specific
temperature requirements which imply the use of heating or cooling fluids. electric filters
consume electricity and almost every processing task requires the attendance of a human
operator. The present paper deals with the class of renewable resources whose availability
levels are replenished after their usage. Examples of renewable resources are manpower.
electricity. heating and cooling flowrates. water and steam. etc. Simple extensions to the
proposed fonnulation can accommodate the existence of consumable resources such as raw
materials. capital. etc.
Let us now define the following sets:
RES = { s I resource s is available to the plant I. S 1. = { i I product i uses resource s I and.
S2. = { m I task m uses resource s }.
Furthennore. let rsimegk denote the utilization level of resource s by task m of product i
perfonned in unit type e and in group g c\uring campaign k. Then. by assuming nonlinear
dependence of the resource utilization level on the split batch size BSimegk • we get the
following equation [7]

rsimegk = 11sime NUimegk + (jsime NUimegk BS imegk IJ....

where SERES; i E SIs; me TAi ( l S2s; e E P im ; g=l •...• NG~ ; k=1 ....• K.
Furthennore. NUimegk denotes the number of units of type e contained in group g that is
assigned to task m of product i during campaign k. and. 11sime. (j.ime and ~.ime are given
constants.
In addition. let RS s denote the utilization level of resource s. Then. the total amount of
resource usage is constrained by RS s as follows [7]:

NC:::
L L rsimegk ~ RSs SE RES; k=1 •... ,K
ieSl, meTA;llS2, eePg" g=l

Finally, let prss denote the cost coefficient associated with the utilization of resource s.
Then. the following tenn will be added to the objective function to account for resource
expansion:

L prss (RS s - RS~in)

seRES

The complete optimization model is shown in Appendix I. The extended fonnulation

exhibits the same characteristics as the original retrofit model, namely. nonconvexity of the
objective function and of the set of feasible solutions which leads to existence of many local
 154

optima, and, combinatorial complexity which results in prohibitively high computation times
for problems of practical scope. Consequently, rather than attempting to solve the proposed
model directly, a formulation specific decomposition of the original MINLP will be
developed.

Solution Procedure

The solution procedure builts on our earlier developments for the retrofit design of
multipurpose plants with no resource restrictions [6]. Specifically, the original MINLP
problem (posed in minimization form) will be decomposed into two subproblems, an upper
and a lower bound (master) subproblem, which are relaxations of the original model and
which will be solved alternately until one of the termination criteria is met. The form of the
relaxed subproblems, however, will be different to accommodate the incorporation of the
resource constraints. Two versions of the decomposition scheme were proposed by [6] with
flowcharts depicted in Figures 1 and 2. The two relaxed subproblems are described in the
following sections.

Master Problem

The master problem is a broad relaxation of the original MINLP problem. The
corresponding formulation is shown in Appendix II. Details on the derivation of constraints
(11.3), (IlA), (II.5), (11.6)-(II.13) can be found in [6], whereas the derivation of (II.2) follows
easily from the definition of the variable VCimegk = NUimegk V~ and constraint (I.19). Integer
cuts corresponding to infeasible assignments identified by the upper bound subproblem
(described below) or cuts corresponding to previously identified assignments may also be
included 'in the formulation. The following proposition describes the sufficient condition
under which the master problem provides a valid lower bound on the optimal solution of the
original MINLP:

Proposition 1. The master problem is a relaxation of the original MINLP model and
provides a valid lower bound on the optimal solution of the original MINLP, if be::; 1 for all
e.

The proof appears in [6].

The master problem is nonconvex due to the nonlinear terms involved in constraints
(1.12), (1.14), (1.15), (1.18), (1.27), (1.29), (1.30), (1l.2) and (I1.3). This form of the problem
cannot be convexified, since most of the variables involved in the formulation have zero as
their natural lower bound. Following the procedure developed in [6], we assume that the
lower bounds on selected variables are equal to E instead of 0, where E is a very small positive
number. Then the following proposition is true:
 MINLP Formulation MINLP Formulation
Fix K Update K Fix K Update K

Fix PR New PR Fix X New X

Upper Upper
MINLP NLP
Bound Bound
Subproblem Subproblem
Ru Ru

01
01
Lower Lower
Bound MINLP MINLP
Bound
RL Master Problem Master Problem
RL

Infeasible Master? No Infeasible Master ? No

K = K max ? No K = K max ? No
Yes Yes
STOP STOP

Figure 1. First version of proposed Decomposition Algorithm Figure 2. Second version of Proposed Decomposition Algorithm.
 156

Proposition 2. If the zero lower bounds on the variables CAPe. VCuneak' NUimeak • Oime.
Tk. TLa.' Yik. nik. npimak and B_~imeak_ are substituted with £ • then (i) the optimal solution to
the master problem will not change and (ii) the optimal profit will be modified by a term 0(£).
provided that £ is a sufficiently small positive number.

Proof of this proposition appears in [6]. A practical guide to selecting the value of £ is to
choose a value which is much smaller than the values of the cost coefficients in the objective
function.
Now the problem can be convexified through exponential transformations of variables
Yik. BSimeak' NUimegk. nPimak' nik. TLa. and Tk• namely

BSimegk = exp (bsimeak )

TLa. = exp (tlik)

After this substitution. constraints (1.11). (1.12). (1.15). (1.17). (1.18). (1.20). (1.27)-
(1.30). (ll.2). (11.3) and (ll.4) take the following form

K
L exp (SYik) = Pi (Ill.2)
1<=1
K
LL exp (Vimeak + bSimegk + snpimgk) :S Onne (Ill.3)
1<=1 g

(Ill.4)

(Ill.5)

(Ill.6)

Ne ~ LLL exp (Vimegk) (Ill.7)

i m g
 157

L exp (Vimegl: + bSimegl:) ::; Bra>: (IIL8)

e

L exp (snpimgt) exp (-snjJ:) ::; (llI.9)

g

LLLL T\sime exp (Vimegl:) + asime exp Utsime bSimegl: + Vimegl:) ::; RSs (ll.1O)
i meg

VCimegl: ~ Sime exp (bsimegt + Vimeg/() (m.ll)

VCimegl:
L L ---"- ~ exp (SYjJ: - snjJ:) (IIL12)
g e Sime

snil: ~ S)'ik -In (BF) (IIL13)

of which only the first is nonconvex (nonlinear equality). To remedy the situation, this
constraint will be replaced by two equivalent constraints
K
L exp (SYjJ:) ::; Pi (m.2a)
1:=1

K
L exp (SYjJ:) ~ Pi (m.2b)
1:=1

the first of which is convex and will be retained in the formulation and the second is
nonconvex and it will be linearly approximated as follows
K
L (Oi/( S)'jJ: + ai/() ~ Pi (llI.2c)
1:=1

exp (sywn) - exp (sy~ax)

where Oil: = --------:---
sy~ax -S)'lF

Notice that a piecewise linear overestimation [2] to more closely approximate this constraint
can also be constructed at the cost of introducing additional binary variables in the model.
Constraint (1.14), however, remains nonconvex. By considering the logical equivalent
of this constraint,
 158

where

t~t ~ t2..e + a;",., exp (~;",., bS;""'gt)

and by rearranging the terms as follows

t·• t
NG:_L > ~X· L (III. 14)
'"'" - TLit IIMg..

we introduce a form which can be linearized according to the procedure introduced by [3] for
t·• k
bilinear constraints involving a continuous variable (in this case .;:;.: ) multiplied by an
integer variable (in this case Xi_gt). For this purpose, the following variables are introduced
and substituted in the model

• ..
_ t·Img
SRG
L

imgt - - - (III. IS)

h ..
(III. 16)

After this substitution, constraint (III. 14) takes the following equivalent form

(III.14*)

and, the bilinear equality (llI.l6) can be substituted by the following equivalent set of linear
inequalities

SRGimgk - SRG~k (1 - X;""'gk) ~ RGitMgk (1l.16a)

SRG:::'~k. (1 - X;""'gk) ~ RG;""'gk (llI.l6b)

SRGimgk - SRG:::'~k (1- XitMgk) ~ RGitMgk (1l.16c)

(III.I6d)

Finally, the nonlinear equality (m.IS) takes the following convex form after
substituting for t :mgt

(m.IS*)
 159

Finally, the bounding constraints (1.16), (1.21), (1.22), (1.31)-(1.33), (II.6), (11.8) and
(II.9) take a slightly different but equivalent form to account for the non-zero lower bound on
the variables and their exponential transfonnation:
NG=:
tlu, ~ (In(Tra"") -lnE) L L L X imLgk + lnE "t i, k (1.16*)
meTA; g=l eeP..

"ti,m,e,g,k (1.21 *)

Vimegk ~ (In(N~"") -lnE) XimLgk + InE "t i , m , g , e , k (1.22*)

bSimegk ~ In(B 7) Ximegk + lnE (1 - Ximegk) "t i , m , e , g , k (1.31 *)

SnPimgk ~ In(np:'gk) Zimgk + lnE (1 - ZimgJ:;) "t i , m , g , k (1.32*)

Snpimgk ~ (In(npl!:.~k) -lnE) Zimgk + InE "t i , m , g , k (1.33*)

K NG::
Qime ~ (InPj"ax -lnE) L L X imLgk + InE "ti, m, e (II.6*)
k=l g=l

(II.8*)

SYi.\: ~ (In(Pj"ax) -lnE) PRi.\: + InE "t i , k (II.9*)

The new formulation (III) consists of minimizing equation (ILl) subject to constraints
(lII.2a), (1II.2c), (III.3)-(III.13), (lII.14*), (1II.15*), CIIU6a)-(lII.I6d), (11.5), (lI.6*), (11.7),
(lI.8*), (II.9*), (II.1O)-(lU3), (1.2)-(1.10), (U6*), (1.23)-(1.26), (1.31 *)-(1.33*), (I.35)-(I.39),
(1.45)-(1.49), (1.51). Clearly, formulation (III) is a relaxation of formulation (II) due to the
linear underestimation of the exponential terms in constraint (III.2c). As already mentioned,
the new formulation constitutes a convex MINLP model which can be solved for its
corresponding globally optimal solution. DICOPT (software implementation of the OA/ER
algorithm [4]) can be used to solve the convex master problem utilizing MPSX to solve the
MILP subproblems and MINOS 5.0 [5] for the NLP subproblems of the corresponding
MINLP. Note that the OA/ER algorithm guarantees the global optimal solution for convex
MINLP's.

Upper Bound Subproblem

In the first version of the decomposition algorithm, the upper bound subproblem
corresponds to the original MINLP with the values of the integer variables PRi.\: fixed.
Consequently, the upper bound subproblem subproblem remains a MINLP model, but
contains less binary variables than the original MINLP, since the product-campaign
 160

assignment is fixed and thus, several sets of binary variables can be eliminated from the
model. In the second version of the decomposition scheme, the upper bound subproblem is a
NLP model, since it corresponds to the original MINLP with the values of the binary
variables Ximegl: fixed. In both cases, the value of the objective function provides an upper
bound on the optimal solution of the original MINLP. However, the problem formulation is
nonconvex and cannot be convexified through variable transformations.
DICOPT++ (software implementation of the AP/ONER algorithm [9]) can be used for
the solution of the nonconvex MINLP upper bound subproblems in the first version of the
decomposition procedure and MINOS 5.0 can be used to solve the NLP upper bound
subproblems in the second version of the algorithm

Example

A multiproduct plant involving 4 products and 4 stages [8] is considered in this

example. Since it is assumed that there is a one to one correspondence between stages and
equipment families, four different equipment families are available in the plant. An initial
equipment configuration involving one unit in each of the stages 1,2 and 4 and 2 out-of-phase
units in stage 3 is given. For each of the existing units, the addition of a single unit in- and
out-of-phase is considered. Consequently, the resulting maximum number of single-unit
equipment groups that are allowed in each stage is 2 for stages 1,2 and 4, and 3 for stage 3. In
addition, there are 14 equipment types available in the plant that are detailed in Table 1. Note
that, since no upper bounds on the equipment sizes have been explicitly given by [8], the
sizes of the existing equipment will be used as upper bounds. In addition, since the proposed
model requires non-zero lower bounds on the equipment capacities, a minimum capacity of
500 has been assumed for each equipment type. The unit processing times (assumed to be
constant in this example) and size factors and the upper bounds on the annual production
requirements and selling prices are given in Tables IT and 1lI. The authors approximated the
capital cost of equipment by a fixed-charge model which is incorporated into our formulation
in the following equivalent form

NEQ
~ ('Y" Ve Ne + I.e Ne )
,,=1
=
The cost coefficients 'Ye and I.e are given in Table IV. Notice that since CAP Ve Ne in the
master problem, the values of coefficients 'Ye and I.e will be used for coefficients Ce and de.
Also notice that the value of coefficient "-G, has been corrected from the value of 10180 to the
value of 44573 to agree with the reponed results.
Additional assumptions that funher simplify the model are that all products must use
all units in the plant and thus, there is no product dependence of the structural variables, no
operating costs are considered and the old units must be retained in the plant. As a
consequence, a two-subscript binary variable suffices for the representation of the structural
decisions that must be made at the design stage
 161

Table I. Available Equipment Items

equipment item capacity range (L)

Rl *,R2,R3 4000 *, 500-4000 1 *,1,1

Ll *,L2,L3 4000 *, 500-4000 1 *,1,1
Fl *,F2 *,F3,F4,F5 3000 *, 3000 *, 500-3000 1 *,1 *,1,1,1
Gl *,02,03 3000 *, 500-3000 1 *,1,1

* : existing units

Table ll. Size Factors (L/kg/batch) and Processing Times (h/batch)

( ) = Processing Time

product / eq. type Rl,R2,R3 Ll,L2,L3 FI-F5 Gl,G2,03

A 7.9130(6.3822) 2.0815(4.7393) 5.2268(8.3353) 4.9523(3.9443)

B 0.7891 (6.7938) 0.2871(6.4175) 0.2744(6.4750) 3.3951(4.4382)
D 0.7122(1.0135) 2.5889(6.2699) 1.6425(5.3713) 3.5903(11.9213)
E 4.6730(3.1977) 2.3586(3.0415) 1.6087(3.4609) 2.7879(3.3047)

Table Ill. Upper Bounds on Demands and Selling Prices

product projected demand (kg/yr) price ($/kg)

A 268,200 1.114
B 156,000 0.535
D 189,700 0.774
E 166,100 0.224
 162

Xt! ={I if unit type e is assigned in equipment group g

g 0 otherwise

The rest of the variables and the constraints in the model are simplified accordingly.
The results obtained after the solution of the model with no resource considerations [6] are
shown in Table V. The corresponding design configuration depicted in Figure 3 yields a profit
of $516,100 and it shows that the optimal policy is to purchase one unit that must operate in-
phase in stage 4. Note that the second version of the proposed decomposition scheme has
been used to solve this problem, since the values of the integer variables PRjJ; are fixed due to
the multiproduct nature of the plant in consideration. In addition, note that the modeling
language GAMS [1] on an mM 3090 was used for the execution of the program.
Let us now assume that resource restrictions have been imposed on the process. The
set of resources includes steam (denoted by m, cooling fiowrates (eL), electricity (EL) and
manpower (MP). The resource utilization rates are assumed to be posynomial functions of the
split batch size, described by (1.29). The corresponding values of the constants Tlse, 9st! and
~se are presented in Table VI. Assume that no resource expansion is considered at this point,
rather there is a maximum availability for each resource (RS':&X) that is shown in Table Vll.
Therefore, the variable RS s will assume a constant value equal to RS':IX and the resource
expansion term in the objective function will be deleted from the formulation. Details on the
problem size and the computational requirements for the master problem and the NLP
subproblem during the decomposition procedure are given in Table VIII.
The results obtained are shown in Table IX. Note that no expansion of the plant is
suggested due to the imposed resource constraints which are rather restrictive in this case.
Consequently, the profit of $461,400 that can be made during the operation of the plant is
lower compared to the profit of $516,100 that can be made after the expansion of the plant
with the addition of one unit in stage 4. This is due to the fact that, in the former case, the
production level of product D is considerably lower than its upper bound value and, although
no new equipment purchase is made, the revenue due to the production levels is lower than
the increased revenue due to the new production targets in the latter case minus the
investment cost for the additional unit.
Let us now solve a different version of this problem by retaining the same maximum
resource availability RS':&X and by slightly changing selected values of the constants TI.... and
~se (Table X). The proposed decomposition algorithm required three major iterations to attain
the optimal solution which now suggests the purchase of two units that must operate in- and
out-of-phase in stage 1. The results obtained are shown in Table XI and the corresponding
design configuration is depicted in Figure 3. Note that the profit of $504,900 made in this case
is again greater than the profit made with no equipment purchase ($461,400) because the new
production target for product D is increased due to the addition of the in-phase unit in stage 1.
Finally, we solve a third version of the problem in which we consider the possibility of
resource expansion. The upper (RS~) and lower (RS,:in) bounds on the utilization level of
resource s, and the resource cost coefficients prss are given in Table XII. The values of the
constants Tlst!, 9se and ~se presented in Table VI are considered in this case. The proposed
 163

Table IV. Capital Cost Coefficients

unit type / cost coefficient 'Ye Ae

R1,R2,R3 0.1627 15280
Ll,L2,L3 0.4068 38200
F1,F2,F3,F4,F5 0.4881 45840
O1,02,G3 0.1084 44573

Table v. Solution with no resource resnictions (only nonzero values)

unit type Ve Ne product nj TL;,h Pj,kg

Rl 4000 A 530.6 6.382 268,200

Ll 4000 B 88.3 6.794 156,000
Fl,F2 3000,3000 1,1 D 122.8 11.921 189,700
01,02 3000,3000 1,1 E 166.6 3.305 142,600

• Net Profit: $516,100

Table VI. Resource Coefficients floe, se ,Ilse a

resource / eq. type Rl,R2,R3 Ll,L2,L3 FI-F5 O1,02,G3

ST 4, l.36e-2, 1 3,3e-2,l 3,3e-2,1.3

CL 4,2e-3,1.2 l,3e-2,O.75 l,3e-2,O.75
EL 3,le-3,L1 2,5e-4,1 2,3.4e-3,L1 2,5e-4,1
MP 3,2e-2,O.75 1,L55e-l,O.3 2,1.83e-5,O.5 l,1.55e-1,O.3

Table vn. Maximum Resource Availability R~

resource Rmax
•
ST 100
CL 50
EL 70
MP 50
 164

(a)

2 groups 1 group

(b)

2 groups
2 groups

Fl

added unit

Figure 3. Design configuration for case with (a) no resource resttictions

and (b) resource resttictions (second version of example).
 165

Table VIll. Decomposition Algorithm Performance

Version Iteration Subproblem Obj. function NoofEqns/V~1 Vars CPU time-(sec)

MINLPI 487,000 361/14719 18.8

NLPI -461,400 108/90 0.3

2 MINLP2 infeasible 363/14719 5.

Total =24.1
2 MINLPI -536,200 361/14719 6.2
NLP1 496,900 116/99 0.5

2 MINLP2 -520,900 362/14719 14.3

NLP2 -504,900 124/104 0.5

3 MINLP3 infeasible 363/14719 29.8

Total =51.3
3 MINLPI 479,300 361/15119 8.1
NLPI -440.000 116/103 0.4

2 MINLP2 462.000 362/15119 17.4

NLP2 422,400 116/103 0.4

3 MINLP3 458,400 363/15119 18.3

NLP3 -456,500 116/99 0.5

4 MINLP4 infeasible 364/15119 13.8

Total =58.9

- IBM 3090

Table IX. Solution with resource restrictions (only nonzero values)

unit type Ve Ne product n·I TL;.h Pj,kg

R1 4000 1 A 530.6 6.382 268,200

L1 4000 1 B 176.5 6.794 156,000
F1,F2 3000.3000 1.1 D 64.9 11.921 54,200
Gl 3000 1 E 194. 3.305 166,100

• Net Profit: $461,400

 166

Table X. Modified Resource Coefficients TJsc, esc , Ilse

resource / eq. type R1,R2,R3 Ll.L2,L3 F1-FS Gl,G2,G3

ST 3,1.36e-2,O.63 3,3e-2,l 3,3e-2,1.3

CL 3,2e-3,1.2 1,36-2,0.75 l,3e-2,O.75
EL 2,le-3,1.1 2,5e-4,1 2,3.4e-3,1.l 2,Se-4,l
MP 3,2e-2,O.7S l,1.55e-I,O.3 2,l.83e-S,O.S l,1.5Se-1,O.3

Table XI. Solution with resource restrictions (case with modified resource coefficients)

unit type Ve Ne product ni TL"h Pi,kg

R1,R2,R3 4000,1029,500 1,1,1 A 467.3 4.739 268,200

Ll 4000 1 B 176.5 6.417 156,000
Fl,F2 3000,3000 1,1 D 179.7 11.921 IS0,200
Gl 3000 1 E 154.4 3.305 166,100

• Net Profit: $504,900

Table XII. Bounds on Resource Utilization and Cost Coefficients

resource RS~ RS~ PfSs

ST 10 150 500
CL 10 100 200
EL 10 140 140
MP 10 100 200

Table XIII. Resource Utilization Levels (case with resource expansion)

resource RS~ RS;

ST 90.62 113.4
CL 19.96 22.2
EL 39.4 41.5
MP 13.7 15.4

1 Case without equipment expansion (initial equipment configuration)

2 Case with equipment expansion (addition one in-phase unit at stage 4)
 167

algorithm required four major iterations to obtain the optimal solution (Table VIII) that
suggests the addition of an in-phase unit at stage 4, similarly to the case without resource
restrictions. A lower profit ($456,500) can be made in this case, however, due to the resource
expansion term added in the objective function. Table XIII shows that the resource utilization
level had to increase in order to accommodate the addition of the new equipment unit in stage
4.
Note that, in all cases, the resource availability led to different plant equipment
inventory expansions. This example shows that the incorporation of resource restrictions into
the retrofit design fonnulation is necessary to more accurately predict the extent of process
modifications during retrofit

Conclusions

The design problem for the retrofit of a general multipurpose plant with resource
restrictions is posed as a nonconvex mixed integer nonlinear program (MINLP) which
accommodates changes in the product demands, revisions in the product slate, addition and/or
elimination of equipment units, batch size dependent processing times and resource
utilization rates and resource expansion. The proposed model is developed as an extension of
the corresponding model for the retrofit design of a general multipurpose plant with no
resource considerations [6].
The complexity of the proposed model makes the problem computationally intractable
for direct solution using existing MINLP solution techniques. Consequently, a fonnulation
specific decomposition strategy is developed, which builds on our earlier developments for
the retrofit design problem with no resource restrictions. The proposed solution strategy
cannot guarantee the global optimal solution due to the nonconvexity of the upper bound
subproblems.
The solution of a test example clearly showed that incorporation of resource restrictions
into the retrofit design formulation has a great influence on the feasibility and quality of
process modifications during retrofit.

Nomenclature

N number of products
E number of batch equipment types
NEQ number of new equipment types
F number of equipment families
K maximum number of campaigns
H total available production time
 168

index on products
m index on tasks
e index on equipment types
g index on equipment groups
k index on campaigns
TAi set of tasks for product i
Pim set of feasible equipment types for task m of product i
Ue set of tasks that can be executed by equipment type e
Lf set of equipment types that belong to equipment family f
SIs set of products using resource s
S2s set of tasks using resource s
Yik amount of product i produced during campaign k
Qi yearly production requirement for product i
Sime size factor of task m of product j in equipment type e
timgk group processing time of task m of product j
in group g during campaign k
t2..."aime, ~ime processing time coefficients
ae , be cost coefficients for equipment type e
Ce , de cost coefficients for eq. type e used in master problem
Pi unit profit for product i
O)ime operating cost coefficient for task m of prod. i in eq. type e
Ve size of units of equipment type e
Ne number of units of equipment type e
Pi production demand for product j
Qime amount of product i produced during task m in eq. type e
X imegk 0-1 assignment variable for task m of product i
in equipment type e in group g during campaign k
number of batches of product i produced during campaign k
number of batches of product i processed by group g during
task m in campaign k
BS imegk split batch size produced during task m of product i
in eq. type e in group g during campaign k
NUimegk number of units of type e that are contained in group g
assigned to task m of product i during campaign k
NGiJnk number of equipment groups assigned to task m of product i
during campaign k
ha limiting cycle time of product i during campaign k
Tk length of campaign k
CAPe total capacity of equipment type e (=Ve N e )
PR ik priority index denoting assignment of product i to campaign k
llsime, 9sime , ~sime resource coefficients
rsimegk utilization level ofresource s by task m of product i
 169

in unit e in group g during campaign k

utilization level of resource s

Appendix I: Original MINLP Formulation (I)

NEQ b N N
min l: ae Ne (Ve) , + l: l: l: CiliIM QiIM -l: Pi Pi l: prss (RS s -RSr;un) (1.1)
e=1 i=1 meTA, eeP.. i=1 seRES
s.t.

K NCr:::
l: l: l: XilMgk ~ 1 i=I •..•• N; meTAi (I.2)
k=1 g=1 eeP..
NCr:::
l: l: XilMgk !;,Ne e=l ••..• E ; k=l •...• K (1.3)
(i.m) e U, g=1
Xjmegk + Ximqjlc !;, 1 't i. k ; m e TAj ; e.q e Pim ; e e Lf
q e Lh ;f-#1 ; g.j=l •...• NGr:t (1.4)

XilMgk + Xim+lIjic + Xim-plqk !;, 2 't i. k; m=2•...• ITAi 1-1 ;p=I ••..• m-l
e e Pim ; e eLf; I e {Pm+1i nPm-qJ
Ie Lh ;f~h; g=l •...• NG~
j=l •.•.• NG:::.~lk ; q=l •...• NG:::.~pk (I.5)
NCr::
Xjmeglc!;, l: l: Xiqegk 't i. k ; m.q e TA j ; e e Pim (I.6)
g=1 eeP.,
Zimgk ~ Zimg+lk 't i. k ; me TA j ; g=I •...• NG~-l (1.7)
Zjmglc ~ Xjmeglc 't i.k; me TAj ; g=l •...• NG~-l ; e e Pjm (I.8)
Zjmglc!;, l: Xjmegk 't i. k; me TA j ; g=I •...• NG~ (1.9)
eeP...

'f i ; me TAj (I.10)

K
l: Yjk =Pj 't i. k (1.11)
k=1
K NCr:::
l: l: NUilMglc BSjmegk nPimgk!;, QiIM 't i ; m e TAi ; e e Pim (I.12)
k=1 g=1
 170

NCr::
L L NUimegk BSimegk npimgk '2 Yik ¥ i, k; m e TAi (I.l3)
g=1 eeP..

timgk
h .. '2 NG imk ¥ i. k; me TAi; e e Pim ; g=l .... ,NG~ (I.14)

o Ximegk + a.il7ll! BS~imegk

timgk '2 time ... ¥ i. k ; me TAi ; e e Pim ; g=I, .... NG~ (I.15)

rc
NC'::t
hi> s x L L L Ximegk ¥ i, k (1.16)
meTA, g=1 eeP..
K
L TksH (1.17)
k=1

Tk '2nik hi! ¥ i, k (1.18)

Ve '2 Sime BSimegk ¥ i. k ; me TAi ; e e Pim ; g=I ..... NG~ (1.19)

NC':t
Ne '2 L L NUimegk ¥ i ; m e TAi ; e e Pim (1.20)
(i.m)e U, g=1

NU imegk '2 Ximegk ¥ i. k ; me TAi ; e e Pim ; g=l ..... NG~ (1.21)

NUimegk ::;; Nr: ax Ximegk ¥ i. k ; me TAj ; e E Pim ; g=I ..... NG~ (1.22)
NC':t
NG imk = L g Wimgk 'f i. k ; m e TAi (1.23)
g=1

Wimgk = Zimgk - Zimg+lk ¥ i, k; m eTAi ; g=l, .... NG~-l (1.24)

Wimgk = Zimgk 'f i, k;me TAi;g=NG~ (1.25)
NCr::
L Wimgk ::;; 1 'f i, k ; me TAi (1.26)
g=1

L NUimegk BSimegk ::;; Blnax 'f i, k ; mE TAi ; g=l, ... ,NG~ (1.27)
eePj,w.
NC':t
nik; '2 L nPimgk 'f i. k ; m E TAi (1.28)
g=1

rsimegk =NUimegk (rtsime + 9sime BSimeg/'-) SERES; i e SIs; m e TA j n S2s

e e Pim ; g=I, .... NG~ ; k=I, .... K (I.29)
NC':t
L L rsimegk S Rs SERES; k=I, .... K (I.30)
ieS I, meTA,1"\S2, eeP .. g=1

BSimegk '2 B i• Ximegk ¥ i, k ; me TAi ; e e Pim ; g=I, ... ,NG~ (I.31)

 171

• Zimgk
nPimgk ~ nimgk 't i, k ; mE TAi ; g=I, ... ,NGr;::t (1.32)

npimgk :S n~ax Zimgk 't i, k ; mE TAi ; g=I, ... ,NGf)3 (1.33)

V~ :SVe $~ax 'rt e (1.34)

o:S N e :S JVfi'ax 'rte (1.35)

PF :S Pi:S rrax 'rt i (1.36)

o:S Qime , Yik :S Pf"u 't i ;m E TAi ; e E Pim (1.37)

O:s NGimk :S max {Nr;ax} "t i, k; m E TAi (1.38)

eeP..

o:S NUimegk :S JVfi'ax "t i, k ; mE TA i ; e E P im ; g=I, ... ,NG~ (1.39)

O::;,T,,::;,H 'rtk (1.40)

o:S h .. :S meTA;
max max {t'/:::}
eeP..
"t i, k (1.41)

~ax ~....
o 0
time :S timgk :S time + Uime -e - 'f i, k; mE TAi ; e E Pim ; g=I, ... ,NG~ (1.42)
Sime
H
O:s nil< , npimg" ::;, -.- "t i, k ; mE TAi ; g=I, ... ,NGr;::t (1.43)
TL ..
vr;ax
o:S BSimegk ::;, -S-.- 'f i, k; mE TAi ; e E P im ; g=I, ... ,NG~ (1.44)
<me

SERES; i E SIs; m E TAi n S2s

e E Pim ; g=I, ... ,NG~ ; k=I, ... ,K (1.45)
SERES (1.46)
"t; i k=l, ... ,K-l (1.47)

PRik ~ Xi Ie lk "t i, k; e E Pi! (1.48)

PRik::;' L Xilelk V i, k (1.49)

eeP;l

v. ~ Ve+1 V e ; e,e+l E Lf (1.50)

N, ~Ne+1 V e ; e,e+l E Lf (1.51)
• t?me
h. = max min {---} "t i, k
.. meTA; ,eP.. NGr;::t
Vrnin
B; = min min - ' -
meTA; eeP.. Sime
"t i
 172

~ax
Bll1ax = min { L N';'ax _e-} "t i. k
meTA; ee P'-"
.. IIJI
I
Sime

Appendix II: Master Problem Formulation (II)

NEQ N N
min L (C e CAPe + de Ne ) + L L L CJ)ime Qime - L pj P j + L prss (RSs -RS~ )(1I.1)
e=1 j=1 meTA.. eeP.. j=1 seRES
(V,;,ax/, _ (vr:in)b,

s.t.

VCimegk ~ Sime NUimegk BSimegk "t i ; mE TA j ; e E Pjm ; g=l •...• NG'{;3 (ll.2)
NC't:I VCimegk Y jk
L L ~ "t i.k ; m E TA j (ll.3)
g=1 eeP.. Sime n~
Y~
n·t ~-- "t i. k (ll.4)
, Bj"ax
N NC't:I
CAPe ~ L L L VCimegk "tk;eEP im (ll.5)
j=1 meTA; g=1
VTax N']'ax H K NC= 1
Qune::; e e L L -.- Ximegk"t i ; m E TA jm ; e E Pjm (ll.6)
Sjme k=1 g=1 TLu.
VCimegk ~ vr: in X jmegk "t i.k; mE TAj; e E P jm ; g=l •...• NG'{;3 (11.7)
VCimegk ::; vr:ax Nr;'ax X imegk "t i.k ; mE TAj ; e E P jm ; g=l •...• NG'{;3 (ll.g)
Y jk ::; Pj"ax PR jk "t i. k (ll.9)
CAPe ~ CAPe+! "t e e.e+1 E Lf (II. 10)
Vr:m Nr:m ::; CAPe::; V,;,ax N';'ax "te (II.l1)
o ::; VC imegk ::; vr:ax Nr;'ax "t i.m.e.g.k (II. 12)
RL ::; Ru (ll.13)

+ (1.2)-(1.12). (1.14)-(1.18). (1.20)-(1.33). (1.35)-(1.49). (1.51)

 173

REFERENCES

1. Brook:, A.; Kendrick:, D.; Meeraus, A.: GAMS, A User's Guide. Redwood City, CA: Scientific
Press 1988.
2. Garfinlc:el, R.S.; Nemhauser, Gl...: Integer Programming. New York:: Wiley 1972.
3. Glover, F.: Improved Linear Integer Programming Formulations of Nonlinear Integer Problems. In
Management Scieoce, 22(4), 455-459 (1975).
4. Kocis, G.R.; Grossmann, IE.: Global Optimization of Nonconvex MINLP problems in Process
Synthesis. In Ind. Eng. Chem. Res., 27, 1407-1421 (1988).
5. Murtagh, BA; Saunders, M.A.: MINOS 5.0 User's Guide. Technical Report SOL 83-20. Stanford
University Systems Optimization Laboraury, 1983.
6. Papageorgili, S.; Reldaitis. G.V.: Retrofittinl!: a General Multipurpose Batch Chemical Plant.
Industrial and Engineering Chemistry Research, Vol. 32, ~45-363 (1993).
7. Tsirulcis. A.G.: Scheduling of Multipurpose Batch Chemical Plants. PhD Dissertation, Purdue
University, W. Lafayette, IN (1991).
8. Vaselenak:, I.A.; Grossmann, IE.; Westerberg, A.W.: Optimal Retrofit Design of Multiproduct
Batch Plants. In Ind. Eng. Chem. Res., 26, 718-726 (1987).
9. Viswanathan. 1.; Grossmann, I.E.: A Combined Penalty Function and Outer Approximation
Method for MINLP Optimization. In Compo Chern. Engng., 14,769-782 (1990).
Design of Operation Policies for Batch Distillation

S. Macchietto and I. M. Mujtaba

Centre for Process Systems Engineering, Imperial College, London SW7 2BY, UK

Abstract: The batch distillation process is briefly reviewed. Control variables, operating decisions
and objectives are identified. Modeling aspects are discussed and a suitable representation for
operations is introduced. Techniques for the dynamic simulation and optimization ofthe operation
are reviewed, in particular the control vector parameterization method. Optimization formulations
and results are presented for typical problems: optimization of a single distillation step, distillation
with recycle of off-cuts, multiperiod optimization, reactive batch distillation and the online
optimization of a sequence of batches in a campaign. Outstanding research issues are identified.

Keywords: Batch Distillation, Modeling, Operation, Dynamic Simulation, Optimization, Optimal

Control.

Introduction

Batch distillation is perhaps one of the oldest unit operations. It was discovered by many ancient
cultures as a way to produce alcoholic beverages, essential oils and perfume, and its basic
operation had been perfected long before the advent of phase equilibrium thermodynamics, let
alone of computer technology. Today, batch distillation is widely used in the production of fine
chemicals and for specialized productions and is the most frequent separation method in batch
processes [49]. Its main advantages are the ability of separating several fractions of a feed mixture
in a single column and of processing several mixtures in the same column. When coupled with
reaction, batch distillation of one or more products permits achieving much higher conversions than
otherwise possible.
Although distillation is one of the most intensely studied and better understood processes in
the chemical industry, its batch version still represents an interesting field for academic and
industrial research, for a variety of reasons: i) even for a simple binary mixture there are many
 175

alternative operations possible, with complex trade-offs as a result of the many degrees of freedom
available, hence there is ample scope for optimization ii) the process is intrinsically dynamic, hence
its optimization naturally results in an optimal control problem, for which problem formulations
and numerical solution techniques are not yet well established. However, advances made both in
dynamic optimization techniques and computing speeds make it possible to consider rather more
complex operation policies iii) advances in plant control make it now feasible to implement much
more sophisticated control policies that was possible with earlier control technology and hence to
achieve in practice any potential benefits predicted iv) finally, batch distillation is also of interest
as merely an excellent representative example of a whole class of complex dynamic optimization
problems.
The purposes of this paper are i) to summarize some recent advances in the development of
optimal operation policies for a variety of batch distillation applications and ii) to highlight some
research issues which are still outstanding. Although, there are obvious interactions between a
batch column design and its operation [52], in the following coverage it will be assumed that the
column design is given a priori and that an adequate dynamic model (including thermodynamic and
physical properties) has been developed. Attention will be focused on the problem of establishing
a priori the optimal values and time profiles of the variables controlling the operation for a given
feed mixture. It is assumed that a suitable control system can be separately designed later for
accurately tracking the optimal profiles predicted. It is in this sense that we talk of "design of
operation policies". The control approach, of establishing the optimal policies on-line in
conjunction with a state estimator, to take into account model mismatch and disturbances, will not
be considered here. Finally, we will concentrate mainly on the optimal operation of a single batch,
rather than of an entire campaign.
The paper is structured as follows: first, a brief reminder is given of the batch distillation
process and of the main control and operation decision choices available. Some representation and
modeling aspects are considered next, followed by simulation issues and optimization issues.
Finally, a set of illustrative examples are given for the optimization of typical batch distillation
problems involving single and multiple distillation steps, the optimization of off-cut recycles, of a
whole batch and of reactive batch distillation. An example is also given of the use of the above
techniques in conjunction with an on-line control system, for the automation of a batch campaign.
Many of the examples summarized here have been presented in detail elsewhere. Suitable
references are given in the main body of the paper.
 176

The Process - A Brief Review

The Batch Distillation Process

The basic operation for processing of a charge in a batch column (a batch) is illustrated with
reference to the equipment in Figure 1 (a general introduction is given in [79]). A quantity of fresh
feed is charged (typically cold) into a still pot and heated to its boiling point. The column is
brought to the right pressure and temperature during an initial startup period, often carried out at
total reflux, during which a liquid holdup builds up in the top condensate receiver and internally
in the column. Initial pressure, flow, temperature and composition profiles are established. A
production period follows when distillate is withdrawn and collected in one or more fractions or
"cuts". The order of appearance of species in the distillate is determined by the phase equilibria
characteristics of the mixture to be separated (for simple distillation, the composition profiles will
follow well defined distillation curves [84, 9]. A typical instant distillate composition profile is
given in Figure 2. The distillate is initially rich in the lowest boiling component (or azeotropic
mixture), which is then progressively depleted. It then becomes richer in the next lowest boiling

Column

Distillate Receivers

Still Pot
Heat (Reaction)

Figure 1. Typical Configuration of a Conventional Batch Distillation Column

 177

A
1. " , . . , - - - -........

Tunc.[1v)

Figure 2. Typical Distillate Composition (mole fraction) Profiles vs. Time, with Fractions Collected

component, etc. Diverting the distillate flow to different receivers permits collecting distillate
product cuts meeting desired purity specifications. Intermediate cuts ("off cuts" or "slop cuts") may
also be collected which will typically contain material not meeting purity specifications. The
operation is completed by a brief shut down period, when the heat supply to the reboiler is
terminated and the liquid holdups in the column collapse to the bottom. The heavy fraction
remaining in the pot may be one ofthe desired products. Valuable constituents in the offcuts may
be recovered by reprocessing the offcut fractions in variety of ways. The column is then prepared
for the next batch. Several variations of the basic process are possible. Additional material may
be charged to the pot during the batch. Reaction may occur in the pot, or sometimes in the entire
column. Esterification reactions are often conducted this way [21, 20, 6]. Vacuum may be applied
to facilitate separation and keep boiling temperatures low, so as to avoid thermal degradation
problems. Two liquid phases may be present in the distillate, in which case the condensate receiver
has the function of a two phase separator. In some cases, the fresh feed is charged to an enlarged
condenser reflux drum which thus acts as the pot and the column is used in a stripping mode
(inverted column), with high boiling products withdrawn from the bottom, as described by
Robinson and Gilliland [77]. Alternative configurations involving feeding the fresh feed in the
middle of the column (which therefore has both stripping and a rectifYing sections) were also
mentioned by [11, 2, 41]. In this paper, attention will be concentrated on the conventional batch
distillation system (Figure 1) since the techniques discussed are broadly applicable to those
alternative configurations with only minor extensions.
 178

Operation Objectives

The purpose of batch distillation is typically to produce selected product fractions having desired
specifications. For each product, these specifications are expressed in terms of the mole fraction
of a key component meeting or exceeding a specified value. Additionally, the mole fraction of one
or more other species (individually or cumulatively) in some fractions should often not exceed
specified values. These quality specifications are therefore naturally expressed as (hard) inequality
constraints.
Additionally, it is typically of interest to maximize the recovery of (the most) valuable species,
to minimize energy requirements and to minimize the time required for all operations (not just for
fresh feed batches, but also for reprocessing off-cuts, if any). Each of these quantities (recoveries,
energy requirements, time) may also have defined upper and/or lower limits. Clearly, there may be
conflicting requirements. We may observe that rather than posing a multi objective optimization
problem, it is much easier to select just one of the desired quantities as the objective function and
define the others as constraints (e.g. maximize recovery subject to purity, time and energy limits).
In fact, the easiest way to combine multiple objectives is to formulate an overall economic
objective function which properly weighs all factors of interest in common, in well understood
monetary terms.

Operation Variables and Trade-ofTs

The maximum vapor boilup and condensate rate that can be produced for a given feed mixture and
the liquid holdups in the column are essentially established by the column design characteristics,
fixed a priori (column diameter, number of equilibrium stages and column internals, pot, reboiler
and condenser type and geometry). For a given charge, the main operation variables available for
control are the reflux ratio, the heating medium flow rate (or energy input to the reboiler or vapor
boilup rate, varied by means ofthe energy input to the reb oiler), the column top pressure and the
times during which distillate is collected in each of the different distillate receivers. Specifying all
these variables determines the amount and composition of each of the fraction collected (hence
recoveries) and other performance measures (e.g. total time, energy used, etc.). A number of
trade-offs must be considered.
Increasing reflux ratio will increase the instant distillate purity, giving a smaller distillate flow
rate and thus requiring longer time to produce a given amount of distillate, with higher energy
 179

requirements for reboiler and condenser. On the other hand, a larger amount of distillate meeting
given purity specifications may be collected this way. Productivity (in terms of the amount of
distillate produced over the batch time) may go down as well as up with increasing reflux ratio,
presenting an interesting optimization problem. A useful upper bound for the top distillate
composition achievable at anyone time is given by the total reflux operation.
Traditionally, constant reflux ratio (on grounds of simplicity of application) and constant
distillate purity (requiring progressively increasing reflux ratio) have been considered. In order to
achieve a given final composition in a specific accumulated distillate fraction, initially richer
distillate may be mixed with lower quality distillate near the end of the production cut. This is
obtained with the constant reflux policy. Thermodynamic considerations suggest that any mixing
is a source of irreversibility and hence will have to be paid for somehow, providing a justification
for the constant distillate composition operation. This argument however ignores the relative costs
of product, column time and energy. In general, some intermediate reflux ratio policy will be
optimal. Other reflux ratio strategies have been used in practice, for example one characterized by
alternating total reflux (no distillate) and distillate product withdrawal (no reflux) [5].
With regards to the vapor hoilup rate, it is usually optimal to operate at the highest possible
energy input, except when hydraulic considerations become limiting (entrainment). When
hydraulics is not considered (at one's own risk), the optimal policy can thus be established a priori.
A constant (top) pressure level is often selected once for the entire batch distillation, or
different but constant pressure levels, if necessary, may be used during each production cut.
Pressure may be decreased as the batch progresses as a way to maintain the boiling point in the pot
below or at a given limit for heat labile mixtures or to increase relative volatility for difficult
separations.
The choice ofthe timingfor each production cut is important. With reference to Figure 2, two
rather obvious points may be noted. First, it is possible to achieve a high purity specification on one
component in an individual fraction by choosing its beginning and end time so as to remove the
lower purity front and tail, respectively, in the previous and in the subsequent cut. This, however,
may make achieving any specifications on the earlier and subsequent cuts much harder and even
impossible. Thus, the operations of all cuts are interacting rather strongly. Second, as already
observed, it is possible to eliminate fronts and tails of undesired components as off-cut fractions.
This however will affect recoveries and will make offcut reprocessing important.
With respect to the o.ffcuts, several choices are available. A fraction not meeting product
 180

specifications may be simply a waste (with associated loss of material and possibly a disposal cost)
or it may be a lower quality by-product, with some residual value. Offcuts may also be reprocessed.
Here, several alternatives are possible. The simplest strategy is to collect all off-cuts produced
during one batch in a single distillate receiver and then add the entire amount to the pot, together
with fresh feed, to make up the next charge. This will increase recovery of valuable species, at the
expense of a reduction in the amount of fresh feed that can be processed in each batch, leading
again to an interesting trade-off[51]. For each offcut, the reflux ratio profile used and the duration
of tile off-cut production step (hence, the amount and composition of the accumulated offcut) must
be defined. The addition of the offcuts from one batch to the pot may be done either at the
beginning of or during the subsequent batch, the time of addition representing one further degree
of operational freedom. The second principle ofthermodynamics gives us again useful indications.
Since mixing is accompanied by irreversibility, the addition of an offcut to the pot should be done
when the compositions of the two are closest. This also suggest an altogether different
reprocessing policy, whereby distinct off-cut fractions from a batch charge are not mixed in the
same distillate receiver, but rather collected in separate receivers (assuming sufficient storage
capacity is available). Each off-cut fraction can then be individually recycled to the pot at different
times during the next batch. In fact, re-mixing of fractions already separated can also be reduced
if the same off-cut material produced in successive batches is collected (in a much larger distillate
storage) and then reprocessed later as one or more batches, with the full charge made up by the
stored off-cut [59, 73]. This strategy is motivated by the fact that each off-cut is typically rich in
just a couple of components, and hence their separation can be much easier this way. In practice,
the choice of a reprocessing policy will depend on the number and capacity of storage vessels.
It may also be noted that unlike its continuous counterpart, a batch distillation column can give
a very high separation even with few separation stages. If a desired purity cannot be achieved in
one distillation pass, an intermediate distillate off-cut can be collected and distilled a second (and
third, etc.) time.
In summary, we may identify two types of operation choices to be made. Some choices define
the structure of the operation (the sequence of products to be collected, whether to produce
intermediate cuts or not, whether to reprocess off-cut fractions immediately or store them for
subsequent distillation, whether to collect off-cuts in a single vessel, thereby re-mixing material
already separated, or to store them individually in distinct vessels. These are discrete (yes/no)
decisions. For a specific instance of these decisions (which we will call an operation strategy),
 181

there are continuous decision variables, the main ones being the reflux ratio profiles and the
duration of each processing step, with possibly pressure and vapor boilup rate as additional control
variables.
Even for simple binary mixtures, there are clearly many possible operation strategies. With
multi component mixtures, the number of such structural options increases dramatically. For each
strategy, the (time dependent and time independent) continuous control variables are highly
interrelated. Selecting the best operation thus presents a rather formidable problem, the solution
of which clearly requires a systematic approach. Formally, the overall problem could be posed as
a mixed integer nonlinear dynamic process optimization problem (MINDPOP), with a general
economic objective function and all structural and continuous operation variables to be optimized
simultaneously. So far, however, only very much simpler subsets of this problem have been
reported. The operation strategy is typically fixed a priori, and just (some) continuous operation
variables are optimized. This approach will be followed in this paper as well. Fixed structure
operations for which optimization solutions have been proposed include the already mentioned
minimum time problem (PI), maximum distillate problem (P2), and maximum profit problem (P3)
(Table 1).
Table 1. Selected References on a priori Control Profile Optimization-Conventional Batch Distillation
Columns

Reference Column Model Mixture Pha~e Equilibria Optimisation Problem

Converse & Gross (1963) Simple l3inary CRY P2

Converse & Huber (1965)
Coward (1967) PI
Robinson (1969)
Robinson (1970) Multicomponent
Mayur and Jackson (1971)
Kerkhof and Yissers (1978) l3inary P3
Murty et al. (1980) P2
Hansen and Jorgensen (1986) PI
Diwekar et aI. (1987) Simple l Multicomponent P2
Mujlaba (1989) Rigorous Rigorous PI/P2/P3
Fahrat el al. (1990) Simple' Simplc P2-multipcriod
Mujtaba and Macchiello (1991) l3imlY Rigorous PI
Diwekar and Madhavan (1991) P3
Logsdon & l3iegler (1992) Simple t CRY P2
Jang (1992) Rigorous Multicomponent Rigorous PlIP2
Diwekar (1992) Simple 2 CRY PlIP2/P3
Mujtaba and Macchiello (1992b)Rigorous Rigorous P3-multiperiod

CRY = Constant Relative Yolatility. I - short-cut model of continuous distillation. 2 - same as 1 but modified
for column holdup and tuned for nonideality. 3.- short-cut modcl, no holdups.
 182

Benefits

That improved perfonnance batch distillation should be achieved by clever manipulation of the
available operation parameters (relatively to simpler, constant control) is intuitively appealing.
However, reviews of previous optimization studies have indicated that benefits are often small
[78]. This argument has been used to dismiss the need for sophisticated operation policies. On the
other hand, these results were often obtained using highly simplified models and optimization
techniques, a limited subset of alternative policies (constant distillate composition vs. constant
reflux ratio) and objective functions only indirectly related to economic perfonnance. Different
benefits are calculated with different objective functions. Kerchief and Vassar [46], for example,
showed that small increases in distillate yield (order 5%) can translate into 20-40% higher profit.
This clearly calls for more comprehensive consideration of operating choices, dynamic models,
objective functions, and constraints.

Representation and Modeling Issues

Representation

There is a need to fonna1ize the description of the operating procedures. It may be convenient to
consider a batch distillation operation as composed by a series of steps, each terminated by a
significant event (e.g. completion of a production cut and switch to a different distillate receiver).
Following [64] the main structural aspects of a batch distillation operation are schematically
represented here as a State Task Network (STN) where a state (denoted by a circle) represents a
specified material, and a task (rectangular box) represents the operation (task) which transfonns

Main-cut Product i

Initial Boltom Intermediate Intermediate

Charge Residue Residue i-I Residue i

Figure J. a) State Task network for a Simple Batch Distillation Operation Producing Two Fractions
b) Basic Operation Module for Separation into One Distillate and One Residue Fractions
 183

the input state(s) into the output state(s). A task may consist of one or more steps. For example,
Figure 3 shows a simple batch distillation operation with one task (Step 1) producing the two
fractions Main-cut 1 and Bottom Residue from the state Initial Charge.
States are characterized by a name an amount and a composition vector for the mixture in that
state. Tasks are characterized by an associated dynamic model and operational attributes such as
duration the time profiles of reflux ratio and other control variables used during the task etc.
Additional attributes of a distillation task are the values of all variables in the dynamic model at the
beginning and at the end of the step. The states in the STN representation should not be confused
with any state variables present in the dynamic model. For example, in Figure 1 the overall amount
and composition (Bo and "Eo respectively) of the Initial Charge STN state are distinct from the
initial pot holdup and composition. The latter may be assigned the same numerical value as the
former (a specific model initialization procedure). It is also possible for simulation and optimization
purposes to neglect the startup period altogether and initialize the distillation task Step 1 by
assuming that some portion of the charge is initially distributed along the column or that the initial
column profiles (holdups composition etc.) are those obtained at total reflux (two other distinct
initialization procedures). Of course, whatever initialization procedure is used the initial column
profiles must be consistent (i.e. mass balance) with the amount and composition of the Initial
Charge state in the STN. The initialization procedure is therefore a mapping between the STN
states and the initial states in the dynamic model for that task. Similarly the amount and
composition BI and xBI of the STN state Bottom Residue are not the holdup and composition of
the reboiler at the end of Step I but those which are obtained if all holdups in the column at the end
of Step 1 are collected as the Bottom Residue STN state. The STN representation originally
proposed in. [45] for scheduling is extended by the use of a dynamic model for a task in the place
of a simple fixed time split fraction model.
The advantages of the above representation are: makes the structure of the operation quite
explicit ii) it enables writing overall mass balances around an individual task (or a sequence of
tasks) iii) suitable definition of selected subsets of states and task attributes make it possible to
easily define a mix of dynamic simulation and optimization problems for a given operation strategy
(STN structure) iv) different initialization procedures or even model equations may be defined for
different tasks v) it enables the easy definition of alternative operation strategies by adjoining a
small set of basic task modules used as building blocks. As an example the batch distillation of a
multi component mixture is represented in Figure 4 as the combination of 2 main distillate and 1
 184

Main-cut Off-cut Main-cut 2

Initial Charge Int. Res. Int. Res. 2 Bottom Product

Figure 4. STN for Batch Distillation Operation with Two Main Distillate Cuts and One Intermediate Off-
cut

off-cut production steps each of these steps being represented by essentially the same generic
"module" given in Figure 3b. Similarly, an operation is shown in Figure 5 consisting of a distillate
product cut (Step 1) followed by an off-cut production step (Step 2). The off-cut produced in a
batch (state Off-cut Recycle of amount R and composition XR) is recycled by mixing it in the pot
residue with the next batch immediately before performing the main cut step to give an overall
mixed charge amount Bc (of composition xBC)- This operation (with states and tasks suitably
indexed as in Figure 3b.) can be used as a building block to define for example the cyclic batch
operation strategy for a multicomponent mixture defined in Figure 6 (states omitted) [63].
Main-cut

previous h~lch current hatch

Offcut Recycle
Figure 5. Basic Module for Separation Operation into One Distillate and One Residue Fractions, with Recycle
of an Intermediate Off-cut to the next Batch

Modeling

Distillation modeling issues have been extensively discussed both for continuous and batch
applications (e.g. [27,44,69,4]) and will not be reviewed here in detail. The models required for
 185

Figure 6. Operation Strategy for Multicomponent Batch Distillation with Separate Storage of Off-cuts and
Sequential Off-cut Recycle to the Next Batch

batch distillation are in principle no different than those required for continuous distillation
dynamics. What is of interest is the ability to predict the time responses of temperatures
compositions and amounts collected for generic mixtures, columns and operations. It may be
argued that models for batch distillation must cover a wider operations range, since startup and
shutdown are part ofthe normal operation. For optimization purposes, the models must often be
integrated many times and, hence, speed of execution is important. Some issues particularly
relevant to batch columns are:
Modeling detail - Short-Cut vs. "Rigorous". As with any modeling exercise, a balance must
be struck between the accuracy of the predicted responses availability of data, and speed of
execution. Therefore, the "right" level of "rigorousness" depends on advances in thermophysical
property predictions, numerical algorithms and computer hardware, and on purpose of use. In the
past, most work on batch distillation used fairly simple short-cut models which relied on a number
of assumptions such as constant relative volatility equimolar overflow no holdup and no hydraulic
models etc. When used, the dynamic mass and energy balances have also been simplified in many
ways. In some cases, accumulation terms have been neglected selectively or even altogether with
dynamics approximated by a sequence of steady state calculations (e.g. [67, 35] for recent
examples). These simplifications were dictated by the available technology and lead to useful
results (e.g. [28]). At the other end of the spectrum, full dynamic distillation models have been
proposed with very detailed phenomena described (e.g. [80]).
Without entering into a lengthy discussion, it appears that at present fairly "rigorous" dynamic
models with full dynamic mass and energy balances for all plates and thermophysical properties
predicted from generic comprehensive thermophysical packages can be readily integrated for use
in batch simulation and optimization. The main justifications for shortcut models (simplicity, speed
 186

of solution, and ability to tailor special solution algorithms) appear to be less and less valid
particularly in the light of the effort needed to validate the short cut approximations. Simplified
thermodynamic and column models may still be necessary for mixtures for which a relative
volatility is all that can be estimated, or for very demanding optimization applications (e.g. [48]).
The desirable approach to follow, however, is no doubt to develop simulation and optimization
tools suitable for as generic a model form as possible, but to leave the user the possibility to adopt
simplifYing assumptions which may be appropriate for specific instances.
Production period. Standard continuous dynamic distillation models can be used, represented
by the general form:

f(x, x', t, u, v) = 0 (eq. 1)

with initial conditions "'0= x( to) and x'o= x'( to ). In eq. 1, f is a vector of differential and algebraic
equations (DAEs), (mass and energy balances, vapor liquid equilibrium relations, thermophysical
property defining equations, reaction kinetic equations, etc.), x and x' are vectors of (differential
and algebraic) state variables and their time derivatives (differential variables only), t is the time,
u is a vector of time varying control variables (e.g the reflux ratio) and v is a vector of time
independent control variables (e.g. pressure).
A modeling issue particularly important for batch distillation regards the treatment ofliquid
holdups. Holdups may have quite significant effects on the dynamic response of the column (e.g.
[19, 62]), and therefore should be considered whenever possible (zero liquid holdup means that
a composition change in the liquid reflux reaches the still infinitely fast, clearly not the case).
Assumptions of constant (mass, molar or volume) holdup are often used, and these already account
for the major dynamic effects (delay in composition responses). Where necessary, (Le. low pressure
drop per plate) more detailed hydraulic models for plates and downcomers should be used. The
extra information added to the model should be considered in light of the low accuracy often
attached to generic hydraulic models and other estimated (guessed?) quantities (e.g. Murphee
efficiencies). High quality relations, regressed from experimental data, are however often available
for proprietary applications and their use is then justified. Vapor-liquid equilibrium has been
invariably assumed so far for batch distillation, but there is no reason in principle why rate models
cannot be used.
Similar arguments apply to the modeling of heat transfer aspects (reboiler, heat losses) and
control loops. If the heat exchange area decreases as the pot holdup decreases, or the heat transfer
 187

coefficients vary significantly during a batch, for example, the assumptions of constant heat supply
or constant vapor boilup rate may no longer be valid and it may be necessary to use a more detailed
heat transfer model, accounting for the reboiler geometry [3]. Ifperfect control action (say, on
the reflux ratio) is not a good assumption, equations describing the control loop dynamics may be
included in eq. 1. In this case, the controller set-points will become the manipulated quantities and
possible optimization variables.
Startup period. This may be defined as the period up to the time when the first distillate is
drawn. It may be divided into two steps. In the first step, the column fills up and some initial
profiles are established. In the second step total reflux is used until the top distillate is suitably
enriched. For the first step, a model must be able to describe the establishment of initial liquid and
vapor profiles, holdups and compositions along the column and in the condenser from an empty,
cold column. Flows may be intermittent and mixing behavior on the plates and in the downcomers
will be initially very different from the fully established situation, with vapor channeling weeping
of liquid, sealing of downcomers, etc. This would demand accurate representation of hydraulic and
phase behavior in rather extreme conditions. The use of vapor-liquid models based on perfect
mixing and equilibrium under these circumstances is clearly suspect. Thermal effects due to the
column being initially cold may be as important as other effects and the exact geometry of the
column is clearly important [57]. Some of these aspects have been modeled in detail in [80], based
on work for continuous distillation [36].
What is more usually done is to assume some far simpler mechanism for the establishment of
the initial vapor and liquid profiles which does not need a detailed mechanistic model. For example:
i) finite initial plate and condenser liquid holdups may be assigned at time zero at the same
conditions as the pot (boiling point) (e.g. [19,48]): ii) the column is considered initially (as a single
theoretical stage, with the vapor accumulating in the condenser receiver at zero reflux. When
sufficient liquid has been accumulated, this is redistributed as the internal plate holdups, filling the
plates from the top down (e.g. [50,38]) or from the bottom up [3]).
For the second startup step the same dynamic model may be used for the production steps.
From an operation point of view, the practical question is whether a total reflux operation should
be used at all, and if so, for how long. The duration of the total reflux operation (if at all necessary)
can be optimized [19, 59].
An even cruder approximation is to ignore the startup period altogether, consider it as an
instantaneous event and initialize the column model using the assumption of total reflux,
 188

steady-state (or as in [4], at the steady-state corresponding to a finite reflux ratio with no distillate
production, obtained by returning the distillate to the pot). Of course, this procedure does not
permit calculating the startup time.
Clearly, different startup models will provide different starting values for the column dynamic
profiles in the first production step. Whether this matters was considered [I], comparing simulation
results with four different models of increasing complexity for the first startup a stage. In all four
cases, the procedure was followed by a second startup step at total reflux until stable profiles were
achieved (steady state). The main conclusions were that the (simulated) startup) time can be
significant (especially for difficult separations), is essentially due to the second step, is roughly
proportional to the major holdups (in the condenser drum and, for difficult separations, those in
the column) and that all four models gave approximately the same results (startup time and column
conditions).
In the absence of experimental data to confirm or reject either approach, the use of the simpler
procedures (i and ii above) to model the first step of the startup period would appear to be a
reasonable compromise.
Shut down period This period is typically not modeled in detail, since draining of the column
and condenser following interruption of heating is usually very rapid compared to all other periods
and separation is no longer affected. However, the final condenser receiver holdup may be mixed
with the last distillate fraction, collected separately or mixed with the bottoms fraction, thus the
exact procedure followed should be clearly defined (this is clearly irrelevant when holdups are
ignored).
Transfers, additions, etc. Fresh feed or recycle addition are often modeled as instantaneous
events. Adding side streams with finite flow rates to a generic column model, if required, is
however a simple matter. These additional feeds (if any) may then be chosen as additional control
variables.
In principle, different models of the type defined by eq. 1 may be defined for distinct distillation
tasks. For example, referring to the operation in Figure 4, the number of species present in the final
separation task (Step 3) may involve only a small subset of the species present in the initial charge.
For numerical reasons, it may well be better to eliminate the equations related to the depleted
species, giving a different set of equations for Step 1 and Step 3. Other modeling formulation
details affect the numerical solution. For example, we found it is better to use an internal reflux
ratio definition (LN, with range 0-1 ) rather than the usual external one (LID, ranged O-infinity).
 189

Simulation Issues
The simulation problem may be defined as follows:

Given: A DAE batch distillation model f(x, x', t, u, v)=O (eq. 1)

Values for all control variables uCt), v
Initial conditions
Tennination conditions based on tr, x(tr), x'(tr), u(tr)
Calculate: Time profiles of all state variables x(t), x'(t)
Dynamic models used in batch distillation are usually stiff. The main techniques used to
integrate the above DAE system are BDF (Gear's) methods [37, 43] and orthogonal collocation
methods [85]. Discontinuities will typically be present due to discrete events such as the
instantaneous change in a control variable (e.g. a reflux ratio being changed between two values)
and the switching of receivers at the end of a step. The presence of algebraic equations adds
constraints between the state variables and their derivatives (and a number of complications). With
respect to the initial conditions, only a subset of all Xo and x' 0 variables may then be fixed
independently, the remaining ones having to satisfy eq. 1 at the initial time (consistent
initialization). A similar situation occurs with the discrete changes. If an algebraic variable can only
be calculated from the right hand side ofa differential equation (for example, the vapor leaving the
stage from the energy balance), then some algebraic equation may have to be differentiated before
the algebraic variable can be calculated (the number of differentiations required being called the
index of the DAE system). A number of ad-hoc solutions had been worked out in the past to
overcome these problems. For example, the derivative term in the energy balance could be
approximated (e.g by a backwards finite difference, yielding an algebraic equation). To avoid
discontinuities when changing reflux ratios, a continuous connecting function may be used [25].
These aspects are now far better understood, and even if it is still not possible to always solve
higher index systems, it is usually possible to reformulate the equations so as to produce index 1
systems in the first place and to address implicit and explicit discontinuities. Requirements related
to the consistent definition of the initial conditions, solvability of the system, DAE index, model
formulation so as to avoid index higher than one and (re)initialization procedures with
discontinuities were discussed in particular by [69, 4].
General purpose simulation packages such as BATCHES [17] and gPROMS [8] allow the
process engineer to build combined discrete-event/differential algebraic models for simulation
 190

studies and have built in integration algorithms dealing with the above issues. In particular, they
are able to deal effectively with model changes between different stages and discrete events. They
are therefore suitable for general batch distillation simulations. General purpose dynamic simulators
for essentially continuous systems such as SPEEDUP [68] can also be used, although the ability
to handle generic events and model changes is more limited. A number of codes have also been
developed specifically for batch distillation simulation (e.g. [4,31, 35]) or adapted from continuous
distillation [80]. A batch simulation "module" is available in several commercial steady state
process simulators (e.g. theBATCHFRAC program [12], Chemcad III, ProsimBatch [71]). These
are tailored to handle the specific events involved (switching of receivers, intermediate addition of
materials, etc.) and in general use a single equation model for the entire batch operation and
generic thermophysical property packages.

Optimization Issues

As noted, there are both discrete and continuous operation decisions to be optimized. At present
dynamic optimization solutions have only been presented dealing with problems with fixed
operation structure and with the same model equations for all separation stages. In the following
we will therefore consider the optimization of the continuous operating decisions for an operation
strategy selected a priori. Some work on the optimization of the operations structure is briefly
discussed in the last section.
Equations and variables may be easily introduced into eq. 1 to define integral performance
measures, for example, the total energy used over a distillation step. This permits calculating
additional functions of all the states, controls, etc. and to define additional inequality constraints
and an objective function, in a general form:
Inequality constraints g = g (tp x(tr), x'(tr), u, v) (eq.2)
Objective function J=J(tp x(tr), x'(tr), u, v) (eq. 3)
The optimization problem may then be defined as follows:
Given: Initial conditions to' Xo and x'o
Find: Values of all control variables v, u(t)
So as to: Minimize the objective function MinJ (eq. 3)
Subject to: Equality constraints (DAE model) f(x, x', t, u, v) =0 (eq. 1)
Inequality constraints g (tp x(tr), x'(tr), u, v) ~ 0 (eq.2)
 191

Upper and lower bounds may be defined on the control variables, u(t) and v, and on the final
time. Termination conditions may be implicitly or explicitly defined as constraints. Additional
inequality constraints may be defined for state variables not just at the end, but also at interior
points (path constraints). For example, a bottom temperature may be bounded at all times. Some
initial conditions may also have to be optimized.
The main problem in this formulation is the need to find optimumfonctio11S' u(t) i.e. an infinite
set of values of the controls over time. The main numerical techniques used to solve the above
optimal control problem are the control vector parameterization (CVP) method (e.g. [58, 34]) and
the collocation method (e.g. [10,26, 74, 48]. Both transform the control functions into discrete
forms approximated by a finite number of parameters.
The CVP method (used in this paper) discretizes each continuous control function over a finite
number, defined a priori, of control intervals using a simple basis function (parametric in a small
number of parameters) to approximate the control profile in each interval. For example, Figure 7
shows a piecewise constant approximation to a control profile with 5 intervals. With the initial time
given, two parameters are sufficient to describe the control profile in each interval a control level
and the final time of the interval. Thus the entire control profile in Figure 7 is defined by 10
parameters. These can then be added to any other decision variable in the optimization problem
to form a finite set of decision variables. The optimal control problem is then solved using a nested
procedure: the decision variables are set by an optimizer in an outer level and, for a given instance
of these variables, a dynamic simulation is carried out to calculate objective function and
constraints (eqs. 1-3). The outer problem is a standard (small scale) nonlinear programming
problem (NLP), solvable using a suitable method such as Sequential Quadratic Programming
(SQP). Parameterizations with linear, exponential, etc. basis functions may be used [34]. Since the

u(t)

Figure 7. Piecewise Constant Discretization of Continuous Control Function

 192

DAEs are solved for each function evaluation, this has been called a feasible path approach. Its
main advantage is that the approach is very flexible. General purpose optimizers and integrators
may be used, and the set of equations and constraints may be very easily changed.
The collocation methods discretize both the control functions and the ordinary differential
equations in the original DAE model, using collocation over finite elements. The profiles of all
variables are approximated using a set of basis functions, with coefficients calculated to match any
specified initial conditions, final conditions and interior conditions (additional conditions being
provided by the continuity of profiles across the finite elements). The end result is a large system
of algebraic equations, which together with constraints and objective function form a large NLP
problem. For the same problem, however, the degree of freedom for this large scale NLP is the
same as for the small scale NLP in the CVP approach. The optimization may be solved using
suitable NLP algorithms, such as a SQP with Hessian decomposition [48]. Since the DAEs are
solved at the same time as the optimization problem, this has been called an infeasible path
approach. The main advantage claimed for this approach is that it avoids the repeated integrations
of the two level CVP method, hence it should be faster. However, comprehensive comparisons
between the collocation and the CVP methods have not been published, so it is not possible to
make definite statements about the relative merit.
The path constraints require particular treatment. Gritsis [39] and Logsdon and Biegler [48]
have shown that they can be handled in practice both by the CVP and orthogonal collocation
methods (in particular, by requiring the constraints to hold at a finite number of points, coincident
with control interval end points or collocation points). A general treatment of the optimal control
of constrained DAE systems is presented in [70].
Computer codes for the optimization of generic batch distillation operations have been
developed by [59, 31]. To my knowledge, no commercial code is presently available.

Application Examples

In this section, some examples are given of the application of the ideas and methods outlined
above, drawn from our own work over the last few years. The examples are used to illustrate the
current progress and capabilities, in particular with respect to the generality of operation strategies,
column and thermodynamic models and objective functions and constraints that can be handled.
 193

Solution Models and Techniques

All examples were produced using a program developed specifically for batch distillation simulation
and optimization by [68] and successively extended to handle more features. For a particular
problem, a user supplies the definition of a batch column configuration (number of theoretical
stages, pot capacity, etc.), defines the fresh charge mixture to be processed and selects a
distillation model and thermodynamic options. Two column models have been used, mainly to
show that the solution techniques are not tied to a specific form of the model. The simpler column
model (MC 1) is based on constant relative volatility and equimolar overflow assumptions. A more
rigorous model (MC2) includes dynamic mass balances and general thermodynamics, with the
usual assumptions of negligible vapor holdup, adiabatic plates, perfect mixing for all liquid holdups,
fixed pressure and equilibrium between vapor-liquid. A total condenser is used with no sub-cooling
and finite, constant molar holdups are used on the plates and in the condenser receiver. To maintain
solution time reasonably low, the energy balances are modeled as algebraic equations (i.e. energy
dynamics is assumed to be much faster than composition dynamics). Thermodynamic models,
vapor liquid equilibria calculations and kinetic reaction models are supplied as subroutines (with
analytical derivatives, if available). Rigorous, general purpose thermodynamic models, including
equations of state and activity coefficient models may therefore be used for nonideal mixtures. A
simple constant relative volatility model (MT 1) may still be used, if desired, in conjunction with
the more rigorous dynamic model MC2.
A desired operation strategy (number and sequence of product cuts, off-cuts, etc) is defined
a priori. A simple initialization strategy is used for the first distillation task. The fresh charge is
assumed to be at its boiling point and a fraction of the initial charge is distributed on the plates and
in the condenser (according to the specified molar holdups). For successive tasks, the initial column
profiles are initialized to be the same as the final column profiles in the preceding task. Adjustments
may be made for secondary charges and to drop depleted component equations from the model.
For each SIN task, the control variables are identified and the number of discretization control
intervals is selected (a CVP method with piecewise constant parameterization is used) and initial
values are supplied for all control levels and control switching times. Reflux ratio, vapor boilup rate
and times are available as possible control variables (additional quantities, if present, may also be
used for control, such as the rate of a continuous feed 10 an intermediate stage). A variety of
specifications may be set for individual distillation steps, including constraints on selected end
 194

purities and amounts collected (other specifications are described in the examples). Finally, an
objective function is selected, which may include those traditionally used (max distillate, min time),
one based on productivity (amounts over time) or on economics (max profit). In the latter case,
suitable cost coefficients must also be supplied for products, off-cuts, feed materials and utilities.
From these problem definitions, programs are generated for computing all required equation
residuals and the analytical Iacobians, together with a driver for the specific
simulation!optimization case.
A robust general purpose SQP code [15] is used to solve the nonlinear programming
optimization. A full matrix version is used here, although a sparse version for large scale problems
(with decomposition) is also available. The DAEs system is integrated by a robust general purpose
code, DAEINT, based on Gear's BDF method. The code includes procedures for the consistent
initialization of the DAE variables and efficient handling of discontinuities. The gradients needed
by the SQP method for optimization are efficiently calculated using adjoint variables [58], requiring
the equivalent of approximately only two integrations of eq. 1 to calculate all gradients (an
alternative, method for generating the gradients would be to integrate the dynamic sensitivity
equations alongside the model equations, [13]. Outputs are in tabular and simple graphical form.

Simulation and Sequential Optimization ofIndividual Distillation Steps

The first example deals with a four component mixture (propane, butane, pentane and hexane) to
be separated in a 10 equilibrium stage column and was initially presented in [12]. The operation
consists of 5 distillation steps with two desired products. A fraction with high propane purity is
collected first, followed by a step where the remaining propane is removed. A high purity butane
fraction is then collected, with removal of pentane in an off-cut in the fourth step. The final step
removes the remaining pentane in the distillate and leaves a high purity hexane fraction as the
bottom residue. A secondary charge is made to the pot after the second step (Figure 8).
Boston et al. presented [12] simulation results for an operation with constant reflux ratio
during each step and different values in distinct steps. Simulations of their operation (same reflux
ratio and duration for each step) were carried out in [59] with two thermodynamic models, MT2
(an ideal mixture model with VLE equilibrium calculated using Raoult's law and Antoine's model
for vapor pressure, ideal gas heat capacities for vapor enthalpies from which liquid enthalpies were
obtained by subtracting the heat of vaporization) and MT3 (SRK equation of state). Results with
the former thermodynamic model and with the more rigorous dynamic column model Me2 are
 195

C3 all C4 prod C50ff

/
" /
/
"

Figure 8. Batch Distillation of Quatemmy Mixture [I2l-Operating Strategy. The two distillation tasks CutI
and Cut3 are composed of two off-cut production steps each

reproduced in Table 2. They are very similar to those reported in the original reference, with
inevitable minor differences that are due to difference in thermodynamic models, integration
tolerances, etc. This operation was taken as a base case.
Mujtaba [12] considered the optimization of the same process, using the above operating
policy as a base case, as follows. In the absence of an economic objective function in the original
reference, it was assumed that the purities obtained in each of the cuts (in terms of the molar
fraction of the key component in that cut) were the desired ones and that the reflux ratio policy
could be varied. A variety of objective functions were defined. Table 3 reproduces the results
obtained for the first distillation step with three objective functions: minimum time, maximum
distillate and maximum productivity, with an end constraint on the mole fraction of propane in the
accumulated distillate ("Propane = 0.981). All results are reported in terms of a common measure
of productivity, the amount of C3 off-I distillate produced over the time for Step 1 and for two
ways of discretizing the control variable, into a single interval and five intervals, respectively. As
expected, the productivity of this Step varies depending on the objective function used, increases
when more control intervals are used and is maximum when productivity itself is used as the
objective function. As a comparison, the base case operation had a productivity of2.0 Ibmollhr for
Step 1 (8.139 Ibmole produced in 4.07 hr). With a single reflux ratio level, the minimum time
problem and the maximum distillate problem resulted (within tolerances) in the same operation as
 196

Table 2. Batch Distillation of Quaternary Mixture [12]. Simulation Results with MC2 Column Model
(dynamic, more rigorous) and MT2 Thermo Model(ideal).Total Time Tf=30.20 hr; Overall
Productivity (A+B)rrf=3.17IbmolJhr

Column
No. of Internal Stages (ideal) 8
Condenser total no suhcooling
Liquid Holdup - stage (lbmol) 4.93 10- 3
Liquid Holdup - condo (lhmol) 4.93 10- 2
Charges Fresh feed ScconcL-uy
(1) Propane 0.1 0.0
(2) Butane 0.3 0.4
(3) Pentane 0.1 0.0
(4) Hexane 0,5 0.6
Amount (lbmol) 100 20
at time initial after Step2
Operation SpeciFied
Task Stepl Step2 Cut2 Step4 StepS
Products C3 off-I C30ff-2 C4prod C50ff-1 C50ff-2
C6prod
Specified:
ReOux ratio (external) 5 20 25 15 25
Time (hr) 4.07 1.81 18.27 4.31 1.78
Distillate rate (lhmollhr) 2 2 2 2 2
Pressure (bar) 1.03 1.03 1.03 1.03 1.03
Operation - Results
Top Vapour raiC (lhmollhr) 12 42 52 32 52
Instant Distillate (mole fraction)
Propane 0.754 0.031 .... .... . ...
Butane 0.246 0.969 0.254 .... . ...
Pentane .... .... 0.745 0.613 0.091
Hexane .... .... . ... 0.387 0.909
Accum. Distillate (mole fraction)
Propane 0.981 0.850 .... .... ....
Butane 0.019 0.150 0.988 0.017 0.012
Pentane .... . ... 0.012 0.940 0.778
Hexane .... .... .... 0.043 0.210
Amount (lhmol) 8.139 11.760 36,548=A 8.619 12.180
Still Pot (mole fraction)
Propane 0.021 .... . .. .... . ...
Butane 0.325 OJ19 0.001 .... ....
Pentane 0.109 0.113 0.133 0.023 0.002
Hexane 0.545 0.567 0.866 0.977 0.998
Amount (lbmol) 91.860 88.240 71.680 63.061 59.380=B

the base case (for the required purity, the system behaves as a binary and the optimization
essentially solves a two-point boundary value problem). Taking both the amount produced and the
time required into account (i.e. the productivity objective function), however, permits improving
this step by over 50010. Further improvements are achieved with more control intervals. Two of the
optimum control policies calculated for the maximum productivity problem are also reported in
Table 3. The desired propane purity is achieved in all cases.
 197

Table 3. Batch Distillation of Quaternary Mixture [12J-Optimization of Step 1 with MC2 Colwnn Model
(dynamic, more rigorous) and MT2 Thermo Model (ideal)

Prohlem Optimise Step I Min Max Max

Time Distillate Productivity
Specified:
Top Vapour rate (lomol/hr) 12 12 12
Pressure (oar) 1.03 1.03 1.03
Product statc C3 off I - mole fraction C3 0.981 0.981 0.981
Amount(lomol) 8.139
Time (hr) 4.07
Controls:
Reflux ratio (external) (r): End time (tl) r, tl r r, t1
Optimal Operation
1 control inte"val a)
End Time: tl (hr) 4.01 1.75
Amount of C3 off-I: C (lhmol) 8.15 5.67
Productivity = Ntl Oomol/hr) 2.02 2.00 3.24
5 control intervals 0)
End Time: tl (hr) 2.82 1.64
Amount of C3 ofr-I: C (lomol) 9.26 5.88
Productivity = Citl (lomol/hr) 2.881 2.275 3.59
Optimal reflux ratio policies (controllevcl/cnd time, level/time, ... )
a) (r, tl = (07:1/1.75)
0) (r, t) = <0.30/0.636. 0.664/0.59.0.695/0.91,0.727/1.22.0758/1.64)

Optimization of the first step, as described, provided not only the optimal values of the control
parameters for the step, but also values of all state variables in the column model at the end of the
step. These are used to calculate the initial values of all the state variables in the model for the next
step. If the same equations model is used in the two steps, the simplest initialization policy for step
2 is simply to set the initial values for step 2 to the final values of step 1 (used in all examples.
unless noted). However, it is also possible to change the model (for example, to use a different
thermodynamic option. to eliminate equations for components no longer present) or to carry out
calculations for impulsive events defined at the transition between the two steps (for example, to
determine the new pot amount and composition upon instantaneous addition of a secondary
charge). The next step may then be either simulated (ifall controls are specified) or optimized (if
there are some degrees of freedom). In principle, a different objective function could be used for
each step.
Results for such a sequential optimization of the five distillation steps in the operation, using
the same distillation model throughout, rhihimum time as the objective function for each step and
the same purity specifications as in the base case gave the results summarized in Table 4, in terms
of overall productivity (total amount ofthe two desired products! total time), Significant increases
 198

Table 4. Batch Distillation ofQuatemary Mixture [12] - Optimization of 5 Steps in Sequence (min. time for
each step) with MC2 Column Model (dynamic, more rigorous) and MT2 Thermo Model (ideal)

Problem: Se(IUential Optimisation Stepl Step2 Step3 Step4 StepS Overall

(min time for each step)
Specified:
Top Vapour ralC (Ihmol/hr) 12 42 52 32 52
Pressure (bar) 1.03 1.03 1.03 1.03 1.03
Product state C3 off! C30ff-2 C4 prod C50ff-l C6prod
(Key component) mole fraction (l )0.981 (1)0.850 (2)0.988 (3)0.940 (4)0.998
Amount (Ihmol) 8.139 11.760 36.548 8.619 59.380
Controls:
Reflux ratio (external), r; times r, tl r, t2 r, 13 r,14 r, l1
Optimal Operation
1 control interval per step:
Time (hr) 4.01 1.56 9.20 3.49 1.55 19.31
Amount (Ibmol) A=C4 prod, 8=C6 prod 36.566=A 59.44=8
Productivity = (A+8)rrr (lhmoUhr) 4.84
S control intervals per step:
Time (hr) 2.82 1.37 2.57 2.83 1.54 11.13
Amount (Ibmol) A=C4 prod, 8=C6 prod 36.567=A 59.44=8
Productivity = (A+8)nr (Ihmol/hr) 8.62

in perfonnance (over 5(010) relative to the base case are achieved even with a single (but optimal)
profiles are given in Figure 9 for the whole operation. The dynamic model used does not include
hydraulic equations, therefore, the optimal operation should be checked for weeping, entrainment,
flooding, etc.

Recycle Optimization

Policies for recycles were discussed in general terms in the initial sections. Off-cut recycle
optimization methods have been discussed for the simpler, binary case in [56, 16,60], as well [51,
72,63] as different special cases of multicomponent mixtures.
With reference to the specific operation strategy in Figure 5, if the off-cut produced in Step
2 in one batch is characterized by the same amount and composition as that charged to the pot in
the previous batch, then, for constant fresh charge and with the same control policies applied in
subsequent batches, the operation described in Figure 5 will be a quasi-steady state one, that is
subsequent batches will follow the same trajectories, produce the same product and intermediate
states, etc. This cyclic condition may be stated in several ways, for example by writing the
composition balance for the charge mixing step (Mix) as:
 199

Instan! Distillate Composition Minimum

Time Problem

cut t

,
, tI
/1'1
. I
'
)
'

11 11 11
t imt • hr

Accumulated Disti lIate Composition Minimum

Time Problem

1.11
0

::;; 1.11
~

;1.91
\
'""
::: 1.61
-0
:;';1.71 - Rtflux Ratio

] I. II - - HlX,nt

~1.51
r
~ I.U --·Buhnt
~
~ l.lI I. . . Propant
::: 1.21 cu t 2 ,cu t J ,.- /
0
cu t 1
I
~ 1.11 , I
u
---------
I. II f==r=-=-=F'::"'--r'---,---,='r--r~r=-r--,--I
.-----:::/-----
I 1 11 II 12
t iOlt • hr

Figure 9. Batch Distillation of Quaternary Mixture(12]-Sequential Optimization of all 5 Steps. Optimum

Reflux ratio Profile and Distillate Composition (mole fractions)
 200

BO xBO + R xR = Bc xBc (eq. 5)

where Bo> xBO and Be> xBc the fresh feed and mixed charge amounts and composition, supply the
initial conditions for Step 1 and R, xR are the amount and composition of the collected off-cut
at the final time of Step 2. Different cyclic operations may be obtained corresponding to different
off-cut fraction amount and composition. An optimal recycle policy can be found by manipulating
simultaneously the available controls (e.g. reflux ratio profile and duration) of all distillation steps
within the loop (Step 1 and Step 2), so as to optimize a selected objective function, while satisfYing
constraints 5 (or equivalent ones) in addition to other constraints (on product purity, etc.). A
solution of this problem for binary mixtures was presented [60] using a two level problem
formulation, with minimum time for the cyclic operation as the objective function. For given values
ofR and xlQ a sequence of two optimal control problems is solved in an inner loop. The minimum
time operation of Step 1 is calculated as described above, followed by the minimum time operation
for Step 2, with purity and amount of the off-cut constraints to match the given values. In an outer
loop, the total time is minimized by manipulating Rand xR as the (bounded) decision variables.
An alternative, single level formulation for this problem was also developed by [59] where the
total time is minimized directly using the mixed charge quantities (Bc and xBc) as decision
variables, together with reflux ratio as control variable, as follows:
Min ]=tl+t2=tf
Bc, xBe> r(t}
subject to: DAB model
bounds on all control variables
interior point constraints, e.g.: XDJ(tl} ~ xDJ*
Dl(tl} ~ D 1'"
end point constraints, e.g.: XB3(tt) ~ xB3*
cyclic conditions (eq. 5)
Here tl and t2 are the duration of production Steps 1 and 2, respectively and tf is the total time
of the cyclic operation. The starred values are specifications (only two of which are independent,
the other two being obtained from a material balance over the SIN states Fresh Charge, Maincut
I and Bottom Product in Figure 5). The control vector is discretized, as usual, into a number of
control intervals. However, the end time of Step 1, tl, is defined to correspond to one of the
control interval boundaries. The constraints on the main distillate fraction are now treated as
interior point constraints and the cyclic conditions are treated as end point constraints. This
 201

03 D4

ofT-CUI I oC(-cuI2
Mixwre: butane(l), propane(2), n-hexane (3)
Slate Amount (comfl\lnent)
main .. cut 1
(lanol)
Fresh feed 6.0
Specifications
mole frnctioo
~O.15. 0.35. 0.50>
j
0.' U
.
- ....... -----------

03
D4
0.9
2.0
(I) 0.935
(2) 0.82 -
~
0.6
;
i ,..-----,
- ReO" Racio
. Top Vapour rate=3.0kmul/hr
~ i ... HCJtIftC
Onrjma! Openlljon
R1 0.64 (1) 0.276 ~ o.e!
-
--Pcnttnc
R2 0.41 (2) 0.40 .~ • t-----I
Min. time formaincull + OffcUll: 2.05 hr :~
.~ :\ -Buunc
Min. time for maincut2 + Offcut 2: 1.69 hr 0. 0.2 _, L -_ _ _...J
~: ..."..~.............. .
Total balCh time 3.74 hr u :. ____ ---------- ._.' _ _ _
O'O+-~T-,~r__,~~~,_--~--~~
0.0 o.s 1.0 1.5 2.0 2.5 J.O l.5 ".0
Time, br
Figure 10. Batch Distillation of Ternary Mixture with production and recycle of intermediate off-cut.
Thermodynamic model MT2 (ideal), Column Model MC2 (dynamic, more rigorous)

fonnulation was found to result in rather faster solution that the two level fonnulation, while giving
the same results.
This approach may also be used for multicomponent mixtures and more complex operation
strategies, such as that shown in Figure 6. An example for a ternary mixture (butane, pentane and
hexane) with specifications on two main products was reported in [63]. The production strategy
involves four steps (two main product fractions, each followed by an off-cut production step, with
each off-cut recycled independently), as shown in Figure 10. Using the single level fonnulation,
the minimum distillation time for the first recycle loop was calculated first, providing initial column
conditions for the second loop, for which the minimum time was then also calculated. The results
for this sequential optimization of individual recycle loops are summarized in Figure 10.
Comparison of the total time required with that for the optimal operation involving just two
product steps and no off-cuts (Figure 11) shows a significant reduction of 32% in the batch time.
The same products (amounts and compositions) are obtained with both operations.
 202

04
Bottom
Residue

Cutl
o

Mixture: bUlane(I). prtlpallc(2). II-hexane (3)

j I~r-~--~====~~
] ,'-----1

r
Stale AmoullI (componcnI)
(kmol) mole frnclioll ,,, .................. ..
0.. !
Fresh feed 6.0
Specifications
<0.15. 0.35. 0.50>
j ! CUll
D3 0.9 (I) 0.935 ~ 0.6/
i
- ReO.. , Ra.io
D4 2.0 (2) 0.82
Top Vapour mlc=3.0J..,nolnlr .•. Hu.anc.
Onrima! Onernlion 0 .• 1 ..... Pcacanc

I :+~_\.~.
Min. time for Cull: 3.8 hr a '8 ~

"T-~"'-: .:- : .:- :.; - :. .- ..,.j-.~-.-

...,... -..-.~
Min. time for Cull: 1.71 hr

Tolal balch time 5.51 hr ---r---.... ••....I-

0.0 U 1.0 3.0 4.0 5.0 6.0

Time. hr

Figure 11. Batch Distillation of Ternary Mixture with no Off-Cuts. Thermodynamic Model MT2 (ideal),
Column Model MC2 (dynamic, more rigorous)

Multiperiod Optimization

In the above examples, optimization was perfonned of distillation steps individually and
sequentially. Clearly, this is not the same as optimizing an overall objective function (say, minimum
total time), nor are overall constraints taken into account (say, a bound on overall energy
consumption). An overall optimization was however discussed for all steps in a single recycle loop
and the same two approaches in that section may also be used to optimize several adjacent steps
or even all the production periods in an operation. We refer to this as multiperiod optimization.
This has been discussed by Fahrat et al. [34], however with shortcut models and simple
thennodynamics and in [64] with more general models.
Only a small set of decision variables is required to define a well posed optimal control
problem for individual steps. With the fresh feed charge given, specification of a key component
purity and an extensive quantity (amount of distillate or residue, or a recovery) for each distillation
step permits calculating the minimum time operation for that step, and hence for all steps in
 203

sequence. Typical quantities of interest for overall economic evaluation of the operation (amounts
produced, energy and total time used, recoveries, etc.) are then explicitly available. A general
overall economic objective function may be calculated utilizing unit value/cost ($/ kmol) of all
products, off-cut materials, primary and (if any) secondary feeds and unit costs of utilities (steam,
cooling water). For example an overall profit ($lbatch) may be defined as:

Jl=Sum (product/off-cut values)-Sum (feed costs)-Sum (utility, setup, changeover costs) (eq. 6)
or in terms of profit per unit time ($lbatch):
12 = Jl / batch time (eq.7)

with significant changeover and setup times also included in the batch time. Overall equality and
inequality constraints may be similarly written.
The decision variables may then be selected in an outer optimization step to maximize an
overall objective function subject to the overall constraints. Many purities are typically specified
as part of the problem definition and recoveries represent sensible engineering quantities for which
reasonably good initial guesses can often be provided. The outer level problem is therefore a small
nonlinear programming problem solvable using conventional techniques. The inner problem has
a block diagonal structure and can be solved as a sequence of small scale optimal control problems
one for each STN task.
This solution method was proposed in [64] who presented results for batch distillation of a
ternary mixture in a 10 stage column with two distillate product fractions (with purity specification)
and one intermediate off-cut distillate (with a recovery specified). Here, the more rigorous dynamic
model was used with thermodynamic described by the SRK equation of state (Thermo model
MT3). The amounts of all fractions the composition of the recycle off-cut and the reflux ratio
profiles were optimized so as to maximize the hourly profit, taking into account energy costs. The
optimal operation is summarized in Figure 12 with details given in the original reference. The
required gradients of objective functions and constraints for the outer problem were obtained by
finite difference (however exploiting the block diagonal structure of the problem) with small effort.
Analytic sensitivity information could be used if available.
For the quaternary example [12] previously discussed the operation in Figure 8 is considered
again this time with Step 1and Step 2 merged into the singe task Cut 1 (propane elimination) and
Step 4 and 5 merged into the single task Cut 3 (pentane elimination). A problem is formulated
whereby the overall productivity (amounts of states C4 prod and C6 produced over the total time)
 204

'mm,!4*,·.t;1!dltq.II,IMi$t!1jjfIf!1§·P'd!t.1·\

t::::J
~

o
R. -
-1.03
0.77~
~~
~ Re-O.7&l

0.875 0311 0333 O.!JI 0.876

~ 6.\7 '.51 LJ5

-10
C •• -1.0 I/Iano~ CIl - C., - 0.0. C k - 3.0 I/Ilr
C D1 - C D1 -:roS/kmoI

Tim!.! .hr

Figure 12. Multiperiod Optimization with Ternary Mixture - Maximum hourly profit, SRK EOS, dynamic
colwnn model MC2, Specifications on maincut I and maincut 2 product purity and Cyclohexane
recovery in Off-cut. Optimal operation and instant distillate composition profiles are shown

is maximized subject to the same product purity constraints considered in the sequential
optimization (Table 4). The pentane purity in the C3 off-fraction butane recovery in Cut 2 and
hexane recovery in Cut 3 are now considered as variables to be optimized in addition to the reflux
ratio profiles and times. A two level multiperiod solution leads to the optimal operation detailed
in Table 5. Smaller amounts of product are produced, however in far less time, leading to an
overall productivity almost twice as high as for the optimal sequential solution and four times
higher than the base case. This is no doubt also due to the use here for all times of the largest vapor
rate (52 Ibmol/hr) which was utilized in the previous cases (only in Steps 3 and 5). This
optimization required 4 functions and 3 gradient evaluations for the outer loop optimization and
less than 5 hrs CPU time on a SUN SPARCIO machine.
 205

Table 5. Multiperiod Optimization of Quatermuy Mixture [12] with 3 Separation tasks. Maximum overall
productivity, Me2 column model (dynamic, more rigorous) and MT2 thermo model (ideal)

Multiperiod Optimisation Cut1 Cut2 Cut3 !Overall

max Productivity - (A+R)/Tf
Specified:
Top Vapour rate (Ibmollhr) 52 52 52
Pressure (bar) 1.03 1.03 1.03
State (Key comp.) mole fraction C4 prod (2) 0.988 C6 prod (4) 0.998
State (Key comp.) task recovery C3 off (I) 0.996
Optimised: initial <bounds initial <bounds> initial <bounds>
State C3 off C4 prod C6 prod
(Key comp.) mole fraction (1)0.981<0.8-0.95
(Key comp.) recovery for task 2)0.90<0.85-0.95 4)0.95<0.90-0.98
Controls
Variables (No. contr. intervals) r, t (3) r, t (3) r, t (3)
Initial guess control level 0.8, 0.8, 0.8 0.8, 0.8, 0.8 0.8, 0.8, 0.8
end time 0.5, 1.0, I.5 1.0, 2.0, 3.0 1.0, 2.0. 3.0
Optimal Operation
Time (hr) 1.66 1.28 3.05 5.99=Tf
Productivity =(A+B)fff (Ibmollhr) 14.55
Product State (mole fraction) C3 off C4prod C5 off C6 prod
Propane 0.869 0.001 ... ...
Butane 0.131 0.988 0.258 ...
Pentane ... 0.011 0.450 ..... 0.002
Hexane ... ... 0.292 0.998
Amount (Ibmol) 11.433 31.361=A 121.186 55.814=B
Optimal control level 0.711,0.987,0.969 P.469,O.599,O.723 p.648,O.877,O.945
end time 0.37. 0.90, 1.66 0.57, 0.96. 1.28 0.59, 1.44. 3.05

As with recycles, it is possible to utilize the single level problem formulation for the general
multiperiod case.

Reactive Batch Distillation

Reactive batch distillation is used to eliminate some of the products of a reaction by distillation as
they are produced rather than in a downstream operation. This permits staying away from
(reaction) equilibrium and achieving larger conversions of the reactants than possible in a reactor
followed by separation. With reference to Figure 1, reaction may occur just in the pot/reactor (for
example, when a solid catalyst is used) or in the column and condenser as well (as with liquid phase
reactions). Suitable reactions systems are available in literature [6].
With the batch distillation model (eq. 1), written in general form, extensions to model the
reactive case amount to just small changes in the mass and energy balances to include a generation
term and to the addition of reaction kinetic or equilibrium equations for all stages where reaction
 206

occurs (pot, or all stages). Energy balances do not even need changing if written in terms of total
enthalpy. Modeling details and simulation aspects are given in many references (i.e. [25,4]).
From the operation point of view, there are some interesting new aspects with respect to
ordinary batch distillation. Reaction and separation are tightly coupled (reaction will affect
temperatures, distillate and residue purities, batch time, etc. while the separation will affect reaction
rates and equilibrium, reactant losses, etc.). There is one more objective, essentially the extent of
reaction, but no new variables to manipulate (unless some reactants are fed semi-continuously, in
which case the addition of time and rate may he new controls), making this an interesting dynamic
optimization problem. Optimization aspects were first discussed in [33] and more recently in [86,
40, 65] who also review recent literature.
With the optimal control problem formulated as above, no change is needed to handle the
reactive case, other than to supply a slightly modified model and constraints. Optimal operating
policies were presented in [65] for several reaction systems and column configurations. Typical
results are summarized in Figure 13 for the esterification of ethanol and acetic acid to ethyl acetate
and water, with maximum conversion of the limiting reactant used as the objective function, reflux
ratio as the control variable, given batch time and a constraint on the final purity of the main
reaction product (80% molar ethyl acetate, separated in a single distillate fraction). This
formulation is the equivalent of the maximum distillate problem in ordinary batch distillation.
Similar results were also calculated for maximization of a general economic objective function,
hourly profit = (value of products - cost of raw materials - cost of utilities)I batch time, with the
batch time also used as a control variable. Again, profit improvements in excess of 40% were
achieved by the optimal operation with respect to quite reasonable base cases, obtained manually
by repeated simulations.

On-Line Application - Optimization of All Batches in a Distillation Campaign

The above examples involved the a priori definition of the optimal control profiles for a batch,
assuming that all batches would be carried out in the same way. Here, we wish to show how such
techniques may be applied in an automated, flexible batch production environment.
The application was developed for demonstration purposes and involves a simple binary
mixture of Benzene and Toluene, to be separated into a distillate fraction and a bottom residue,
each with specified minimum purity (benzene mole fraction = 0.9 in the distillate, toluene mole
fraction = 0.9 in the residue). A batch column is available, of given configuration. A quantity of
 207

(II (1) (3)

Acetic Acid + Elhanol - Ethyl AccLlle + Water
~:~:1K) 191.1 lSt.s 350.3 113.1

....~.................-.----..------.--....
2 t:::::l
1----13 to 12 '4 16
EIlIyIA""....
Time, hr

---
,- O.Q12S mol X ~, - 0.10 or 0.10

0.,

"
N-I0
:~. ~------
/.·i··...~.::;. ~.:::::::----__
II.J
IU
'" .,/
-1
~'ctd· <=5.0 kInol
Composition <=<OAS,OAS,O.O,OJ> . "l~~-:':':::::
CoTomD Prcssurccl.OJ3 b:tr °0 1 4.' 11011'4"
Time, hr

Figure 13. Reactive Batch Distillation - Maximum Conversion of Acetic Acid to Ethyl Acetate, with purity
specification on the Ethyl Acetate (Distillate) product. Column model MC2, correlated K-values
and kinetic rates

feed material becomes available, its exact composition being measured from the feed tank. Other
tanks (one for each product plus one for an intermediate) are assumed initially empty. The quantity
offeed is such that a large number of batches are required. The following procedure is adopted:

1. An operation structure is selected for the typical batch, in this case involving an off-cut
production and recycle to the next batch (Figure 5). Given the measured fresh feed charge
composition (mole fraction benzene = 0.6), pot capacity and product purity specifications, the
minimum total time, cyclic operation (reflux ratios profiles, times, etc.) is calculated as 0.6 kmol
of distillate, optimal off-cut of 0.197 kmol at mole fraction benzene = 0.63, reflux ratio r = 0.631
for 2.64 hr (benzene production) then r= 0.385 for 0.51 hr (off-cut production), leaving 0.6 kmol
of bottom product. A single control level was chosen for each step for simplicity, with a
multiperiod, single level formulation for the optimization problem.
2. First batch - Since the off-cut recycle tank is initially empty, the cyclic policy cannot be
implemented at the beginning and some way must be established to carry out one or more initial
 208

batches so as to reach it. One way (not necessarily the best one), is to run the first batch
(according to the operation strategy in Figure 14 (secondary charge and off-cut production.).
The off-cut from the previous batch, OFF-O is known (zero for the first batch). The desired off-cut
OFF-I, is specified as having the optimal cyclic amount and composition determined in I above.
With distillate and residue product purifies specified, a material balance gives the target values for
distillate and product amounts (e.g. 0.4744 Ianol of distillate for the first batch). With these targets,
a minimum total time operation for this batch is calculated. In an ideal world, the cyclic operation
could then be run from batch two onward.
3. The optimal operation parameters for this batch are passed to a control system and the batch is
carried out. The actual amounts and compositions produced (distillate cut, off-cut and residue cut)
will no doubt be slightly different than the optimal targets, due to model mismatch, imperfect
control, disturbances on utilities, etc. Product cuts from the batch are discharged to their
respective tanks. The actual composition in the product and in the off-cut tanks is measured.
4. The optimal operation for the next batch is calculated again using the operation structure in
Figure 14 and the calculation procedure outlined for the first batch but using the measured amounts
and composition of the off-cut produced in the previous batch. Because of this, the optimal policy
for the batch will be slightly different from that calculated in step 1. Any variations due to
disturbances control problems missed targets etc. in previous batches will also be compensated by
the control policy in the current batch. Steps 3 and 4 are repeated until all the fresh feed is
processed (a special operation could be calculated for the last batch so as to leave the off-cut tank

Benzene

Figure 14. Batch Distillation of Binary Mixture (Benzene, Toluene)-Operating strategy with addition of
secondary charge (off-cut from previous batch) and production of intermediate off-cut. Column
model MC2 (dynamic, more rigorous), thermo model MT2 (ideal)
 209

Charge fresh Charge

feed recycle offcut
Distill Benzene product
control parameters: rI. t 1.. ..
Distill offcut
control parameters: r2. t2 ....
Calculate Optimal
Control parameters
for next batch

Dump distillate Dump hOlloms

Figure 15. Control Procedure for the Automatic Execution of one Batch by the Control System

empty).
The procedure controlling the actual batch execution is shown schematically in Figure 15. The
main control phases correspond to the distillation tasks in the STN definition ofthe operation with
additional control phases for all transfers and details with regards to the operation of individual
valves control loops etc. After the main distillation steps (and quality measurements on the
fractions produced) the sequence automatically executes an optimal control phase kicking off the
program for the numerical calculation of the optimal policy for the next batch. The resulting
parameters are stored in the control system database and used in the control phases for the next
batch.
A complete implementation of this strategy within an industrial real time environment (the
RTPMS real time plant management and control system by IBM) was presented in [53] with the
optimal control policies calculated as previously discussed batch management carried out by the
SUPERBATCH system [22, 23, 54] and standard control configurations (ratio control on reflux
rate and level control on the condenser receiver all PID type). Actual plant behavior was simulated
by a detailed dynamic model implemented in a version of the Speedup general purpose simulator
directly interfaced to the control system [68].
 210

This application indicates that the dynamic optimization techniques discussed can indeed be
used in an on-line environment to provide a reactive batch-to-batch adjustment capability. Again,
the operation strategy for each batch was defined a priori and optimization of individual batches
in sequence, as just described, is not the same as a simultaneous optimization of all the batches in
the campaign, so there should be scope for further improvements.

Discussion and Conclusions

We may now draw a number of general conclusions and highlight some outstanding problems.
1) With regards to modeling and simulation, recent advances in handling mixed systems of
algebraic and differential equations with discrete events should now make it possible to develop
and solve without (too many) problems the equations required to model in some detail all the
operations present in batch distillation. The question remains, in my opinion, of how detailed the
models need be in particular to represent the startup period, heat transfer effects in the reboiler,
and hydraulic behavior in sufficient detail.
2) It is presently possible to formulate and solve a rather general class of dynamic optimization
problems with equations modeled as DABs, for the optimization of individual batch distillation
steps. Both the control vector parameterization method utilized in this paper and collocation over
finite elements appear to work well. These solution techniques are reaching a certain maturity, that
is, are usefully robust and fast (a few minutes to a few hours for large problems). Algorithmic
improvements are still needed to handle more efficiently very large scale problems and for some
special cases (e.g. high index problems). Some attention is also required to formulation of
constraints and specifications so that the dynamic optimization problem is well posed. A variety
of problems for which specialized solution methods were previously developed can now be all
effectively solved in a way which is largely independent from the specific column and
thermodynamic models objective function and specifications used. These advances should shift the
focus of batch distillation studies towards the use of more detailed dynamic column models and
towards optimization of more difficult processes (reactive azeotropic extractive with two liquid
phases etc.) which did not easily match the assumptions (simple thermodynamics etc.) of the short
cut methods.
3) With regards to the optimization of several batch distillation steps in a specified sequence
(multiperiod problem), two approaches were presented here, one based on a two level
 211

decomposition, taking advantage of the natural structure of batch distillation, the other on a single
level formulation. In this area, further work is needed to establish whether one approach is better
than the other or indeed for altogether new approaches.
4) The problem of choosing the optimal sequence of steps for the processing of one batch (an
operation strategy), as well as the control variables for each step, has not been given much
attention, no doubt because of the difficulty of the mathematical problem involved. Systematic
methods are needed to select in particular the best strategy for reprocessing off-cuts and more in
general, for processing multiple mixtures. Some initial results were presented in literature [82, 83]
where a nonlinear programming (NLP) formulation was proposed. The use MINLP for simulation
and solution was suggested, but not developed. In this area, there is clearly scope for novel
problem formulations and solutions. Similarly, the optimization of each batch in a campaign so as
to maximize the performance of the whole campaign does not appear to have been considered
other than in the context of scheduling, with extremely simplified "split fraction" and fixed time
models, e.g. in [45].
5) One of the assumptions made initially was that of perfect control response. The integration of
open loop operation design and closed loop optimization is clearly relevant, as are the sensitivity,
controllability and robustness properties of any optimal operation policies. These issues are beyond
the scope of this paper, but some initial results are discussed in [73, 81], while an interesting
method for model based control of a column startup was presented in [7].
6) Current developments in hardware speed, optimal control algorithms and control and
supervisory batch management systems are such that sophisticated optimal operations can be
calculated and implemented on-line, not only with respect to the optimal control policies for a
batch, but also with regards to batch-to-batch variations, as demonstrated by the last application
example. "Keeping it constant", which used to be a practical advantage, is no longer a constraint.
7) Finally, while a number of earlier studies indicated that performance improvements obtained by
optimizing the reflux ratio policies were often small, if not marginal, more recent work appears to
point to double digit benefits in many cases. As discussed above in one of the examples, this is
possibly due to the consideration of a wider range of operating choices, more realistic models and
objective functions. Whether the benefits predicted using the more advanced operation policies are
indeed achieved in practice is an interesting question which awaits confirmation by the presentation
of more experimental, as well as simulated results.
 212

supervisory atc
management
Reid Tirne Plant (SUPERBATCH)
\Management System
RTPMS

3
plant control 5
real time optimisation
software
database software
(ACS) 6

Data flows
1 control commands and parameters
2 measurements
3 phase commands and parameters
4 phase status
5 optimal control requests and problem data
6 optimal control solutions (phase parameters) for next batch

Figure 16. Schematic Structure of Control Software

Acknowledgments

This work was supported by SERC!AFRC, whose contributions are gratefully acknowledged.

References

I. Abdul Aziz, B. B., S. Hasebe and I. Hashimoto, Comparison of several startup models for binary and
ternary batch distillation with holdups. in Interactions Between Process Design and Process Control,
IFACWorkshop,I.D. Perkins ed., PergamonPress,pp 197-202,1992
2. Abram, H.I., M. M. Miladi and T. F. Attarwala. Preferable alternatives to conventional batch
distillation. IChemE. Symp Series No. 104. IChemE, Rugby, UK, 1987
3. Albet, J.,. Simulation Rigoureuse de Colonnes de Distillation Discontinue a Sequences Operatoires
Multiples. PhD Thesis, ENSIGC, Toulouse, 1992
4. Albet, J., I. M. Le Lann, X. Joulia and B. Koehret. Rigorous simulation of multi component multi
sequence batch reactive distillation. Proceedings Computer Oriented Process Engineering, Elsevier
Science Publishers B.V., Amsterdam, p. 75. 1991
5. Barb, D. K and C. D. Holland, Batch distillation. Proc. of the 7th World Petroleum Con., 4, 31,.1967
6. Barbosa, D. and M. F. Doherty, The influence of chemical reactions on vapor-liquid phase diagrams.
Chem Eng. Sci.,43 (3), 529, 1988
7. Barolo, M., G.B. Guarisc, S. Rienzi and A. Trotta, On-line startup of a distillation column using generic
model control. Comput. Chem. Engng., 17S, pp 349-354,(1992
8. Barton, P. and C. C. Pantclides, The modeling and simulation of combined discrete/continuous
processes. Proc., PSE'91, Vol. I, pp.20., Montebello, Canada, 1991
9. Bernot, C., M.F. Doherty, and M. F. Malone. Patterns of composition change in multicomponent batch
 213

distillation. Chern. Eng. Sci., 45 (5), 1207, 1990

10. Biegler, L.T., Solution of dynamic optimization problems by successive quadratic programming and
orthogonal collocation. Comput. Chern. Engng., 8, pp 243-248, 1984
11. Bortolini and Guarise (1970). Un nuovo metodo di distillazione discontinua. Ing. Chim. Ital., Vol. 6,
pp. 1-9,1970
12. Boston J.P., H.J. Britt, S. Jirapongphan and v.B. Shah, An advanced system for the simulation of batch
distillation operation. FOCAPD, 2, p. 203, 1981
13. Caracotsios, M. and W.E. Stewart, Sensitivity analysis of initial value problems with mixed ODEs and
algebraic equations. Comput. Chern. Engng., 9, pp. 359-365, 1985
14. Chang, YA. and J.D. Seider, Simulation of continuous reactive distillation by a homotopy-continuation
method. Comput. Chern. Engng., 12. 12, p. 1243, 1988
15. Chen, c.L., A class of successive quadratic programming methods for flowsheet optimization. PhD
Thesis. Imperial College. University of London, 1988
16. Christensen, F.M. and S.B. Jorgensen, Optimal control of binary batch distillation with recycled waste
cut. Chern. Eng. J., 34, 57, 1987
17. Clark, S. M. and G. S. Joglekar, General and special purpose simulation software for batch process
engineering. This volume, p 376
18. Converse, A.O. and G.D. Gross, Optimal distillate-rate policy in batch distillation. lEC Fund, 2(3),
p. 217., 1963
19. Converse, A.O. and C.I. Huber, Effect of holdup on batch distillation optimization. lEC Fund., 4 (4),
475, 1965.
20. Corrigan, T.E. and W.R. Ferris, A development study of methanol-acetic acid esterification. Can. 1.
Chern. Eng., 47(6), 334,1969
21. Corrigan, T.E. and J.H. Miller, Effect of distillation on a chemical reaction. IEC PDD, 7(3), 383, 1968
22. Cott, B. J. , An Integrated Management System for the Operation of Multipurpose Batch Plants. PhD
Thesis, Imperial College, University of London, 1989
23. Cott, B.1. and S. Macchietto, An integrated approach to computer-aided operation of batch chemical
plants. Comput. Chern. Engng., 13, 111l2, pp. 1263-1271,1989
24. Coward, I.,. The time optimal problem in binary batch distillation. Chern. Eng. Sci., 22, 503, 1967
25. Cuille, P.E. and G.Y. Reklaitis, Dynamic simulation of multicomponent batch rectification with
chemical reaction. Comput. Chern. Engng., 10,4,389,1986
26. Cuthrell, 1.E. and L. T. Biegler, On the optimization of differential-algebraic process systems, AIChE
1.,33, 1257, 1987
27. Distefano, G.P., Mathematical modeling and numerical integration of multicomponent batch distillation
equation. AIChE 1., 14, I, p. 176, 1968
28. Diwekar, UM., Unified approach to solving optimal design-control problems in batch distillation.
AIChE 1.,38(10),1571,1992
29. Diwekar, U.M. and 1.R. Kalagnanam, An application of qualitative analysis of ordinary differential
equations to azeotropic batch distillation, AIChE Spring National Meeting, New Orleans, March 29
- April2, 1992
30. Diwekar, UM. and K.P. Madhavan, Optimal design of multicomponent batch distillation column.
Proceedings of World Congress III of Chemical Engng., Sept., Tokyo, 4, 719,1992
31. Diwekar, UM. and KP. Madhavan, BATCHDIST- A comprehensive package for simulation, design,
optimization and optimal control of multi component, multi fraction batch distillation columns design.
Comput. Chern. Engng., 15 (12), 833,1991
32. Diwekar, UM., KP. Madhavan and R.K Malik, Optimal reflux rate policy determination for
multicomponent batch distillation columns. Comput. Chern. Engng., 11,629, 1987
33. EgIy, H., V. Ruby and B. Seid, Optimum design and operation of batch rectification accompanied by
chemical reaction. Comput. Chern. Engng., 3, 169, 1987
34. Farhat, S., M. Czernicki, M., L. Pibouleau, L. and S. Domenech, Optimization of multiple-fraction
batch distillation by nonlinear programming. AIChE J., 36(9), 1349, 1990
35. Galindez, II. and A. Fredenslund, Simulation of multi component batch distillation processes. Comput.
 214

Chem Engng., 12(4),281,1988

36. Gani, R., C.A. Ruiz and I. T. Cameron, A generalized model for distillation columns - I. Model
description and applications. Comput. Chern. Engng., 10(3), 181, 1986
37. Gear, C. W., Simultaneous numerical solution of differential-algebraic equations. IEEE Trans. Circuit
Theory, CT- 18,89,1971
38. Gonzales-Velasco, 1. R., M.A. (Gutierrez-Ortiz, 1. M. Castresana-Pelayo and J A. Gonzales-Marcos.
Improvements in batch distillation startup. IEC Res., 26, pp. 745, 1987
39. Gritsis, D., The Dynamic Simulation and Optimal Control of Systems described by Index Two
Differential-Algebraic Equations. PhD. Thesis, Imperial College, University of London, 1990
40. Gu, D. and AR. Ciric, Optimization and dynamic operation of an ethylene glycol reactive distillation
column. Presented at the AlChE Annual Meeting, Nov. 1-6, Miami Beach, USA, 1992
41. Hasebe, S., B.B. Abdul Aziz, 1. Hashimoto and T. Watanabe, Optimal design and operation of complex
batch distillation column. in Interactions Between Process Design and Process Control, IFAC
Workshop, J.D. Perkins ed., Pergamon Press, p 177,1992
42. Hansen, T. T. and S.B. Jorgensen, Optimal control of binary batch distillation in tray or packed
columns. Chem. Eng. 1.,33,151,1986
43. Hindmarsh, AC., LSODE and LSODI, two new initial value ordinary differential equation solvers. Ass.
Comput. Mach., Signam Newsl., 15(4), 10, 1980
44. Holland C.D. and A. I. Liapis, Computer methods for solving dynamic separation problems.
McGraw-Hill Book Company, New York, 1983
45. Kondili, E., C.C. Pantelides, R.W.H. Sargent, A general algorithm for scheduling of batch operations.
Proceedings 3rd intI. Syrnp. on Process Systems Engineering, pp. 62-75. Sydney, Australia, 1988
46. Kerkhof, L.H. and H.J.M. Vissers, On the profit of optimum control in batch distillation. Chern. Eng.
Sci., 2, 961,1978
47. Logsdon, IS. and L.T. Biegler, Accurate solution of differential-algebraic optimization problems. Ind.
Eng. Chem. Res., 28, 1, pp. 1628-1630,1989
48. Logsdon, IS. and L.T. Biegler, Accurate determination of optimal reflux policies for the maximum
distillate problem in batch distillation. AIChE National Mt., New Orleans, March 29 - April 2, 1992
49. Lucet, M., A Charamel, A Chapuis, G. Guido and J. Loreau, Role of batch processing in the chemical
process industry. This volume, p. 43
50. Luyben, W.L., Some practical aspects of optimal batch distillation design. IEC PDD, 10,54, 1971
51. Luyben, W.L., Multicomponent batch distillation. 1. Ternary systems with slop recycle. IEC Res., 27.
642, 1988
52. Macchietto, S., Interactions between design and operation of batch plants. in interactions between
process design and process control, IFAC Workshop, J.D. Perkins ed.,Pergamon Press, pp. 113-126,
1992
53. Macchietto, S., B. J. Colt, 1. M. Mujtaba an J C. A Crooks, Optimal control and on line operation of
batch distillation. AIChE Annual Meeting, San Francisco, Nov. 5-10, 1989
54. Macchietto, S., C. A Crooks and K. Kuriyan, An integrated system for batch processing. This volume,
p. 750
55. Mayur, D.N. and R. Jackson, Time optimal problems in batch distillation for multicomponent mixtures
and for columns with holdup. Chem. Eng. 1.,2, 150, 1971
56. Mayur, D.N., R.A. May and R. Jackson, The time-optimal problem in binary batch distillation with a
recycled waste-cut. Chern. Eng. J. ,1,15,1970
57. McGreavy, C. and G.H. Tan, Effects of process and mechanical design on the dynamics of distillation
column. Proc. IFAC Symposium Series, Boumemouth, England, Dec. 8-10, p. 181,1986
58. Morison, K.R., Optimal control of processes described by systems of differential and algebraic
equations. PhD. Thesis, University of London, 1984
59. Mujtaba, I.M., Optimal operational policies in batch distillation. PhD Thesis, Imperial College,
London, 1989
60. Mujtaba, I.M. and S. Macchietto. Optimal recycle policies in batch distillation - binary mixtures.
Recents progres en Genie des Proceeses, (S. Domenech, X. Joulia and B. Koehret Eds.) Vol. 2, No.6,
 215

pp. 191-197, Lavoisier Technique et Docwnentation, Paris, 1988

61. Mujtaba, I.M. and S. Macchietto, Optimal control of batch distillation. in IMACS Annals on
Computing and Applied Mathematics-Vol 4: Computing and Computers for Control Systems, (P.
Borne, S.G. Tzafestas, P.Breedveld and G. Daophin Tanguy, eds.), I.C. Baltzer AG, Scientific
Publishing Co., Basel, Switzerland, pp.SS-58, 1989
62. Mujtaba, I.M. and Macchietto, S, The role of holdup on the performance of binary batch distillation.
Proc., 4th International Symp. on PSE, Vol. 1, 1.19.1, Montebello, Quebec, Canada, 1991
63. Mujtaba, I.M. and S. Macchietto, An optimal recycle policy for multicomponent batch distillation.
Computers Chern. Engng., 16S, pp. 273-280, 1992
64. Mujtaba, I.M. and S. Macchietto, Optimal operation of multicomponent batch distillation. AIChE
National Mt., New Orleans, USA, March 29-ApriI2, 1992
6S. Mujtaba, I.M. and S. Macchietto, Optimal operation of reactive batch distillation", AIChE Annual
Meeting, Nov. 1-6, Miami Beach, USA, 1992
66. Murty, B.S.N., K. Gangiah and A Husain, Performance of various methods in computing optimal
control policies. Chern Eng, 1.,19, p. 201, 1980
67. Nad, M. and L. Spiegel, Simulation of batch distillation by computer and comparison with experiment.
Proceedings CEF'87, 737, Taormina, Italy, 1987
68. Pantelides, C. C. Speedup - Recent Advances in Process Simulation. Computers Chern. Engng., 12 (7),
745, 1988
69. Pantelides, C.C, D. Gritsis, KR. Morison and R. WH. Sargent, The mathematical modeling of transient
system using differential-algebraic equations. Comput. Chern. Engng., 12(5),449, 1988
70. Pantel ides, C. C. Sargent, R. W. H. and V. S. Vassiliadis, Optimal control of multistage systems
described by differential-algebraic equations. AIChE Annual Meeting, Miami Beach, Fla, USA, 1988
71. ProsimBatch, Manuel Utilisateur. Prosim S. A, Toulouse, 1992
72. Quintero-Marmol, E. and WL. Luyben, Multicomponent batch distillation. 2. Comparison of
alternative slop handling and operating strategies. IEC Res., 29,1915,1990
73. Quintero-Marmol, E. and W.L. Luyben, Inferential model-based control of multicomponent batch
distillation. Comput. Chern. Engng., 47(4),887,1992
74. Renfro, J.G., A. M. Morshedi and O.A. Asbjornsen, Simultaneous optimization and solution of
systems described by differentiaValgebraic equations. Comput. Chern. Engng., 11(S), S03, 1987
7S. Robinson, E. R., The optimization of batch distillation operation. Chern. Eng. Sci., 24, 1661, 1969
76. Robinson, E.R. , The optimal control of an industrial batch distillation column, Chern Eng. Sci., 25,
921, 1970
77. Robinson, C.S. and E.R. Gilliland, Elements of fractional distillation. 4th ed., McGraw-Hill, 19S0
78. Rippin, D. W T., Simulation of single and multiproduct batch chemical plants for optimal design and
operation. Comput. Chern. Engng., 7, pp. 137-1S6, 1983
79. Rose, L.M., Distillation design in practice, Elsevier, New York ,198S
80. Ruiz, C.A, A generalized dynamic model applied to multicomponent batch distillation. Proceedings
CHEMDATA 88,13-15 June, Sweden, p. 330.1, 1988
81. Sorensen, E. and S. Skogestad control strategies for a combined batch reactorlbatch distillation process.
This volwne, p. 274
82. Sundaram, S. and L.B. Evans, Batch distillation synthesis. AlChE Annual Meeting, Chicago, 1990
83. Sundaram, S. and L.B. Evans, Synthesis of separations by batch distillation. Submitted for publication,
1992
84. VanDongen, D.B. and M.F. Doherty, On the dynamics of distillation processes: Batch distillation.
Chern. Eng. Sci., 40, 2087, 1985
8S. Villadsen, 1. and M. I. Michelsen, Solution of differential equation models by polynomial
approximation. Prentice Hall, Englewood Cliffs, NJ, 1978
86. Wilson, JA Dynamic model based optimization in the design of batch processes involving
simultaneous reaction and distillation. IChemE Symp. Series No. 100, p. 163, 1987
Sorption Processes

Alirio E. Rodrigues I and Zuping Lu

Laboratory of Separation and Reaction Engineering, School of Engineering, University of Porto

4099 Porto Codex, Portugal

Abstract: The scientific basis for the design and operation of sorption processes is reviewed.

Examples of adsorptive processes, such as liquid phase adsorption, parametric pumping and

chromatography, are discussed. A new concept arising from the use of large-pore materials is
presented and applied to the modeling of HPLC and pressure swing adsorption processes.
Numerical tools to solve the model equations are addressed.

Keywords: Sorption processes, chromatography, intraparticle convection, modeling, parametric

pumping, pressure swing adsorption, high pressure liquid chromatography.

Chapter Outline

In the first part, we will present a definition and the objectives sorption processes. Then, the
fundamentals of design and operation of such processes will be reviewed, examples of liquid phase
adsorption and parametric pumping presented, and chromatographic processes will be discussed
showing various modes of operation. The concept of Simulated Moving Bed (SMB) will be
introduced. The second part will be devoted to a new concept in High Pressure Liquid
Chromatography (HPLC) and Pressure Swing Adsorption (PSA) processes will be shown. Finally,
numerical tools used for solving model equations will be briefly reported and ideas for future work
will be given.

I Lecturer
 217

1. Sorption Processes

1.1. DEFINITION AND OBJECTIVES

Sorption processes are a sub-set of percolation processes ,i.e., processes in which a fluid
flows through a bed of particles, fibers or membranes exchanging heat/mass or reacting with
the support [1]. Examples of percolation processes are ion exchange, adsorption,
chromatography, parametric pumping, pressure-swing adsorption (PSA). Sorption processes
can be carried out with different objectives: i) separation of components from a mixture, ii)
purification of diluents, iii) recovery of solutes.

1.2. BASIS FOR TIIE DESIGN OF SORPTION PROCESSES

The basis for the design of sorption processes, as for any other chemical engineering
operation, are:

i) conservation equations (mass, energy, momentum ,electric charge)

ii) kinetic laws (heat transfer, mass transfer, reaction)
iii) equilibrium laws at the interfaces
iv) boundary and initial conditions
v) optimization criteria

The factors governing the behavior of sorption processes are: equilibrium isotherms,
hydrodynamics, kinetics of film mass!heat transfer, kinetics of intraparticle mass/heat transfer
(diffusion and convection) [2]. Sorption in fixed-bed is really a wave propagation process.
Figure la shows the evolution of the concentration front at different times (concentration
profiles in the bed), starting with a clean bed until complete saturation; Figure 1b shows the
concentration history at the bed outlet, i.e., solute concentration as a function of time. Several
quantities can be defmed at this point. The breakthrough time tbp is the time at which the outlet
solute concentration is, say, 0.01 Cin; the stoichiometric time tst is the time corresponding to
complete saturation if the concentration history was a discontinuity. The useful capacity
(storage) of the bed Su is the amount of solute retained in the column until tbp; the total capacity
Soo is the amount of solute retained in the bed until complete saturation. The length of the mass
transfer zone (MTZ) is Ushock tMIT and the Length of Unused Bed (LUB) is L( 1- tbp/tsU. The
stoichiometric time is the first quantity to be calculated in designing a fixed-bed adsorption
process; it is obtained from an overall mass balance leading to:

£V 1-£ 9.in
tst=U( 1+( Cin) [1]
 218

U Ctn

II
C. t

C'n

C.

t BP tr time
Figure 1. Concentration profiles and concentration history

where V is the bed volume, U is the flowrate, qin is the amount sorbed in equilibrium with the
inlet concentration of the sorbed species Cin and £ is the interparticle porosity. Introducing the
£V 1-£ .9.i.n
space time t =U and the capacity factor I; =T Cin ' we get tst =t (1 +1;).

1.3. EQUILIBRIUM MODEL. COMPRESSNE AND DISPERSIVE FRONTS.

The main features of fixed-bed sorption behavior can be captured by using an equilibrium
model based on the assumptions of instantaneous equilibrium at the fluid/solid interface at any
point in the column, isothermal operation, plug flow for the fluid phase and negligible pressure
drop.

For a single solute, tracer system, the model equations for the equilibrium model are:
mass balance for species i :

[2]

Sorption equilibrium isotherm

 219

[3]

The velocity of propagation of a concentration Ci is:

[4]

It appears that the nature of the sorption equilibrium isothenn is the main factor governing
the behavior of the fixed-bed process. For unfavorable isotherms, f'(Ci» 0; therefore, uCi
decreases when Ci increases and the front is dispersive [3]. For favorable isotherms, f'(cj) < 0
and so Uci increases with Cj; the front is compressive leading to a shock which propagates with
the velocity Us given by:

v [5]
us- 1-£ &}
1+£&;

where ~q and ~c are calculated between the feed state (Cjn,qin) and the presaturation state of the
bed; for a clean bed the presaturation state is (0, 0).

1.4. KINETIC MODELS

Sorption kinetics has been included in the analysis of sorption processes in two different
ways. The first one is through a kinetic law similar to a reaction rate as in the Thomas model [4]:

~ = k, [ Cj (qO - qi) - r qj (CO - Cj)] [6]

where r is the reciprocal of the constant separation factor. The Thomas model can be simplified
to the Bohart model [5] in the case of irreversible isotherms (r=O), to the Walter model [6] in the
case of linear isotherms, etc. as shown elsewhere [7]. Thomas model has been applied recently
in the area of affinity chromatography [8, 9]. This class of inodels is what we call "chemical
kinetic type models". The second way of treating sorption kinetics is by describing intraparticle
mass transport. This class of models is called "physical kinetic type models". A typical model in
this group is Rosen's model [10]. It considers homogeneous diffusion inside particles, film
diffusion and linear isothenns. However, the general particle equation is:

[7J

where q> is the flux of species i through the sphere at radial position r and eve is the solute
concentration in the volume element. According to the particle structure and the model used to
describe diffusion inside particles we may have:
 220

i) homogeneous diffusion:

ii) pore diffusion:

iii) pore + surface diffusion in parallel

dC' ~
cp =-epDp ~ -Db (Jrl; Cve=EpCpi +qi
iv) pore + surface diffusion in series
dC' -
cp =-epDp ~; Cve=£pCpi +qi
where qi is the average adsorbent concentration in the microspheres calculated with the
homogeneous diffusion model.

1.5. METHODOLOGY FOR TIlE DESIGN OF SORPTION OPERATIONS

The methodology for the design of sorption processes is based on the measurement of
sorption equilibrium isotherms (batch equilibration), film mass transfer ( shallow-bed
technique), intraparticle diffusivity (Carberry type adsorber operating in batch or as a CSTR)
and axial dispersion (fixed-bed) by simple, independent experiments. Model parameters are
then introduced in the adsorber model in order to predict its dynamic behavior under conditions
different from those used at the laboratory scale [11, 12].

1.5.1. Adsorption from liquid phase

This methodology was tested for a single component system: adsorption from liquid phase
of phenol onto a polymeric adsorbent (Duolite ES861; Rohm and Haas). Adsorption
equilibrium isotherms measured at 20 0C and 60 °C are shown in Figure 2.

........ 40 FENOL/ES-861
•..
Cl
........ 20 0 e
Cl 40 0 e
E
'-'"

* C"
20

o 9) 100 19) 200

c*(mg/l)
Figure 2. Adsorption isothenns at 20°C and 60 °C for the system: phenoVduolite ES-861
 221

Intraparticle effective diffusivities were measured in a stirred adsorber of Carberry type

shown in Figure 3. The response of the batch adsorber was measured for two different particle
diameters dp=o.077 cm dp=o.06 em and dp=o.034 em (Figure 4).

Figure 3. Sketch of the Carberry type adsorber

O.S

0.1

10 20 30 40 SO
t (min.)

Figure 4. Batch adsorption phenol on to Duolite ES·861 for different particle diameters
 222

Film mass transfer coefficients were measured by using the shallow bed technique shown
in Figure 5. The response of the system to a step in phenol concentration at the inlet is shown in
Figure 6. Modeling studies showed that the initial part of the response curve does not depend
on the value of the intraparticle diffusivity and can be used to estimate the number of fIlm mass
transfer units Nc.

11M:~1=-1
11Q;C)(M1t-- 2
11":'!'~'!t--
:...::..,::....: 43

DQ~t-- 5

Figure 5. Shallow-bed

(1 - Column; 2&5 - glass beads; 3 - inox gauze; 4 - shallow bed; 6 - porous glass support)

X(l) Nf
"-0.0
o.s L,;:;:====:=:::::::
.. 1.0

• 2.0

o.s 1.0
t (min.)
Figure 6. Response of a shallow-bed: outlet concentration versus time
 223

Finally, breakthrough curves were obtained in flxed bed runs. Experimental and model
predicted results are shown in Figure 7. We should stress that model paranieters were measured
by independent experiments; no fltting of breakthrough curves was involved. Table I
summarizes operating conditions and model parameters.

Table I. Operating conditions and model parameters for fixed-bed runs

Run U(ml/min) cin(mg!l) ~ Nr Nd Pe

158.7 82 93.6 36.3 0.261 18.7

2 115.2 91.6 79.7 35.8 0.395 24.0
3 54.4 91.6 82.7 57.0 0.824 36.0
4 16.8 82.4 95.1 132.6 2.468 68.0

X( 1,e)

RUN
0.5 1 •
I •
••
4 •

o
o 0.5 1.0 1.5 2.0 e
Figure 7. Breakthrough curves of phenol in a fixed-bed of Duolite ES-861

Scaling for large beds is not a major problem. In fact only the parameter related with
hydrodynamics will be affected.

Sorption processes are cyclic in nature in the sense that saturation, adsorption or load is
followed by regeneration or desorption. In the case of phenol/polymeric adsorbent system the
regeneration is made with sodium hydroxide and therefore a problem of desorption with
chemical reaction arises. The basis for the regeneration is that the amount of phenol adsorbed
as a function of the total concentration (phenol + phenate) changes with pH as shown in Figure
8.
 224

When both steps (saturation and regeneration) are understood it is quite easy to predict the
cyclic behavior. Model and experimental results are shown in Figure 9; the model used for the
regeneration step is a shrinking core model in which the inner core contains phenol and the
outer core contains phenate from the reaction between NaOH and phenol.[13.14] .

......
......!!D"I
0
E 0.4
...S
0- a.,
0.2
10.S
~---
11.S

00 1 2 ,
cr (mmole/I)
Figure 8. Effect of the pH on the adsorption equilibrium of phenol on Duolite ES-861

........
......
~

51)00

-
Cl
E
o phenate
U • phenol
1000

o 20 40 100 120 140 110

t (min.)
Figure 9. Cyclic behavior

1.5.2. Parametric pumping

Parametric pumping was invented by Wilhelm et al. in 1966 [15]; it is a cyclic separation
process based upon the effect of an intensive thermodynamic variable (temperature. pressure or
pH) on the adsorption equilibrium isotherm together with a periodic change of the flow
 225

direction. There are two modes of operation of thennal parametric pumping: a) direct mode, in
which the cyclic temperature is imposed through the column wall and b) the recuperative mode
in which the temperature change is imposed at the bed extremities being canied by the fluid.

A linear equilibrium model enables us to understand the behavior of a thermal parapump.

Model equations are:

mass balance for species i :

[8]

sorption equilibrium isothenn

qj = K(T) Cj [9]

The key parameter in the analysis is the separation factor

b __ a_
[lO]
-1+mo

h m(Tl)-m(T2) m(Tl)+m(T2) d (T) l-e K(T)

werea= 2 ,rna= 2 an m =T .

The concentration velocity in each half-cycle is

(j=1 cold downward flow; j=2 hot upward flow) [11]

'Y T*
It has been shown [16] that b = tanh [( 2 1 -T*/4] where T*= 6T/T and y=(- 6H )/RT.

For the system phenoIlDuolite ES861 b=O.32 to 0.36.

Laboratory work in narrow columns (1.5 cm in diameter) was carried out many years ago
[17,18] with good results. For a semicontinuous parametric pump the reduced concentration in
the bottom reservoir after n cycles is given by:

[12]

and the top product concentration at steady-state is :

<ytp>oo = 1+'l>b [13]

Yo 'l>l

where % and 'l>t are the reflux ratios to the bottom and top of the column, respectively.
 226

Recent work [19] was carried out in wide columns ( 9 cm in diameter). The completely
automated unit is shown in Figure 10. The system was characterized from the hydrodynamic,

......... ---....,
T 0-- - - - - - , IBM
computer
p~~~~~
'Y --".. DT707-T DT280S

RS232

Figure 10. Experimental set-up

1.- glass column G90-Amicon. 2- Feed reservoir. 3.- Top reservoir. 4.- Bottom reservoir. 5.-
Fraction collector. 6-7.- Heat exchange. 8-12.- two-ways solenoid valves. 13-14. - three-ways
solenoid valves. 15-19.- Peristhaltic pumps. Tl-T5. - Thermocouples type K. PI-P3.- Pressure
transducers

IO.r---------------------------------~

I. . ~
a a • • • • • • •

r
oJ

,g::: .1
II

0.01
2 3 4 5 6 7 8
cycle

Figure 11. Model and experimental results for semicontinuous paranletric pwnping
 227

heat/mass transfer point of view. A typical result is shown in Figure 11. Conditions were: feed
concentration = 98mg/l; bed initially equilibrated with feed composition at 60 oC; average
flowrate = 290 rnl/min; top product flowrate = 12 rnl/min; bottom product flowrate = 30 rnl/min;
average cycle tirne=156 min; 4>1>=0.1; «I>t=O.04; Vu=20300 mI; Vo=24900 mI; U1t/w=22400 mi.

A complete model for parametric pumping was developed including all relevant
mechanisms; results are in Figure 1nor comparison with experimental results.

1.6. CHROMATOGRAPHY
1.6.1. Operating modes

Chromatography is an old operation discovered by M.Tswett [20]. The definition of

chromatography given by IUPAC is a wordy and lengthy one; however, the main characteristic
of chromatography, i.e, separation occurs as a result of different species velocities, is missing.
Several modes of operation are listed below [21]:

a) Elution chromatography
The sample to be separated is injected in a continuous stream of eluent; the main problem is
the dilution of components to be separated.

b) Gradient elution
The elution of the samples is carried out with different eluents.

c) Frontal chromatography
The column is continuously fed with the mixture to be separated until complete saturation
of the sorbent which is then regenerated.

d) Displacement chromatography
After loading with the sample, a displacer with higher affinity is fed to the column.

Figure 12 shows the above mentioned operating modes of chromatography.

Elution chromatography has several drawbacks: a)dilution of the species as they travel
down the bed and b) only a fraction of the bed is effectively used.

Several operating modes can be envisaged to improve process performance such as recycle
chromatography, mixed recycle chromatography, segmented chromatography and two-way
chromatography [22, 23]. Figure 13 shows schematically these operating modes. Figure 14
compares results obtained in the separation of a binary mixture by elution, mixed recycle and
two-way chromatography. Process performance can be improved using these new operating
modes, namely in terms of eluent consumption.
 228

input output

1tJ\
elution

E
Eq

time time
frontlll

B
Eq

time time
E2 gradIent

time time

Eq Dis.

time

Figure 12. Operating modes of chromatography

1.6.2. Simulated moving bed (SMB)

Along the line of improving the use of the adsorbent the idea of 5MB was developed; it is
in my opinion one of the more interesting ideas in chemical engineering [24].

In a moving bed sketched in Figure 15 both solid and fluid phases flow in counter-current
However, there is a problem of attrition. The idea was then to simulate the behavior of a moving
bed by keeping the particles fixed in a column and moving the position offeed and withdrawal
streams. Schematically, the 5MB can be represented as in Figure 16. Many applications of this
technology are currently being used in industry such as the Parex process for the recovery of
p-xylene from a mixture of Cs isomers, Sarex process, etc. [25,26]. Modeling and experimental
studies in 5MB have been worked out by Morbidelli et al. [27]. Figure 17 shows a typical
diagram for the separation of p-xylene from a mixture containing m-xylene, o-xylene and
ethyl benzene.
 229

1&.
Simple Recycle Chromatography
T I
1&. Mixed Recycle Chromatography

Feed'~~_ _----,I
Feed 1 1lXl.
-\:;&'-----~=_=_~'51::;-~.
Segmented Chromatography

Feed lr-~~ri&lF
~ I IIIIII~'
Multi Segmented Chromatography

Inversion

Figure 13. Enhanced operating modes of chromatography

B Elution

time
A
pure without dilution

pure

time
.J, recycle

Figure 14. Comparison between elution and recycle chromatography for separation of binary mixture
 230

I I
I I
I I
I
I
I Zone IV
I Removal A
I
I
----+----------------- .....
C
I QJ
I ::::J

t
I
Zone III
QJ
A _+_B--M------------------ QJ
"0
I
I
Zone"
~
QJ
I
I Removal B ~
I
I
_---4-----------------
I
I
I Zone I
I
I ~I~---~I Eluent
IL __________ lI I ___ L _ ~ _+__

Figure 15_ Moving-bed

I
I
Q) I Q)
c: I c:
o I 0
N I N

eJoC]------eJoC]-<\
~
Eluent Zone IV ~
Figure 16_ Simulated moving-bed
 231

PX purity 99.6596
PX recovery 91.896

100.

~ 10.
~
Z
0
E
en
0
c.
2:
0
to)
1.

A tt-X
B O-X
C P-X
.1 o TOL
1 S S 7 9 1113 1S 1719 21 2S 1 E ED
G PDED
FEED BED LOCATION H SATs
Figure 17. A typical profile in a Parex process

2. A NEW CONCEPT: DIFFUSIVITY ENHANCEMENT BY INTRAPARTIa.E

CONVECflON. APPLICATION TO HPLC AND PSA PROCESSES.

The need for large pore ttanspon pores in catalyst preparation was recognized by Nielsen et
al. [28] and others. The imponance of flow thrOugh the pores was addressed by Wheeler in his
remarkable paper [29]. He concluded that viscous flow due to pressure drop would only be
imponant for high pressure gas phase reactions. He also wrote the correct steady state model
equation for the panicle taking into account diffusion, convection and reaction. Unfonunately, he
did not solve the model equation and so he missed the point recognized later by Nir and Pismen
[30] : in the intermediate region of Thiele modulus the catalyst effectiveness factor is enhanced
by intrapanicle convection (Figure 18).
 232

o
Figure 18. 1') de I TJ d versus Thiele modulus <I> s and intraparticle Peclet number Am

The important parameter in the analysis is the intraparticIe PecIet number ')...=VotlDe relating
the intraparticIe velocity va with diffusivity De in a slab with half thickness e. In 1982 Rodrigues
et al. [31] showed that the reason was that diffusivity is augmented by convection; therefore the
augmented diffusivity is De is:

- 1
De = De f(')...) [14]

where
311
£0.. ) = X (tanh')... -X) [15]

The enhancement factor of diffusivity is thus 1/f(I..), shown in Figure 19.

100~-----------------------------,

D.ID.
10

1~------~~------~------~~
.1 10 100

Figure 19. Enhancement factor versus intraparticle Peclet number A

 233

This is the reason why perfusion chromatography using large-pore packings has improved
perfonnance compared with conventional supports, although the original patent [32, 33] failed to
mention the key result in Equation 14. In fact, in HPLC using large-pore, permeable packings
the classic Van Deemter equation for HETP (height equivalent to a theoretical plate) has to be
modified and replaced by Rodrigues equation [34, 35, 36] to take into account the contribution
of intraparticle flow.

2.1. High Pressure Liquid Chromatography (HPLC)

A model for linear HPLC includes the following dimensionless equations in terms of axial
coordinate x=z/L, particle coordinate p=z'/l and 9=tlt :

species mass balance in the outer fluid phase:

[16]

l-Ep UL
where b=l+ E m and Pe=.,.. n .
P <-D'-'ax

mass balance inside particle for species i:

[17]

equilibrium adsorption isotherm

qi'=mCi'
Boundary and initial conditions
x=O q'=Md(9)
x Ci limited
9=0 q=q'=O x and p
p=O and p=2, Ci'=Cis'

The HETP is defined as cr2l.JIlI and is given by:

[18]

or
HETP = A + - + C
B
u
m. )u Rodrigues equation [19]
 234

2 Ep(1-Eb)J>btd . 8 . ed b Rodri I [34] In th I .

where C = 3' [Eb+Ep(1-Eb)b]2' Equatton 1 was denv y gues et a . e c aSSlC

Van Deemter equation [37] f(;1.)=1 (no intraparticle convection). At high superficial velocity or
high A, f(A)= t and the HElP reaches a plateau at:

[20]

since Vo=alU.

It appears that intraparticle convection contributes to a better efficiency of the

chromatographic column since we obtain lower HElP than with conventional adsorbents and
also the speed of separation can be increased without loosing efficiency.

Figure 20 shows Van Deemter equation for HETP of columns packed with conventional
supports (dashed line) and Rodrigues equation for large-pore packings (full line). Figures 22a
and b show the response of a chromatographic column to an impulse and step input of
concentration, respectively. Dashed lines refer to conventional supports and full lines to large-
pore supports.

c..
f-
UJ
:c
1
• u
Figure 20. HETP as function of the superficial velocity
 235

25
a
E(S) 20 1\
15

10

0
0.4
-,/ 1'""t-"-
0.6 0.8 a 1.0

1.0

7./~' b
F(S) ,.
,,
,/
0.5
,J
,•
I

..,//)
0.0
0.4 0.6 0.8
a 1.0

Figure 21. Response of a HPLC column to impulse and step functions

(Dashed lines: conventional supports; full lines: large-pore packings)

2.2. Pressure Swing Adsorption (PSA)

In gas phase adsorption processes intraparticle convective flow should also be considered if
there are important large pressure gradients as it is the case of pressurization and blowdown
steps in PSA. We have been involved in modeling PSA and details can be found in a recent
series of papers [38-42]. Model equations in dimensionless form are given below using the
following variables:

z z' cP c'P' u v t
x =L' P=I' f =Co =Po • r =Co =Po' u* =Uo • v* =Vo' S='to
 236

where Co= ~ is the total concentration at atmospheric pressure, Vo = ~ ~ is a reference

intraparticle velocity, to= £L is the reference space time and Uo is the bulk: fluid superficial
Uo
velocity at the bed inlet in steady-state at a given pressure drop Mlo = Ph - Pe (here Pe= Po), i.e.:

- Lai/Ph + ~ (Lal/Ph)2 + 2La2( 1 - (Pe/Ph)2)

Uo = 2La2 [21]

with

[21a]

Intraparticle diffusion + convection model

Mass balances inside the adsorbent particle:

l.. (_f_ ~ + Y1. CJf) '1 CJ(V*f'YA) = N_ (1 + "p) CJ(~9yj.J [22]

CJp b4 + f CJp b4 CJp - I\() CJp "'" '> 0

CJ (1... CJf) _ CJ(v*f') _ [CJf' "CJ(~y•.\) ]

I"O-V -
'1
dp b4 dp 0.0 de + '>p 9 [23]

where SP = ~
£
p
m is the adsorbent capacity parameter. The boundary conditions are:

p=<>, yA= YA - ~ ~R ; f' =f - M~R [24]

p = I, YA=YA+~~R; f'=f+¥Xh [25]

and initial condition:

9=0, 'v'x,p [26]

where fo = fe for pressurization and fo = fh for blowdown.

Momentum equation for the fluid inside particle:

v*=~=-
Vo
~e~
op
[27]
 237

Mass balances for the bulk fluid phase in a bed volume element:

Species A

a (u*f~) _o(u*fYA) =O(fyA) + 1-£ N [28]

dx Pe ()x ax "d6" £ A

Overall

O~*f) ()f 1-£ N 0 [29]

x +d9 + T =

where dimensionless fluxes of species A and ovemll are given by:

()p - U.
- (- b4 f+ f' ~ ()f
b4 dp '1
+ ",<>V *f' ')I]
YA p:l [30]

N= ~ [ (- ~4 ~ + Aov*f') Ip:o - (- ~4 ~ + Aov*f') Ip:l ] [31]

Momentum equation for the bulk fluid:

()f
-dx =bsu* + b6 f (u*)2 [32]

Boundary conditions associated with eqs.28 and 29 are :

Pressurization:

x=o, [33]

x= 1, [34]

The initial condition is:

9=0, [35]

where fo =ft for pressurization.

The definitions of the parameters, ~R, ~. bl. b2. b3, b4, bS. 1>6. Pe, A, ~. n, no, ~ can be

found in Table II.

 238

Intraparticle diffusion model

In the absence of intraparticle convection, Ao = 0; therefore, the intraparticle diffusion +

convection model reduces to the intraparticIe diffusion model.

Equilibrium model:

If there are no mass transfer resistances inside the particle, the intraparticle diffusion model
reduces to the equilibrium model.

Model equations are :

In the bulk fluid phase:

a (u*f~) a(~YA)
dx Pe ax - a(U*fYA)
ax --[1 + Ep l-E
E
(1 l')]
+ '>P e [36]

[37]

Table II. The definition of parameters

--~--------,----------------

~e =E~~X =ubif + bz
A. = Aov*(b4 + f)
Ao- tft2~
- DmoEpe
(l = (loU*(b4 + f)
tc t2 !!Q.
no = Dmo eL
t
PR =2L
Pe =i IPo
Simulations show that the final profile of the mole fraction of the adsorbable species
calculated with the intraparticIe diffusion/convection model is between profIles predicted by the
diffusion and the equilibrium model (Figure 22).
Figure 22. Final axial mole fraction profiles in pressurization with equilibriwn (I), diffusion (3)
and diffusion/convection model

3. CONCLUSIONS

Sorption processes are becoming more important in chemical process industries as a result
of biotechnology developments, energy and environment constraints.

The first message we tried to give is on the methodology used to study sorption processes
based on the measurement of parameters governing the behavior of fixed beds by simple
independent experiments ;those parameters are then used in the fixed-bed adsorber model.

The second message is related with modeling. It is first of all a learning tool before it
becomes a design, operation and process optimization aid. Numerical methods used for the
solution of various model equations include: collocation methods, finite differences, moving
finite element methods, Lax-Wendroff with flux correction.

Available packages are PDECOL, COLNEW, FORSIM together with our own software
[43-44].

The third message is to encourage development of new processes and concepts. New
processes arise sometimes by coupling separation and reaction, or by using flow reversal or by
simulating moving bed as Broughton did. The new concept of augmented diffusivity by
convection is a rich one. In separation processes large-pore packings contribute to the
improvement of sorption efficiency; in fact, column responses are driven from diffusion-
controlled to equilibrium-controlled limits by intraparticle convective flow. The result is
equivalent to that observed with reactions in large-pore catalysts; in such case conversion at the
 240

reactor outlet moves from the diffusion-controlled to the kinetic-controlled limit as a result of
intrapanicle convection.

HPLC is a separation process in which the effect of intraparticle convection is important

when large-pore particles are used. Proteins separation contributed to focus the attention of
researchers; therefore the interest on perfusion chromatography increased recently.

PSA is an industrially important process. It has been shown that intraparticle convection is
important in pressurizationlblowdown steps. It is expected that the cycle performance will be
influenced by intraparticle convective flow when cycles are short as in rapid pressure swing
adsorption.

The last message is related with multicomponent systems. A crucial step is the prediction of
multicomponent sorption equilibria. Most of the calculations are based in multicomponent
equilibrium isotherms which are easy to implement in fixed-bed models. However ,even if a
model based on Ideal Adsorption Solution (lAS) is used fixed-bed calculations become time
consuming since at each step one has to solve the iterative algorithm for lAS leading to
eqUilibrium composition. We believe that we have to focus on the representation of
multicomponent equilibrium first; then describe it by some working relationships to be finally
used in the fixed-bed model avoiding the time-consuming iterations.

References

I. Rodrigues, A: Modeling of percolation process. In Percolation Processes: Theory and Applications, ed. A
Rodrigues andD. Tondeur, pp 31-81, Nijtholfand Noordhoff, 1981
2. Rodrigues, A: Mass transfer and reaction in fixed·bed heterogeneous systems: application to sorption operations
and catalytic reactors. In Disorder and Mixing, ed. E. Guyon, et. ai, pp 215-236, Kluwer Acad. Pubs., 1988
3. De Vault, D.: The theory of chromatography. J. Am. Chern. Soc., 65,532 (1943)
4. Thomas, H. C.: Heterogeneous ion exchange in a flow system. J. Am. Chern. Soc., 66,1664 (1944)
5. Bohart, G. and Adams, E.: Some aspects of the behavior of charcoal with respect to diclorine. 1. Am. Chern. Soc.,
42,523(1920)
6. Walter, 1.: Rate dependent chromatography adsorption. J. Chern. Phy., 13,332(1945)
7. Rodrigues, A: Theory oflinear and nonlinear chromatography. In Chromatographic and Mernbrane Processes in
Biotechnology, ed. C. Costa and 1. Cabral, Kluwer Academic. Pubs., 1991
8. Arnold, F., Schofield, S. and Blanch, H.: Analytical affinity chromatography. 1. Chroma!. 355, 1-12 (1986)
9. Chase, F.: Prediction of performance of preparative affinity chromatography. 1. Chroma!. 297, 179-202 (1984)
10. Rosen, J.: Kinetics of fixed-bed systern for solute dilfusion into spherical particles. Ind. Eng. Chern. 20(3).
387 (1952)
II. Rodrigues, A: Percolation theory I - Basic principles. In Stagewise and Mass Transfer Operations, ed. J. Calo and
E. Henley, AIChEMI., 1984
12. Rodrigues, A and Costa, C.: Fixed-bed processes: a strategy for modeling. In: Ion Exchange: Science and
Technology, pp 272-287, M.Nijthoff, 1986
13. Costa, C. and Rodrigues, A: Design of cyclic fixed-bed adsorption processes 1. AIChE J. 31, 1645-1654 (1985)
14. Costa, C. and Rodrigues, A: Intraparticle dilfusion of phenol in macroreticular adsorbents. Chern. Eng. Sci., 40,
983-993 (1985)
 241

15. Wilhelm, R Rice, A and Bendelius, A.: Parametric pwnping: A dynamic principle for separating fluid mixtures.
Ind. Eng. Chern. Fund.,5, 141-144 (1966)
16. Ramalho, E., Costa, C., Grevillot, G. and Rodrigues, A.: Adsorptive parametric pwnping for the purification of
phenolic effluents. Separations Technology, 1,99-107 (1991)
17. Costa, C., Grevillot, G., Rodrigues, A and Tondeur, D.: Purification of phenolic waste water by parametric
pwnping. AlChE 1., 28, 73 (1982)
18. Almeida, F. Costa, C, Grevillot, G. and Rodrigues, A: Removal of phenol from waste water by recuperative mode
parametric pwnping. In: Physicochemical Methods for Water and Wastewater Treatment. ed. 1. Pawlowski,
pp.169-178, Elsevier, 1982
19. Ferreira, 1. Costa, C and Rodrigues, A: Scaleup of adsorptive parametric pwnping. Annual AlChE mt., LA, 1991
20. Tsewtt, M.: On a novel class of adsorption phenomena and their use in biochemical analysis. Trudi Varshavskogo
obstchestva estestvoispitatelei, 40, 20-39 (1903)
21. Nicoud, R and Bailly, M.: Choice and optimization of operating mode in industrial chromatography. In: PREP-92,
ed. M. Perrut, 1992
22. Bailly, M. and Tondeur, D.: Reversibility and performance in productivity chromatography. Chem. Eng. Process.,
18,293-302 (1984)
23. Bailly, M. and Tondeur, D.: Two way chromatography. Chern. Eng. Sci., 36, 455-469 (1981)
24. Broughton, D.: Adsorptive separations-liquids. In: Kirk-OthmerEnc. of Chern. Techn., VoU, 1. Wiley, 1978
25. De Rosset, A: Neuzil, R and Broughton, D.: Industrial application of preparative chromatography. In: Percolation
Processes: Theory and Applications, ed. A Rodrigues and D. Tondeur, pp 249, Nijtholf and Noordhoff, 1981
26. Johnson, 1.: Sorbex: continuing innovation in liquid phase adsorption. In: Adsorption: Science and Technology,
pp383-395, M.Nijthoff, 1989
27. Stord, G., Masi, M. and Morbidelli, M.: On counter current adsorption separation processes. In Adsorption: Science
and Technology, pp 357-382, M. Nijhoff, 1989
28. Nielsen, A, Bergd, S. and Troberg, 8.: Catalyst for the synthesis of ammonia and method of producing it. US Patent
3,243,386, (1966)
29. Wheeler, A: Reaction rates and selectivity in catalyst pores. Adv. in Catalysis, 3, 250- 337 (1951)
30. Nir, A. and Pismen, L.: Simultaneous intraparticle convection, diffusion and reaction in a porous catalyst. Chem.
Eng. Sci., 32, 35-41 (1977)
31. Rodrigues, A, AIm, B. and Zoulalian, A: Intraparticle forced convection effect in catalyst diffusivity measurement
and reactor design. AlChE 1., 28, 541-546 (1982)
32. Afeyan, A Regnier, F. and Dean, R: Perfusive chromatograph. US Pat. 5,019,270, (1990)
33. Afeyan, A Gordon, N., Mazsareff, I. ,Varady, 1.,Fulton, S., Yang, Y. and Regnier, F.: Flow through particles for
the HPLC separation ofbiomolecules: Perfusion chromatography. 1. Chromat., 519, 1-29 (1990)
34. Rodrigues, AE., Lu, l.P. and Loureiro, 1.M.: Residence Time Distribution of Inert and Linearly Adsorbed Species
inFixed-Bed Containing 'Large-Pote' Supports: App. in Sep. Eng. Chern. Eng. Sci., 46, 2765-2773 (1991)
35. Rodrigues, AE., Lopes, 1.C., Lu, l.P., Loureiro 1. M. and Dias, M.M.: Importance oflntrapartic1e Convection on
the Performance of Chromatographic Processes. 8th. Intern. Symp. Prep. Chromat. "PREP-91", Arlington, VA; 1.
Chromatography, 590. 93-100 (1992)
36. Rodrigues, A: An extended Van Deemter equation (Rodrigues equation) for the performance of chromatographic
processes using large-pore, permeable packings. submitted to LC-GC, 1992
37. Van Deemter, 1. luiderwveg, F. and K1inkenberg, A: Longitudinal diffusion and resistance to mass transfer as
causes of nonideality in chromatography. Chern. Eng. Sci., 5, 271-289 (1956)
38. Lu, l.P., Loureiro, 1.M., LeVan, M.D. and Rodngues, AE.: Effect of IntraparticIe Forced Convection on Gas
Desorption from Fixed-Bed Containing 'Large-Pore' Adsorbents. Ind. Eng. Chern. Res., 35, 1530 (1992)
39. Lu, l.P., Loureiro, 1.M., LeVan, M.D. and Rodrigues, AE.: Intraparticle Convection Effect on Pressurization and
Blowdown of Absorbers. AlChE 1.,38, 857-867 (1992)
40. Rodrigues, AE., Loureiro, 1.M. and LeVan, M.D.: Simulated Pressurization of Adsorption Beds. Gas Separation
and Purification,S, 115 (1991)
41. Lu, l.P., Loureiro, J.M., LeVan, M.D. and Rodrigues, AE.:lntrapartic1e Diffusion! Convection Models in
Pressurization and Blowdown: Nonlinear Equilibrium", Sep. Sci. Technol., 271 (4), 1857-1874 (1992)
42. Lu, l.P., Loureiro, 1.M., LeVan, M.D. and Rodrigues, AE.: Pressurization and Blowdown of an Adiabadc
Adsorption Bed: IV Diffusion!Convection Model. Gas Sep & Purif. 6 (2), 89-100, (1992)
43. Sereno, C. Rodrigues, A and Villadsen, 1.: Solution of partial differential equations systems by the moving finite
element method. Computers Chern Engng., 16,583-592 (1992)
44. Loureiro,1. and Rodrigues, A: Two solution methods for hyperbolic systems of partial differential equations in
chemical engineering. Chern Eng. Sci.,46, 3259-3267 (1991)
 Monitoring Batch Processes

John F. MacGregor and Paul Nomikos

Department of Chemical Engineering, McMaster University, Hamilton, Ontario, Canada LaS 4L 7

Abstract: Two approaches to monitoring the progress of batch processes are

considered. The first approach based on nonlinear state estimation is reviewed
. and the problems in implementing it are discussed. A second approach based on
multi-way principal components analysis can be developed directly from
historical operating data. This approach leads to multivariate statistical process
control plots which are very powerful in detecting subtle changes in process
variable trajectories throughout the batch. This method is evaluated on a
simulation ofthe semi-batch emulsion polymerization of styrenelbutadiene.

Keywords: Batch monitoring, state estimation, statistical process control, fault

detection

Introduction

Monitoring batch reactors is very important in order to ensure their safe

operation, and to assure that they produce a consistent and high quality product.
Some of the difficulties limiting our ability to provide adequate monitoring are
the lack of on-line sensors for product quality variables, the highly nonlinear
nature and finite duration of batch reactors, and the difficulties in developing
accurate mechanistic models that characterize all the chemistry, mixing, and heat
transfer in these reactors.
Current approaches to achieving consistent, reproducible results from
batch reactors are based on the precise sequencing and automation of all the
stages in the batch sequence. Monitoring is usually confined to checking that
 243

these sequences are followed, and that certain reactor variables such as
temperature are following acceptable trajectories. In some cases, on-line energy
balances are used to keep track of the instantaneous reaction rate, and the
conversion or the residual reactant concentrations in the reactor [2, 11, 19J.
In this paper, we consider two advanced model-based approaches to batch
monitoring. The first approach is based on state estimation, and combines
fundamental nonlinear dynamic models of these batch processes with on-line
measurements in order to provide on-line, recursive estimates of the fundamental
states of the process. Since this approach is well known and has an abundant
literature, we only provide an overview of its main ideas and some of the key
references.
The second approach is based on empirical multivariate statistical models
which are easily developed directly from past production data on the large number
of process variables such as temperatures, pressures and flows measured
throughout the batch. This new approach is based on Multi-way Principal
Component Analysis (MPCA) and Projection to Latent Structure (PLS) methods.
It will be discussed in greater detail and some examples will be presented.
Both of the above approaches rely upon the observability of the states or
events of interest. If the data contain no information or very little information on
certain states or events, then no effective monitoring scheme for them is possible.

State Estimation

Theoretical or mechanistic models for batch or semi-batch reactors usually take

the form of a set of ordinary nonlinear differential equations in the form
dx d
- = c<i (x u t) (1)
dt ' ,

y = hex, u, t) (2)

Xo = x(t = 0) (3)

where x represents the complete vector of internal states, xd is the deterministic

subset of the differential state vector x described through eq. (1), y is the vector of
measured outputs that is related to x through eq. (2), and u is a time-varying
vector of known manipulated inputs. The complete state vector x is assumed to be
made up of a deterministic component xd and a stochastic component xs. Included
 244

in X s are model parameter and disturbance states that may vary with time in some
stochastic manner and may be unknown initially.
The objective of the state estimation problem is to predict the internal
states x(t) from a limited set of sampled,and noise-corrupted measurements y(tk).
It is assumed that some elements of xo will be initially unknown and that some
states will be time-varying disturbances and/or fIxed parameters that must be
estimated.
These are several requirements needed for the successful application of
state estimation. One must have a good mechanistic model that captures the
main physical and chemical phenomena occurring in the reactor. A set of
measurements (y) is necessary that not only makes the states of interest
observable, but also ensures that the state estimation errors will be small enough
to detect the state deviations of interest. As pointed out by MacGregor et al [11]
and Kozub and MacGregor [8], a common error in formulating fIlters is neglecting
to incorporate adequate disturbance and/or parameter states (x s). These are
needed to eliminate biases in the state estimates and to yield a robust estimator
when there are modelling errors or unknown disturbances in the system. Failing
to incorporate such nonstationary stochastic states leads to a proportional type of
state estimator without the integral behaviour necessary to eliminate the biases.
The most common form of nonlinear state estimator is Extended Kalman
Filter (EKF) but various other forms of second order fIlters, reiterative Kalman
Filters, and nonlinear optimization approaches have been suggested. Kozub and
MacGregor [8, 9] investigated the use of the EKF, the reiterative EKF and a
nonlinear optimization approach in monitoring semi-batch emulsion polymeriza-
tion reactors, and concluded that the EKF with reiteration and a second filter to
estimate the initial states was the preferred approach. In other situations linear
Kalman Filters based on semi-empirical models are adequate. Stephanopoulos
and San [17] used such filters with non-stationary growth parameter states to
monitor fed-batch fermentation reactors.
The state estimator provides on-line recursive estimates of important
process states Xd(tkltk) and stochastic disturbance or parameter states xs(tk I tk)
thereby enabling one to monitor the progress of the batch reactor. The number of
non-stationary stochastic states or parameters estimated cannot be greater than
the number of measured output variables (y). One can also monitor the per-
formance of the filter itself by plotting the innovations (y(tk) -y(tk I tk-l». The
variance ofthe innovations is sometimes used to adapt the Kalman Filter gain.
On-line reactor energy balances are very effectively implemented using
Kalman Filters [11, 2, 14]. The objective is to combine the energy balance
 245

equations with simple flowrate, and temperature measurements taken around the
reactor and its cooling system to track stochastic states such as the instantaneous
heat release due to reaction (qR)' and the overall heat transfer coefficient (UA).
Unsafe or undesirable batch reactor conditions such as the beginning of a run-
away reaction, or excessive fouling of the reactor heat transfer surfaces can then
be detected. In semi-batch reactors where reactants are being fed continuously
these state estimates can be used to detect the unsafe accumulation of reactants in
the reactor that may occur if there is a temporary reaction break-down due to
poisoning, etc. [15].
Kalman Filters based on models which include more detailed 'kinetic
phenomena and more informative sensors can be used to monitor species concen-
trations and molecular properties. Kozub and MacGregor [8] illustrate this in
monitoring a semi-batch styrene-butadiene emulsion polymerization reactor.
Particle size and concentration, and polymer composition and structure were
monitored together with stochastic states such as impurity concentrations. Such
detailed models allow for detection of more specific problems such as particle
coagulation, impurity contamination, or feedrate errors. These state estimators
can also be used to implement nonlinear control over polymer property
development [9].
An alternative approach aimed at providing more specific fault detection
and diagnosis is to run several parallel filters, each based on a different set of
plausible events or faults. Based on the innovations (X(tk) - X(tk I tk -1» from
each filter the posterior probability of each model being valid can be evaluated at
each sampling interval. A high probability for any model which characterizes an
undesirable event would lead to an alarm and a response. Such an approach was
used by King [7] to monitor batch reactors for the onset of undesirable side
reactions.
A major difficulty in practice with monitoring batch reactors by state
estimators is the need for detailed mechanistic models, and some specific on-line
sensors related to the variables of interest. Even when these models are developed
some parameters in both the models and the filters must be adjusted to ensure
that the resulting filters can track the actual processes. The advantage of such an
approach is that it is "directional" in nature since the incorporation of mechanistic
understanding into the choice of the state vector x = (xd, xs)T allows one to make
inferences about the nature of any faults as well as their magnitudes.
 246

Empirical Monitoring Approaches

Although good theoretical models of batch processes and on-line sensors for funda-
mental quality properties are often unavailable, nearly every batch process has
available frequent observations on many easily measured process variables such
as temperatures, pressures, ariaflowrates. One may have up to 50 measurements
or more every few seconds throughout the entire history of a batch. Furthermore,
there is usually a history of many past successful (and some unsuccessful) batches.
From this data it should be possible to build an empirical model to characterize
the operation of successful batch runs. The major difficulties are how to handle
the highly correlated process variables, and the large number of multivariate
observations taken throughout the batch history.
In this section, we develop some multivariate statistical process control
methods for monitoring and diagnosing problems with batch reactors which make
use of such data. The approach used in based more on the statistical process
control (SPC) philosophy of Shewhart [16] than that of feedback control. In SPC
one usually assumes that under normal operation, with only common cause
variations present, the system will operate in some stable state of statistical
control, and will deviate from this behaviour only due to the appearance of special
causes. The approach is therefore to develop some statistical monitoring proce-
dures to detect any special event as quickly as possible, and then look for an
assignable cause for the event. Through such a procedure one can gradually make
continuous improvements to the process. Traditionally univariate SPC charts
such as the Shewhart chart have been used to monitor single variables. However,
these approaches are inappropriate when dealing with large multivariate
problems such as the one being treated here. Statistical process control charts
based on multivariate PCA ad PLS methods have been developed for steady-state
continuous processes [10], but for batch processes where the data consists offinite
duration, time varying trajectories little has been done. Therefore, in this paper
we develop monitoring procedures based on multi-way principal components
analysis (MPCA). This method extracts the essential information out of the large
number of highly correlated variables and compresses it into low-dimensional
spaces that summarize both the variable and time histories of successful batches.
It then allows one to monitor the progress of new batches by comparing their
progress in these spaces against that of the past reference distribution.
Multivariate factor analysis methods (closely related to principal
components) have recently been used by Bonvin and Rippin [3] to identify
 247

stoichiometry in batch reactors. This represents an approach which combines

some fundamental knowledge with a multivariate statistical approach to monitor
more specific features in batch reactors, but does not provide a general framework
for on-line monitoring of the progress of batch reactors.

Multi-Way Principal Components Analysis (MPCA)

The type of historical data one would usually have available on a batch process is
illustrated in Figure 1. For each batch (i = 1, ... ,1), one would measure J
variables at K time periods throughout the batch. Thus one has a three-
dimensional array of data ~(i,j, k); i = 1, ... , I;j = 1, ... , J and k = 1, ... , K. The top
plane in the array represents the data on the time trajectories for all J variables in
the first batch. Similarly, the front plane represents the initial measurements on
all J variables for each of the I batches.

K
time
k

2
1~______~~__~
2

batches
i

I
1 2 j J
variables

Figure 1: Data array ~ for a typical batch process.

If we only had available a two-dimensional matrix (X) such as the matrix

of variables versus batch at a given time k, then ordinary principal components
 248

analysis (PCA) could be used to decompose the variation in it into a number of

principal components. After mean centering (ie subtracting the mean of each
variable), the fIrst principal component is given by that linear combination of
variables exhibiting the greatest amount of variation in the data set (tl = X PI).
The second principal component (t2) is that linear combination, orthogonal to the
fIrst one, which exhibits the next greatest amount of variation, and so forth. With
highly correlated variables, one usually fInds that only a few principal
components (tl, t2, ... , tA) are needed to explain most of the signifIcant variation in
the data. The (IxJ) data matrix can then be approximated as sum of A rank one
matrices
A
X =T pT = '"
L t 8 pT
8
8=1

where the "score" vectors ta are mutually orthogonal and represent the values of
the principal components for each object (i). The loading vectors Pa show the
contribution of each variable to the corresponding principal component. Principal
components analysis is described in most multivariate statistics texts [1, 6].
However, the projection aspects of PC Aand the NIPALS algorithm for computing
principal components sequentially that are used in this paper are best described
in Wold etal [18].
Since in the batch data array of Figure I, we are interested in analyzing
the variation with respect to variables, batches and time a three-dimensional
PCA is needed. Such methods have been developed for the analysis of
multivariate images [5], and we use a variant of this approach in this paper.
Multi-way PCA decomposes the X array into score vectors (ta ;
a = 1,2, ... , A) and loading matrices P a such that
A

~ =L t8.® p. + E
8=1
There are three basic ways in which the array can be decomposed but the
most meaningful in the context of batch monitoring is to mean center the data by
subtracting the means of the variables for each time over the I batches. In this
way, the variation being studied in MPCA is the variation about the average
trajectories of the variables for the I batches. In this way, the major nonlinear
behaviour of the batch process is removed through subtracting the mean
trajectories, and linear MPCA will be used to analyze variations about these mean
trajectories.
The loading matrices P a (a = 1, 2, ... , A) will then summarize the contri-
butions ofthe variables at different times to the orthogonal score vectors tao These
 249

new variables ta =! 0 P a are those exhibiting the greatest variation in the

variables over the time ofthe batch.
The NIPALS algorithm for MPCA fonows directly from that for ordinary
PCA, and the steps are given below.
o. scale the array!, set!! = !
1. take randomly a column from!! and put it as t
Start of minor iterat!uns
2. P = !!'.t
3. P = PIIIPII
4. t =!!o P
5. if the t has converged then go to step 6 else go to step 2
6. !! = !!-t® P
7. go to step 2 for the calculation of the second principal component

Post.Analysis and On· line Monitoring of Batch Processes Using MPCA

Plots

The information extraction and data compression ideas of MPCA can be used to
perform post·analysis of batch runs to discriminate between similar and
dissimila.r runs, and to develop on·line methods for monitoring the progress of new
batches. We shan concentrate here on the development of on·line monitoring
methods. The approach follows closely that of Kresta et al [10] for monitoring the
operating performance of continuous processes.
From the historical data base, a representative sample of successful batch
runs can be selected. These would normally comprise all those batches that
resulted in good product quality. The variable trajectory data array (~) as shown
in Fig. 1 can be assembled using the data from these runs, and an MPCA analysis
performed on this array.
The progress of these "good" batches can be summarized by their
behaviour in the reduced principal components space T = (tI, t2, "., tA). The
behaviour of these batches with time will be confined to a region in this space.
This region will therefore define a reference distribution against which we can
assess the performance of other past batches or new batches. The principal
components calculated for other good batches should fall close to the hyperplane
defined by T, and they should fall in the region of this plane defined by the
 250

previous good batches. The acceptable regions in the T-space can be defmed using
multivariate Normal distribution contours with variances calculated from the
estimated score vectors (ta ), or if the reference sample contains a sufficient
number of batches an approximate 99% contour can be defmed directly as the
contours enclosing approximately 99% of the scores from these batches. Post-
analysis,' that is the analysis past batches for which the complete history is
known, can be summarized by plotting the fmal value of the t-scores for any given
batch and comparing it against this reference distribution of t-scores from other
batches.
A problem arises in the on-line monitoring of new batches because
measurements on the variables are not available over the complete batch as they
were with past batch runs. Instead, measurements are only available up to time
interval k. There are several approaches to handling this missing data, and we
shall use here the rather conservative approach of setting all the values of the
scaled, mean-centered variables beyond the current time k to zero. This means
that we are giving the new batch the benefit of the doubt by implying that the
remaining portion of the batch history will have no deviation from the mean
trajectory. Therefore, in monitoring a new batch the following procedure is used:
1. Take the new vector of measurements at time k
2. Mean center and scale them as with the reference set.
3. Add this new observation as the k-th row in Xnew and set the rows from
(k + 1) onward equal to zero. A
4. Calculate the new scores ta = Xnew P a; E = Xnew - ~ ta ® P a
a=1
5. Return to step 1.
In monitoring the progress of a new batch there are several ways in which
an excursion from normal operation can show up. Hthe process is still operating
in the same way as the batches in the reference data base, but simply exhibits
some larger than normal variations, this behaviour should show up as the scores
(ta's) for the new batch moving outside the control region in the T-space.
However, if a totally new fault not represented in the reference data base were to
occur, at least one new principal component vector would be needed to describe it.
In this case the computed score values of the new batch would not be predicted
well by the MPCA model since they would fall off the reduced space of the
reference T-plane. To detect all such new events that would show up in this way
we plot the squared prediction error (SPE) or equivalently the squared perpen-
dicular distance from the reference T-plane for each new observation from the new
batch. To assess the significance of any increase in SPE we place an upper control
 251

limit on the SPE above which there is only approximately a 1% probability of

occurring if the new batch is on target. This control limit can be calculated in
various ways from the variations ofthe calculated SPE's in the reference set [13].

Example: Semi-Batch Styrene-Butadiene Emulsion Polymerization

Styrene-butadiene rubber (SBR) is made by semi-batch emulsion

polymerization for use in adhesives, coatings, footwear, etc. A detailed modelling
study on the SBR processes was performed by Broadhead et al [4]. A modification
of this model was used in a simulation study to evaluate these MPCA monitoring
methods. Using typical variations in the initial charge of materials and
impurities, and in the process operations a number of batches were simulated.
Fifty batches which gave final latex and molecular properties within an
acceptable region were selected to provide a reference data array. On-line
measurements were assumed to be available on nine variables: the feed rates of
styrene and butadiene monomers, the temperatures of the feed, the reactor
contents and jacket contents, the latex density, the total conversion, and the
instantaneous heat release from an energy balance. Using 200 time increments
over the duration ofthe batch the reference data set! was a (50 X 9 X 200) array.
To evaluate the ability ofMPCA to discriminate between "good" and "bad"
batches a post-analysis was performed using the 50 good batches plus one bad
batch. In one case the ''bad'' batch had a 33% higher level of organic impurities in
the butadiene monomer feed to the reactor right from the beginning its operation.
The other "bad" batch had a 50% higher level of organic impurities, but this time
the contamination started half-way through its cycle (at time = 100). The score
plots for the first two principal components (tl, t2) are shown in Figure 2. The two
''bad'' batches, denoted as point "51", clearly do not belong to the family of normal
batches. Therefore in this case the MPCA plots were easily able to detect
abnormal operation of the batch.
In order to implement an on-line monitoring scheme for new batches a
MPCA model was developed from the historical records of 50 good batches. Four
significant principal components were needed to capture the predictable variation
about the average trajectories of the variables in the batches. Plots of the first two
principal components (tl, t2) and the SPE versus time are shown in Figure 3 for
the evolution of a new "good" batch run. The principal components remain
 252

30 .12

Batch with initial problem

20

10
•
Sl
/ •26
•1(~~ 6
••
:,l.1J.:
29. ~~7
.1i
•

25
.17 ·l3 • •
'"
.. iJ...~ .
.JiO
I-
0
~9~ ..
•3~
4~3

.
-10

33
•3S 48
-20 •'0 •50 l

-30 39
•
-50 0 50

T1

20

N
I-
o

-20

-w Batch with problem hllif-way

51 ------- through its operation
•
-60 -w -20 o 20 60

T 1

Figure 2: Post analysis of batch data in the score space.

 SP£ OI.J'T Of' \Jl,Ilf I T O<IT or lMfS 0 TOVTorwiTS 2
l51.-------~--------_r--------~------_,

JO

~ 2.
~
...~
~
20 ~ ~
~ ! ~
Ii!
I( .5
I~I ~ Ii!
g:
. E E
.. ... I\)
~
.Or M _! •
- I.·.
• '.,/
~
.. -: . .._.
~
.;
...
.. .. CJo
W

,~/-: ~
. .'r-t. ...- ;....~.'" ..'~:..:.J... ~ I'
. ~. .~-.: -:-., "...._.1;,
.....' .
0'o ;0 100 1~ 200
50 150 2CO
T"'( T1IJ( nUE

Figure 3: Monitoring of a new "good" SBR batch.

 254

between their upper or lower control limits, and the SPE lies below its control
limit throughout the duration of the batch indicating that the progress of this new
batch was well within the acceptable range of variation dermed by the reference
distribution. The SPE control limits shown here are approximately 99% limits
based on the reference distribution with only 50 samples, and are therefore quite
erratic. Improved estimates of this upper control limit can be obtained.
Figure 4 shows the same monitoring plots for the new SBR batch in which
there is a 33% higher level of organic impurities in the butadiene monomer feed to
the reactor starting at time zero. The principal component plots and the SPE plot
detect this event very quickly.
Figure 5 shows the monitoring plots for the new SBR batch in which at
time 100 the level of organic impurities in the butadiene monomer feed increased
. by 50%. The final product from this batch did not meet specifications. Although
the principal component plots do not detect a change, the SPE plot rapidly detects
the occurrence of this new event around t = 100.
To further verify that these multivariate monitoring methods are very
powerful for detecting problems which occur in batch processes, the trajectories of
the individual variable measurements for the three batch runs just considered are
plotted in Fig. 6. It can be seen that there is not much observable difference
among the three runs. If all the trajectories from the good batches in the reference
set were plotted on this Figure any differences between these and the bad batches
would be almost undetectable through visual inspection. The power of the
multivariate MPCA method results from using the joint covariance matrix of all
the variable trajectories. By doing this, it utilizes not just the magnitude of the
deviation of each trajectory from its mean, but the correlations among them in
order to detect abnormal operation.

Summary

Methods for on-line monitoring of the process of batch processes have been
presented. Theoretical model-based approaches based on state estimation were
briefly reviewed, and new empirical methods based on multi-way principal
components analysis were presented. The latter statistical monitoring approach
was evaluated on simulations of an SBR semi-batch emulsion polymerization
reactor, and was shown to provide rapid detection of operating problems.
 sp( OOfM """ 137 1 ()Of M uuns 1 ()Of M IAIns '00
!IO
..,
<5 '. ~
<0 , . "
, , ,
~ \ :
35 ",:' ~
I~ ~
"," ,
~ 30 ~ "
20 ~
~ 25 ~
!i1 ~
20
E! ~
~ ~
. 001
i
.S .. -20 <I
~ .. I\)
<.n
'0 ~_I <.n
..
0
0 !IO '00 '!IO 700

nU[ nU[
tII.It

Figure 4: Monitoring of a new SBR batch with impurity contamination in the butadiene feedrate
starting at time zero.
 SPE OUT OF l..MT ee T OUT Of" lMI"S 0 T OUT OF lNII'S 47
lro'i------~--------~------~------__,

140
~
..
&! 120 . '\.
. &! ~
~ 100 . ,. . "
~ ~
~ eo ~
~ ~
~ ~
A. ~
ro !
!.. ,. ... ;
~ ~
1- .'"
~
.i- (J1
""
'"
01;,,'\--- ~
o ..,
~ ~
l1li£

Figure 5: Monitoring of a new SBR batch with impurity contamination in the butadiene feedrate
starting at time 100.
 257

IICU

I-; lOA
0
U
oj
<l) IO.l
I-;
'-H
0
<l) IO.l
C
I-;
'r;;
....
;::l
(iJ ~
<l)

..
I-;
<l) -0
0.
S
<l)
f--<
(~'7 -6"0' ....- "CI<:7

.,.

.... 0 .. .oo

time time

=r
'" OJ

..
I-;
<l)
(iJ
~

.
bIJ
,9 ~
0
-00 'r;;
0
'-H
0 .,
I-;
<l)
>
~
...
0

..
<l) 0
I-; cu
OJ
~I-;
2
<l) Q.2
0.
S
<l) ...
.. .•
0. •
f--<

.oo
C
'" "XI .."

time time

'"
..,
.
....6

-...
<l)
.!<:
u
oj

'-H
0 ....
..
<l)
I-;
S
...
;::l
(iJ

-
;::l
I-; r:/J

8' c.s
<l)

<l)
f--< c
"'"
....
.
JOt)

..,
c .00

time
.oo "XI
• '" '00 ':>Q .."

time
Figure 6: Trajectories of some of the measurements variables during one good batch (solid) and
the two batches with early (dotted) and late (dashed) impurity contamination,
 258

Literature
1. Anderson, T.W.: An Introduction to Multivariate Statistical Analysis, John Wiley &
Sons, New York (1984).
2. Bonvin, D., P. de Valliere and D. Rippin: Application of Estimation Techniques to Batch
Reactors. Part I: Modeling Thermal Effects. Compo Chem. Eng., 13, pp. 1-9 (1989).
3. Bonvin, D. and D.W. Rippin: Target Factor Analysis for the Identification of
Stoichiometric Models, Chem. Eng. Sci.,!§' 3417-3426 (1990).
4. Broadhead, T.O., A.E. Hami!llec and J.F. MacGregor: Dynamic Modelling of the Batch,
Semi-Batch and Continuous Production of StyrenelButadiene Copolymers by Emulsion
Polymerization, Makromol. Chem., Suppl.l0/11, pp. 105-128 (1985).
5. Geladi, P., H. Isaksson, L. Linkqvist, S. Wold and K. Esbensen: Principal Component
Analysis of Multivariate Images, Chemometrics and Intelligent laboratory Systems, 5,
209-220 (1989).
6. Jackson, J.E.: A User's Guide to Principal Components, John Wiley and Sons, New York
(1991).
7. King, R.: Early Detection of Hazardous States in Chemical Reactors, IFAC Symp.
DYCORD'86, pp. 93-98, Bournemouth, U.K., Pergamon Press (1986).
8. Kozub, D. and J.F. MacGregor: State Estimation for Semi-Batch Polymerization
Reactors, Chem. Eng. Sci., 47, 1047-1062 (1992a).
9. Kozub, D., and J.F. MacGregor: Feedback Control of Polymer Quality in Semi-Batch
Copolymerization Reactors, Chem. Eng. Sci., 47, 929-942 (1992b).
10. Kresta, J.V., J.F. MacGregor and T.E. Marlin: Multivariate Statistical Monitoring of
Process Operating Performance, Can. J. Chern. Eng., 69, 35-47 (1991).
11. MacGregor, J.F.: On-line Energy Balances via Kalman Filtering, Proc. IFAC Symp.
PRP-6, pp. 35-39, Akron, Ohio, Pergamon Press (1986).
12. MacGregor, J.F., D. Kozub, A. Penlidis, A.E. Hamielec: State Estimation for
Polymerization Reactors, IFAC Symp. DYCORD '86, Bournemouth, pp. 147-152, U.K.,
Pergamon Press (1986).
13. Nomikos, P.: Multivariate Statistical Process Control of Batch Processes, Ph.D. transfer
report, Dept. ofChem. Eng., McMaster University, Hamilton, Canada (1992).
14. Schuler, H. and C.U. Schmidt: Calorimetric State Estimators for Chemical Reactor
Diagnosis and Control: Review of Methods and Applications, Chem. Eng. Sci., fl, 899-
915(1992).
15. Schuler, H. and K. de Haas: Semi-batch Reactor Dynamics and State Estimation, IFAC
Symp. DYCORD'86, Bournemouth, UK, pp. 135-140, Pergamon Press (1986).
16. Shewhart, W.: Economic Control of Quality, Van Nostrand (1931).
17. Stephanopoulos, G. And K.Y. San: Studies on On-line Bioreactor Identification, I:
Theory, Biotechnol. Bioengng., g§, 1176 (1984).
18. Wold, S., K. Esbensen and P. Geladi: Principal Component Analysis, Chemometrics and
Intelligent Laboratory Systems, 2, 37-52 (1987).
19. Wu, R.S.: Dynamic Thermal Analyser for Monitoring Batch Processes, Chem. Eng.
Progress, Sept. 1985, pp. 57-61 (1985).
Tendency Models for Estimation, Optimization and Control
of Batch Processes

Christos Georgakis

Chemical Process Modeling and Control Research Center and Department of Chemical Engineering. Lehigh University,
Bethiehem,PA 19015, USA

Abstract: This paper summarizes recent progress in the area of estimation and control of batch
processes. The task of designing effective strategies for the estimation of unmeasured variables
and for the control of the important outputs of the process is linked to our need to optimize the
process and its success is depended upon the availability of a process model. For this reason we
will provide a substantial focus on the modeling issues that relate to batch processes. In particular
we will focus attention on the approach developed in our group and referred to as "tendency
modeling" that can be used for the estimation, optimization and control of batch processes.
Several batch reactor example processes will be detailed to illustrate the applicability of the general
approach. These relate to organic synthesis reactors and bioreactors. The point that distinguishes
tendency modeling from other modeling approaches is that the developed Tendency Models are
multivariable, nonlinear, and aim to incorporate all the available fundamental information about
the process through the use of material and energy balances. These models are not frozen in time
as they are allowed to evolve. Because they are not perfectly accurate they are used in the
optimization, estimation and control of the process on a tentative basis as they are updated either
between batches or more frequently. This iterative or adaptive modeling strategy also influences
the controller design. The controller performance requirements and thus the need of a more
accurate model increase as successive optimization steps guide the process operation near its
constraints.

Keywords: Tendency modeling, batch reactors, process control, state estimation, process
optimization
 260

Introduction
Batch processing is an important segment of the chemical process industries. A growing
proportion of the world's chemical production by volume and a larger proportion by value is made
in batch plants. In contrast to continuous processes, batch processes related to the production of
fine and specialty chemicals, pharmaceuticals, polymers, and biotechnology are characterized by
the largest present and future economic growth among all sections of the chemical industry [1].
This trend is expected to continue as industry pursues the manufacture oflow volume, high value
added chemicals, particularly in developed countries with few indigenous raw materials.
In comparison to continuous processes, batch processes are characterized by a greater
flexibility of operation and a rapid response to changing market conditions. Typically, a single
piece of process equipment may be used for manufacturing a large variety of products utilizing
several different unit operations such as reactors, distillation columns, extraction units, etc. As a
result, batch plants have to be cycled frequently and monitored carefully, thereby, requiring higher
labor costs per unit volume of product throughput. At the same time most of the batch processes,
particularly those related to the production of fine and specialty chemicals, are characterized by
significant price differences between the reactants and the products.
Unlike continuous processes, batch processes seldom operate at steady state. This results in
a lack of reproducibility. Most batch processes suffer from batch-to-batch variation in the quality
of the product due to imprecise measurement and control of operating conditions. In the case of
continuous processes, the off-specification material produced during the start-up and transient
operation of the plant can be blended with the good products during normal operation. Batch
processes do not enjoy this luxury. The batch process is in transition most of the time, and the
limited time cycle of a batch does not allow for many corrective actions. If the product resulting
from a batch is not of desired quality, it usually has to be discarded. Because the added value of
the product in relationship to the reactants is so high, the economics of process improvements are
dependent more on whether the batch made an acceptable product rather than whether the amount
of energy or reactants used was the minimum possible. There are significant economic benefits
which can be realized from the optimization of a great variety of batch and semi-batch processes.
The challenge, however, is different from that of the traditional and continuously operated
chemical industry.
Many batch processes are characterized by small annual volumes of production. Frequently,
 261

the annual requirement for a particular product can be manufactured in a few weeks. The plant is
then adapted, and if necessary, re-configured to produce the next product. This makes the
development of detailed models for each process or product economically unattractive. The
frequent process changes that characterize this technology seldom provide enough time for the
development of any model at all. In the absence of such a systematic organization of our
knowledge of the process through a process model, the operation is quite sub-optimal. On-line
information is limited to the available process measurements and the most often used control
action is temperature and/or pressure control. In some rare cases, the controller might utilize the
energy balance of the unit [20, 21]. In almost all cases, one lacks any information as to how much
the process operation can be improved. Any improvement attempts have to be based on a trial and
error approach without the benefit of the quantitative suggestions of a model.
In the chemical and pharmaceutical industries, emphasis is presently placed on the quality
control ofthe product. For example, in emulsion polymers the demand for special properties and
improved performance has recently led to the increased interest in the more detailed understanding
of the inner workings of the process. In the area ofbioreactors, the need for substantial on-line
information is presently needed to ensure that each batch meets regulatory guidelines and provides
a reproducible product quality. Such time dependent knowledge can induce improvements in the
quality of the product produced and can also help increase the process productivity. The more
readily available time dependent information are those provided through the on-line measurements
ofthe process. These measurements might not be directly related to product quality, necessitating
the off-line measurements at the end of the batch. At that time, one can find if the product quality
is appropriate but can do nothing to correct an undesirable situation. For this reason, an on-line
estimation of the quality related parameters is needed through the help of a process model. Since
the development of a detailed and accurate model might be uneconomical, one important issue to
consider is how accurate the model needs to be to achieve the appropriate estimation of the
unmeasured quality variables.
To account for some of these difficulties the investigation of a general purpose modeling,
estimation, optimization, and control strategy has been initiated by some researchers [13, 14, 15,
25, 26, 28, 34] which is directly applicable to several batch processes. This strategy aims to
properly account for our lack of detailed process knowledge and the great variety of batch or
semi-batch reactors. At the same time it takes advantage of new plant data collected during the
process operation to update the model as many times as necessary. The iterative modeling and
 262

optimization methodology proposed initially by Filippi et al. [14, 15], as well as the alternative
approach proposed by Hamer [17] have been modified and extended to develop a more complete
methodology as detailed by Rastogi et al. [28,31]. For example, a systematic procedure was
proposed, based on statistical design of experiment techniques for determining preliminary
experimental runs that will provide the data for initialization of the modeling and optimization
cycle.
In the following sections, we will detail a comprehensive framework needed for the modeling,
optimization and control of batch processes based on the use of a tendency model of the process.
Since such a model is not always very accurate, emphasis will be placed on the model updating
strategy. For this reason, we have called the proposed strategy "Tendency Modeling, Optimization
and Control". In defining such strategy, we have tried not to be constrained by present practice
and hardware limitations. On the other hand, we feel that the strategy defined is a realistic and
achievable one. Several successful industrial applications of the approach have provided
considerable positive feedback.
Even though all the technical details concerning such a strategy are not presently resolved,
substantial progress has been achieved. The following sections will provide a brief overview of the
most recent progress.

Generic Characteristics of TeMOC

In this section, we will examine generic research issues related to the proposed strategy of
tendency modeling, optimization and control of batch processes. These include modeling, state
estimation, contro~ and optimization. The overall Tendency Modeling, Optimization and Control
(TeMOC) strategy can be summarized in the diagram of Figure 1. This diagram describes the
comprehensive structure of the proposed approach for the modeling, optimization, and control of
chemical reactors or other units. Activities involved in each of the schematic boxes will be
summarized here. The central box in the middle ofthe diagram depicts the process. Both on-line
and off-line measurements are assumed available from the process. Because measurements are not
always correct and reliable, two additional boxes could have been added to denote that some data
reconciliation and gross error detection algorithm should be employed right after the
measurements become available. Since off-line measurements are not as frequent and are
characterized by substantial time delays between the time the measurement is made and the
 r---
PROCESS OBJECTIVE

'If
ON and OFF-LINE ON and OFF-LINE
OPTIMIZATION .J MODEL UPDATING
.---- ~
"'

I\)

H
'"
'"
If
• ,
I

SET MODEL ST...·n! I

BASED lPl~OClESS ON·UN!! I!STIM ",ll!l) ... COMPARISON
POINTS . I!STIM"'TION
~ CONTROL ~ MI!"'SURI!MI!NTS ,. V... RI ... BLES

, JI'
I Ol'l'-UN\! MI!"'SUIU!MI!NTS

FT:P.DDACK OF OH-1.1N1! MI!ASUREMF.NTS

mEDII ...CK OF IlSnM ...·I1!\) V",RIAIIJJ!S
--- --- ---------------------- ------

Figure 1: Comprehensive Schematic Diagram of the TeMOC Methodology

 264

laboratOlY results are available, an important task, denoted by the state estimation box, has been
introduced. This algorithm utilizes the on-line measurements and a dynamic or steady state model
of the process to estimate process variables that are not available on-line. Such variables can and
should include the off-line measurements but are not limited to these only. Depending on how this
algorithm is designed, it can handle the existence of noise in the on-line measurements and can also
cope with a dynamic model of the process that is not perfectly accurate. The existence of a model
is critical for the success of this approach. In the past the models used were accurate first principle
ones and thus required substantial effort for their development. This has limited the use ofthis
technique to applications for which the development of the fundamental model is a straightforward
task. In the future, more efficient modeling methods will need to be developed to facilitate the
application of state estimation techniques. In most batch plant applications, one usually expects
that there will be substantial initial process-model mismatch. This can be quantified by the
difference between the estimated and the actual values of the off-line measurements of the process.
This difference can be used to update the model either off-line or on-line. We will return to this
point in a short while.
On the top part of the diagram, the indicated task represents the definition of the process
objectives that are used in the subsequent task of process optimization. The optimization task is
also achieved with the use of an appropriate process model. The results of the optimization task
are expressed in terms of desired set point or set point profiles with time. The model based
controller ensures that these set points are indeed met during the operation of the process. For this
purpose the model base controller will utilize, in general, both direct on-line measurements as well
as those that are estimated on-line through the state estimation task.
It is quite obvious that there are three important tasks where some model of the process is
utilized: State Estimation, Control and Optimization. Since the model development and updating
activities are time consuming and technical manpower intensive, there is a very substantial
incentive to initially develop and subsequently update the same model for the control, estimation
and optimization purposes.
In the following sections, we will first describe the progress made to date on the
methodological issues of the TeMOC approach. We will also comment on the additional
challenges that need to be met in order to further the comprehensive character of the strategy
described above. After this has been achieved we direct attention to some references where
application examples demonstrate the successful use of the proposed approach.
 265

Modeling

Despite the central role that individual batch unit operations, and in particular batch reactors, play
in the overall perfonnance of an industrial batch process, their description in the fonn of a
mathematical model is either nonexistent or not very accurate. While steady state models for
continuous units are widely used for the design and operation of such processes, the development
and use of models for batch processes is not as widely practiced. Since batch process units are
never in steady state, the necessary models need be dynamic ones quantitatively describing the
time evolution of the unit or process. Dynamic models have presently started to become widely
used in studies of continuous processes. Small efficiency improvements in such processes can
result in economical benefits that more than compensate for the development costs of the process
model. Many other chemical and specialty chemical processes have not widely benefited from
such dynamic models because of the apprehension that the development cost of a specific process
model might outweigh the potential process benefits. This apprehension is also present in batch
processes.
To make models more widely available, one needs to address and resolve two issues. The first
issue relates to the cost of model development: It should be reduced! The second issue relates to
the effective use of the developed model: It should be used for more than one purpose, i.e.,
optimization, estimation of unmeasured variables, as well as control. Reduction of model
development costs can be achieved by development of a systematic modeling approach that we
have called "Tendency Modeling". Past modeling strategies, usually for continuous processes,
implied that the model was developed once and remained valid for a substantial part of its process
life. This was justified because processes did not change their operational characteristics very often
and the amount of available on-line data were not substantial. These assumptions are no longer
valid. Nowadays, the operation of many continuous processes varies week by week and even day
by day. This was and is true with batch processes, but it is quickly becoming true for continuous
processes as well. As feed-stock characteristics and product specifications change, the model
structure and its parameters also need to change. A model must be flexible enough to adapt to
these changes. Meanwhile, the wide use of digital control computers in plants has availed a much
larger number of experimental data that can be used to update and improve the model.
The proposed "Tendency Modeling" approach that has been developed over the past few
years can serve as the primary vehicle for answering the modeling needs for batch and other
 266

processes. It is based on the following main principles:

* i) The initial and subsequent versions of the "Tendency Model" should involve a set of
material and energy balances of the unit. For this purpose utilize as much initial information
as is available on the reaction stoichiometry, chemical kinetics, heat and mass transfer,
thermodynamics, and mixing characteristics of the unit. If the initial information is not
sufficient then design and perfoan additional experiments to collect the necessary information.

* ii) Estimate the accuracy of the model and take this information into account in all possible
uses of the model, as for the design of estimation, optimization, and/or control algorithms.
* iii) As new experimental data become available from the operation of the unit, update the
model parameters as well as its structure by the effective use of all relevant on-line and
off-line process data.
In contrast to previous ones, the proposed "Tendency Modeling" approach is an evolutionary
one and aims to systematize the ad hoc model updating that is widely used in industry today. A
simpler version of such an evolutionary modeling approach has been used in the past in a different
and restrictive manner - adaptive control. However, the types of models used in adaptive control
are linear input/output dynamic models that do not provide any fundamental knowledge about the
internal workings of the process. Because of their restrictive character, these models have not been
used for purposes other than control. They are only linear and cannot be used for optimization;
their input/output nature make them inappropriate for estimation of unmeasured parameters. There
is also some similarity between the proposed "Tendency Modeling" approach and evolutionary
optimization (EO) in that the model is used for the optimization of the process and that it evolves
with time. On the other hand, evolutionary optimization uses input-output statistical and often
static models compared to the dynamic and nonlinear models that are used in the Tendency
Modeling approach. Further more, evolutionary optimization models do not utilize any of our
knowledge of the inner workings of the process. This means that EO utilizes identical statistical
input-output models for the optimization of a batch reactor as for the optimization of a batch
distillation column. Models used in evolutionary optimization cannot be used for the estimation
of unmeasured variables or for the control of the unit.
The proposed "Tendency Modeling" methodology aims to develop nonlinear models based
on material and energy balances around the unit operation of interest, such as a chemical reactor.
The nonlinear character of these models enables their use in optimizing the reactor or process.
Since these models are developed by writing material and energy balances, they provide the
 267

possibility of estimating process variables that are not measured on-line. Periodic updating of
tendency models, either on-line or off-line, will combine and enhance the advantages of their
fundamental character by continuously increasing their accuracy. There are generic research
challenges that the tendency modeling approach must still resolve to become a useful and
successful tool in all process examples. They are summarized in the following open research
questions:
* i) When is parameter updating of the nonlinear tendency model sufficient and how does
one select the parameters that should be updated?
* ii) When is it necessary to update the structure of the model, i.e., change the reaction rate
expression, consider partial mixing, or introduce heat or mass transfer limitations?
While methods for parameter updating of linear models have been considered in the literature,
work needs to be done for updating the structure of the process model using nonlinear models.
Here updating of the model structure implies a discrete change of one model to another in order
to take more detailed account of, for example, imperfect mixing or mass (and/or heat) transfer
limitations.

State Estimation

State estimation techniques for linear models were first introduced by Kalman [22, 23] about thirty
years ago. They were extended to nonlinear systems and found their most extensive use in
aerospace control applications. It is not totally clear why such technologies have not yet found a
more extensive use in chemical process applications. One can argue that the model developmental
cost in the aerospace industry is distributed over a number of identical products (e.g. airplanes),
while chemical plants are usually one of a kind. One can further speculate that the nonlinear
character of chemical processes has been an inhibiting factor. Unlike aerospace applications, the
available model of a chemical process is not always as accurate. Recent research on the control
of emulsion copolymerization [7, 8, 9, 10, 11] has demonstrated that the use of a less-than-perfect
model can lead to successful estimation of unmeasured parameters-with substantial economic
benefit resulting from the simultaneous increase in the product quality and process productivity.
The ever-increasing emphasis nowadays on controlling product quality variables that cannot be
measured on-line further increases the need for successful application of state estimation
techniques.
 268

The following issues need to be addressed to achieve significant progress in the use of state
estimation techniques:
* i) Practicing research engineers need to be more effectively informed about the power and
limitations of the method. Tutorial papers need to be written, simple processes examined as
examples, and demonstration software provided for application engineers to develop an
intuitive rather than mathematical understanding ofthe power of the method.
* ii) The success of existing state estimation techniques depends on the accuracy of the model,
which needs to be examined in a more systematic and generic fashion. Motivated by the
specific application results in the reactor control area, this activity shows promise for a
significant technical contribution. Understanding what model accuracy is needed to make state
estimation successful will further enhance the applicability of this technique to chemical
processes.
* iii) Methods need to be developed that utilize the observed process model mismatch, usually
called the innovation residual, to help in the on-line updating of the process model and further
enhance the usefulness of state estimation techniques. Model updating can mean parameter
updating or updating the form of the model. Parameter adaptive state estimation techniques
can be used in updating the model when we are certain which major parameters are to be
updated. However, new techniques are needed for updating the model structure. Orie such
technique is to perform two state estimations in parallel with two different models in order
to see which one provides a more confident estimation of the unmeasured variables.

* iv) In most ofthe past research activities one has assumed that a dynamic model is needed for
the estimation ofthe unmeasured variables from the ones that are measured on-line. Because
a steady state model is easier to develop than a dynamic model one needs to examine when
this simplification can be made without a substantial sacrifice in the accuracy of the estimate.
Furthermore one needs to explore the development of input-output models between measured
and estimated variables so that the dependence of the estimation task on a fundamental model
of the process is not as critical.

Control

In the design of control strategies for batch processes, one needs to first focus on the selection of
the proper variables to be controlled. As in most other processes, the objective of the controller
 269

is to ensure that the desired quality of the product is achieved in spite of disturbances that might
enter the process. Since the final product quality is affected by the operation of several units, many
of which are downstream from the- one under consideration, it is not easy to define and select the
quality related variables that need to be controlled in each unit. Furthermore, even if these
variables are identified it might not be possible to measure then on line. For example, one knows
very well that the quality of emulsion polymers is often dependent on the molecular weight, which
is impossible to measure on-line. This then necessitates the development of an estimation
algorithm discussed in the previous section. If the expected process and products improvements
are substantial, then the recommended approach is to design and implement an estimation
algorithm and then control the estimated variables. In many cases, the possible process or product
improvements have not been considered and estimated. The easy way out then is to accept that
we can only control the variables that we can directly measure. It is often hoped that this will
indirectly help in the control of the product quality and in some cases it does. However, it is quite
possible that the maximum process benefits might not be achieved through this approach.
Once the selection of the controlled and manipulated variables is made, the remaining task is
to design the controller strategy. The challenges here are that the relationship between manipulated
and controlled variables is almost always nonlinear and often quite multivariable. The controller
design becomes more challenging when more than one variable related to product quality, such
as polymer composition, particle size, and molecular weight, is controlled simultaneously. Often
the challenge becomes even greater if temperature runaway, due to the exothermicity of the
reaction, can lead to an explosion. In this case, the possibly more urgent task of temperature
control must be coordinated with the economically more important product quality control
strategy. This is not an easy task in many applications, such as bulk polymerization or certain
organic synthesis reactions, because of the substantial interactions between temperature and
compositions. Many of the model based control strategies that have been developed for the control
of continuous processes could be utilized in the control of batch processes. Their use of the
available model of the process in the prediction of the future differences between the measured
variables and their appropriate set points could be an effective way to design the controller
algorithm. The major limitation that can be cited is that most, but not all, model predictive control
strategies utilize a linear model of the process. Because batch processes are nonlinear in their
dynamics, substantial room exists for the application of nonlinear model predictive control
strategies. One can mention the use of the Reference System Control (RSC) strategy that has been
 270

proposed by Bartusiak et al. [3,4], and further examined by Bartee et al. [2]. This is an almost
identical strategy to the one proposed by Lee and Sullivan [24] and often referred to as Generic
Model Control.

Optimization

Batch process optimization is the final and most important task that needs to be undertaken in
order to improve performance of the process. Such optimization tasks can be divided into two
broad and overlapping categories. The first one deals with the optimization of the operation of the
process unit. The second category of optimization challenges relates to the optimal scheduling of
the different unit tasks to perform the overall process objective. While the second class of
problems is very important as well, we will focus attention here onto the first class. The issues
examined here with respect to the optimization of the operation of each batch units, are also of
relevance to the optimal scheduling of the overall process.
Improvements in the process will either reduce operating costs, equivalently increase the
productivity, or increase the quality of the product produced, or both. A substantial number of
mathematical optimization techniques are available in the literature [5, 18, 19,33, 12,6] and can
be readily utilized if an accurate model is available. One needs to also mention the substantial
progress made recently by the application of Sequential Quadratic Programming on the operation
of batch processes. In this case only a tendency model will be available, and process optimization
will be performed concurrently with efforts to increase the accuracy of the model. Our interests
then should be to develop algorithms that will ensure simultaneous convergence of model updating
and process optimization tasks. We need to develop algorithms which determine the global, rather
than local, process optimum and result in the largest process improvement. To achieve this, one
needs to develop a strategy that decides whether the next batch run should be used to either
improve the process or increase the model's accuracy. As Rippin, Rose, and Schifferli [32] have
demonstrated, the optimization of the process through an approximate model can be trapped into
a local optimum. These authors have also proposed an extended design procedure which continues
model parameter improvement with performance verification in the neighborhood of the predicted
optimum. While achieving the global rather than the local optimum is an important issue, one
should keep in mind that by guiding three or four process units to some local optimum might be
more beneficial rather than guiding one process unit to its global optimum. Nevertheless, the
 271

impact ofthe accuracy of the model on the process optimum needs to be studied further and we
might consider denoting such a challenge as "Tendency Optimization".

Example Process Applications

To properly elucidate the arguments provided above, one needs to also refer to some specific
example applications of the proposed Tendency Modeling approach. We will provide here some
comments about the application to organic synthesis reactions done as part of the doctoral thesis
by Rastogi [28, 30, 31]. The reaction considered was the epoxidation of oleic acid to epoxide and
the evolution of the Tendency Model and the optimization ofthe process are summarized in Figure
2. Here the value ofan economic performance index ($/day/liter processed) is plotted against the
number of experimental runs. The initial eight runs were designed by a factorial design of
experiments procedure [28] to provide the initial data needed to start the Tendency Modeling
approach. All experiments indicate that operation of the process was not economical resulting in
a loss rather than a profit. It is worth also mentioning that all experiments were operated in a batch
mode, all reactants were fed in the reactor at initial time. Utilizing these data, the first version of
the tendency model (MO) of three reactions and power-law kinetics was identified. Because a
negative reaction order was calculated on the oleic acid reaction, it was decided to feed the oleic
acid in semibatch mode [30]. The optimization of the next batch through this model predicted a
possible profit of73.4 $/day/liter but the corresponding experiment only achieved a profit of6.9
$/dayJliter. One needs to remark that inaccuracies of the MO model are to blame for the substantial
difference between the predicted and the experimentally achieved value of the profit function. At
the same time one can easily observe that this tendency model, however inaccurate, guided the
process to profit making operation (Run 9) as compared to the previous eight runs. With this
additional data and a closer look at the process model mis-match, the structure of the kinetic
equations was changed [29, 30] and Model Ml was obtained. Optimization of the next batch run
through this model led to the prediction that a profit of 74.2 $/day/liter can be achieved.
Experimental run 10 implemented the optimal profile of model MI and resulted in a profit of 53.2
$/dayJliter. With these additional experimental data, Tendency Model M2 was obtained by refitting

the parameter of model MI. With experimental Run 11, it was shown that both model predictions
and experimental data converged very close to each other successfully ending the cycle of
tendency model updating and process optimization.
 272

100 M1 M2
MO
Q-_·-o-·-.
50 /
......
t
Break Even Point
---------- --/- i
--------- --- -----_ .
"O'i:;'
~~ -50
4)'-
u~
c: «I
"'''0
E~ -100
0'-'
'e4)
Po. • EXDerilllelllal
-lSO
o Model

-200
a 2 4 6 8 10 12

Experiment Number

Figure 2. Summary of the Evolution of the Performance Index with Experimental Nwnber

An additional and successful comparison of the overall approach to experimental data of a

process of industrial interest was also presented by Marchal [27]. The applicability of the
Tendency Modeling approach to the case ofbioreactors has been addressed by Tsobanakis et al.
[35]. Rastogi [29] also reports of a very successful application to an industrial process of Air
Products and Chemicals, Inc. Productivity was increased by 20.
One might end by making the comment that while additional industrial applications are
expected to be undertaken and completed in the future, the challenge of further extending the
methodology to more quantitatively account for the accuracy of the model is a real and worthy
one.

References

I. E. Anderson. Specialty chemicals in a mixed bag of growth. Chem. Eng. News, 20,1984
2. J.F. Bartee, KF. Bloss, and C. Georgakis. Design of nonlinear reference system control structures. Paper presented
at AIChE National Meeting, San Francisco, 1989
3. RD. Bartusiak, MJ. Reilly, and C. Georgakis. Designing nonlinear control structures by reference system synthesis.
Proceedings of the 1988 American Control Conference, Atlanta, Georgia, June 1988,
4. R.D. Bartusiak, M.1. Reilly, and C. Georgakis. Nonlinear feedforwardlfeedback control structures designed by
reference system synthesis. Chem. Eng. Sci., 25,1989
5. Denbigh. Optimal temperature sequence in chemical reactors. Chem. Eng. Sci., 8: 125-132, 1958
6. M.M. Denn. Optimization by Variational Methods. Robert E. Krieger Publishing Company, 1978
7. 1. Dimitratos, M El-Aasser, A. Klein, and C. Georgakis. Composition control and kalman filtering in emulsion
 273

copolymerization. Proceedings at the 1988 American Control Conference, Atlanta, Georgia, 1988
8. 1. Dimitratos, M. EJ-Aasser, A. Klein, and C. Georgakis. Control of product composition in emulsion
copolymerization. Proceedings, 3rd International Workshop, Polym. React. Eng., Berlin, Sep. 27-29, 1989
9. 1. Dimitratos, M. EI-Aasser, A. Klein, and C. Georgakis. Digital monitoring, estimation and control of emulsion
copolymerization. Proceedings of the 1989 American Control Conference, Pittsburgh, PA, 1989
10. 1. Dimitratos, M. EI-Aasser, A. Klein, and C. Georgakis. Dynamic modeling and state estimation for an emulsion
copolymerization reactor. Compo Chern. Engng., 13 :21-33, 1989
11. 1. Dimitratos, M. EI-Aasser, A. Klein, and C. Georgakis. An experimental study of adaptive kalman filtering in
emulsion copolymerization. Chern. Eng. Sci., 46:3203-3218, 1991
12. T.F. Edgar and D.M. Himmelblau. Optimization of Chemical Processes. McGraw-Hill, 1988
13. C. Filippi, 1. Bordet, 1. Villermaux, S. Marchal-Brassely, and C. Georgakis. Batch reactor optimization by use of
tendency models. Compo and Chern. Engng., 13:35-47, 1989
14. C. Filippi, 1.L. Graffe, 1. Bordet, 1. Villermaux, 1.L. Bamey, P. Bonte, and C. Georgakis. Tendency modeling of
semibatch reactors for optimization and control. Chern. Eng. Sci., 41 :913, 1986
15. C. Filippi-Boissy. PhD thesis, L' Institute National Polytechnique de Loraine, Nancy, France, 1987
16. V. Grassi. Communication at pmc's industrial advisory committee meeting, October. 1992
17. J. W. Hamer. Stoichiometric interpretation of multireaction data; application to fed-batch fermentation data. Chern.
Eng. Sci., 44:2363-2374, 1989
18. 1. Horak, F. Jiracek, and L. Jezova. Adaptive temperature control in chemical reactors. a simplified method
maximizing productivity of a batch reactor. Czech. Chern. Comm., pages 251-261, 1982
19. H Hom Feasibility study of the application of self-tuning controllers to chemical batch reactors. Oxford Univ. Lab
Report, 1978
20. M.R. Juba and J. W. Hamer. Process and chaIJenges in batch process control. Paper presented at the third
International Conference on Process Control, 1986
21. A. Jutan and A. Uppal. Combined feedforward-feedback servo control scheme for an exothermic batch reactor.
Proc. Des. Dev., 23:597-602, 1984
22 RE Kalman. A new approach to linear filtering and prediction problems. 1. basic Eng, March:35-46, 1960
23. RE. Kalman and RS. Bucy. New results in linear filtering and prediction theory. 1. Basic Eng., March:95-J08,
1961
24. P.L. Lee and GR Sullivan. Generic model control- theory and applications. Paper presented at IF AC Workshop,
June 1988, Atlanta, 1988
25. S. Marchal-Brassely. PhD thesis, L' Institute National Polytechnique de Loraine, Nancy, France, 1990
26. S. Marchal-Brassely, 1. Villermaux, 1.L. Houzelot, 1.L Bamay, and C. Georgakis. Une methode interative efficace
d'optimisation des profils de temperature et de debit d'alimentation pour la conduite optimale des reacteurs
discontinus. Proc.of 2eme Congres Frcncais De Genie Des Procedes, Toulouse, France, pages 441-446, 1989
27. S. Marchal-Brassely, 1. Villermaux, J.L. Houzelot, and J.L. Bamey. Optimal operation of a semi-batch reactor by
self-adaptive models for temperature and feed-rate profiles. Chern. Eng. Sci., 47:2445-2450, 1992
28. A. Rastogi. Evolutionary optimization of batch processes wing tendency models. PhD thesis, Lehigh U., 1991
29. A. Rastogi. Personal communications, September. 1992
30. A. Rastogi, 1. Fotopoulos, C. Georgakis, and H.G. Stenger. The identification of kinetic expressions and the
evolutionary optimization of specialty chemical batch reactors using tendency models. Paper presented at 12th
In!. Symposium for Chemical Reaction Engineering, Torino, Italy. Also in Chern. Eng. Sci., 47:2487-2492.,1992
31. A. Rastogi, A. Vega, C. Georgakis, and H. G. Stenger. Optimization of catalyzed epoxidation of unsaturated fatty
acids using tendency models. Chern. Eng. Sci., 45:2067-2074,1990
32. D.W.T. Rippin, LM. Rose, and C. Schifferli. Non-linear experimental design with approximate models in reactor
studies for process development. Chern. Eng. Sci., 35:356-363, 1980
33. CD. Siebenthal and R Aris. Studies in optimization vii. the application ofPontryagin's method to the control of
a batch and tubular reactor. Chern. Eng. Sci., 19:747-746, 1964
34. P. Tsobanakis, S. Lee, J. Phillips, and C. Georgakis. Adaptive stoichiometric modeling and state estimation of batch
and fed-batch fermentation processes. Presented at the 1989 Annual AIChE Meeting, San Francisco, 1989
35. P. Tsobanakis, S. Lee, J. Phillips, and C. Georgakis. Issues in the optimization, estimation, and control of fed-batch
bioreactors using tendency models. 5th In!. Con. on Computer Applications in Fermentation Tech. and 2nd
IFAC Sym. on Modeling and Control of Biotechnology Processes, Keystone, Colorado, March 29 - April 2, 1992
Control Strategies for a Combined Batch Reactor/Batch
Distillation Process

Eva S0renson and Sigurd Skogestad

Department of Chemical Engineering, University ofTrondheim-NTH, N7034 Trondheim

Abstract: A batch reactor may be combined directly with a distillation column by distilling
off the light component product in order to increase the reactor temperature or to improve
the product yield of an equilibrium reaction. The controllability of such a system is found to
depend strongly on the operating conditions, such as reactor temperature and composition of
distillate, and on the time during the run. In general, controlling the reactor temperature (one
point bottom control), is difficult since the set point has to be specified below a maximum value
in order to avoid break-through of heavy component in the distillate. This maximum value
may be difficult to know a priori. For the example considered in this study control of both
reactor temperature and distillate composition (two-point control) is found to be difficult. As
with one point bottom control, the reactor temperature has to be specified below a maximum
value. However, energy can be saved since the vapor How, and thereby the heat input to the
reactor, can be decreased with time. Controlling the temperature on a tray in the column (one
point column control) is found to give the best performance for the given process with no loss of
reactant and a high reactor temperature although no direct control of the reactor temperature
is obtained.

Keywords: Reactive batch distillation, controllability, control strategies

1 Introduction
Batch distillation is used in the chemical industry for the production of small amounts of
products with high added value and for processes where flexibility is needed, for example,
when there are large variations in the feed composition or when production demand is
varying. Batch reactors are combined with distillation columns to increase the reaction
temperature and to improve the product yield of equilibrium reactions in the reactor by
distilling off one or more of the products, thereby driving the equilibrium towards the
products.
Most often the control objective when considering batch processes is either i) to min-
imize the batch time or ii) to maximize the product quality or yield. Most of the papers
published on batch distillation focus on finding optimal reflux ratio policies. However,
sometimes the control objective is simply to obtain the same conditions in each batch.
This was the case for the specific industrial application which was the starting point for
our interest in this problem and which is to be presented later.
 275

Few authors have considered the operation of batch distillation with chemical reaction
although these processes are inherently difficult to control. The analysis of such systems
in terms of controllability has so far only been considered by S0rensen and Skogestad [l1J.
Roat et al. [8J have developed a methodology for designing control schemes for contin-
uous reactive distillation columns based on interaction measures together with rigorous
dynamic simulation. However, no details about their model were given.
Modelling and simulation of reactive batch distillation has been investigated by Cuille
and Reklaitis [2J, Reuter et al. [7J and Albet et al. [1]. Cuille and Reklaitis [2] developed a
model and solution strategies for the simulation of a staged batch distillation column with
chemical reaction in the liquid phase. Reuter et al. [7] incorporated the simulation of PI-
controllers in their model of a batch column with reaction only in the reboiler. They stated
that their model could be used for the investigation of control structure with the aid of
Relatjve Gain Array analysis (RGA) but no details were given. Albet et al. [1] presented
a method for the development of operational policies based on simulation strategies for
multi component batch distillation applied to reactive and non-reactive systems.
Egly et al. [3], [4] considered optimization and operation of a batch distillation column
accompanied by chemical reaction in the reboiler. Egly et al. [3] presented a method for
the optimization of batch distillation based upon models which included the non-ideal
behavior of multi component mixtures and the kinetics of chemical reactions. The column
operation was optimized by using the reflux ratio as a control variable. Feeding one of the
reactants during the reaction was also considered. In a later paper [4], they also considered
control of the column based upon temperature measurements from different parts of the
column. The optimal reflux ratio policy was achieved by adjusting the distillate flow
using a non-linear control system. However, no details where given about neither the
column/reactor nor the control system.

The purpose of this paper is to investigate the possible difficulties in controlling a

coupled system of a reactor and a distillation column, and also to give some alternative
control strategies based on an industrial example. First, a model of the industrial process,
consisting of a batch reactor with a rectifying column on top, is developed. Based on a
linearized version of this model, we compare different operating points to show how the
model differs, that is, whether the same controller settings can be used for different reactor
conditions or reactor temperatures. In the various operating points we also consider the
stability of the system and the response to step changes in flows. We consider two-point
control, when both the top and the bottom part are controlled, as well as one point control,
when only one part of the column/reactor is controlled. A Relative Gain Array (RGA)
analysis is used for the investigation of control structures in two point contro!' Finally, the
similarities and differences between our process and a conventional continuous distillation
column is considered.
The reaction is reported to be of O'th order and due to limited data, we also assume
the rate to be independent of temperature. However, interesting observations can still
be made concerning the coupling between the formation of product in the reboiler and
the separation in the column above. Indeed, later work [6], has confirmed that this
simplification does not affect the conclusions. The influence of disturbances on the system,
e.g. in reaction rate or in temperature measurements, has not been considered in this
study.
 276

Column: 6 trays + condenser

Reaction: 0.5 Rl + 0.36 R2 + 0.14 R3 -> P(s) + W
Volatile components: W (Tb = 100°C) and R2 (n = 188°C)
Non-volatile components: Rl (Tb = 767°C), R3 (Tb = 243°C) and P (solid)
Vapor pressure: R1 : InPR, = -4009.3 + 176750.0/Ti + 6300.0 log Ti
- 0.51168Ti (Pa)
R 2 : In PRo = 25.4254 - 6091.95/( -22.46 + Ti) (Pa)
R3: In PR3 = 231.86 - 18015.0/T; - 31.753 log Ti
+ 0.025Ti (Pa)
w
W: In P = 23.1966 - 3816.44/( -46.13 + T;) (Pa)
Relative volatility (Wand R2): 8-32
Startup time: 30 min
Total reaction time: 15 hr
Pressure in column/reactor: 1 atm/1.2 atm
Reaction rate, r: 1.25 kmol/hr
Initial vapor flow, V: 16.8 kmol/hr
Hydraulic time constant, r: 0.0018 hr =6.5 s
Initial holdups: reactor: 24 kmol
condenser: 1.6 kmol
trays: 0.09 kmol
Initial amounts in reactor: R 1 : lOA kmol (Acid)
R 2 : 7.5 kmol (Alcohol) (20 % excess)
R3: 3.2 kmol (Alcohol)
P: 0.0 kmol (Ester)
W: 2.5 kmol (Water)

Table 1: Process data for simulation.

2 Process example
The motivation for this study was an industrial equilibrium esterification reaction of the
type
e1R1 + e2R2 + 6R3;=: P(s) + W
where Rl is a dibasic aromatic acid, R2 and R3 are glycols, P is the solid polymer product
and W is the bi-product water. The reaction takes place in a reactor heated by a heating
jacket with heat oil. The equilibrium is pushed towards the product side by distilling off
the low boiling by-product W from the reactor. Only reactant R2 and the by-product W
are assumed to be volatile, and the binary separation between these two components takes
place in the column. The reaction rate was reported to be of zero order; independent of
compositions. Due to lack of data we also assume the rate is independent of temperature.
A summary of the process data is given in Table 1. In the industrial unit the amount
of reactant R2 in the feed was 20 % higher than necessary to yield complete conversion
of the reaction, and this was also assumed in most of our simulations. This was done to
account for the possible loss of the reactant in the distillate.
The existing operating practice was to use one-point top control; the temperature at
the top of the column TT was kept constant at about 103°C which gave a distillate
composition of 0.004 (about 2 weight%) of the heavy component R2 and thereby a loss
of this component. The vapor flow was kept constant by using maximum heating of the
 277

240
Reactor TB
220 'r-f
200
I"_......._ ..........._ _ ' ... ..._ ......................_ ... r ....._
2,3,4
.................... ,..... ...... .......... .

Q
~ ~_ ~

180
5
160

140 - ______________________________ 9_______________ _

120

lOOO~···~·==~2~···~·I~.QPc4~·II~·~·.. ··~····~··6~····~·····~····~····~···8~~·~····~·1~~==~12~..~····~·····714~····~····-··-716

Time, [hrJ

Figure 1: The existing temperature profile in column/reactor.

reactor and the condenser level was controlled by the distillate flow D. The temperature
profile at different locations in the column as a function of time is given in Fig. 1. The
reactor temperature TB is almost constant at the beginning but increases as reaction
proceeded. The conditions on tray 2, 3 and 4 are practically equal because the column
has more stages than needed for the desired separation. With the existing control scheme
there is no direct control of the reactor temperature, TB and more severely, it gives a
varying loss of the heavy component, reactant R 2 • This leads to a varying quality of the
product P between batches.

3 Mathematical model
In this section we consider the mathematical description of the batch distillation column
and reactor shown in Fig. 2 and described in the previous section. The equations for the
individual stages consist of the total mass balance, the mass balance for each component,
tray hydraulics and phase equilibrium and are valid under the following assumptions:

Al A staged model is used for the distillation column.

A2 A multi component mixture in the reactor, but a binary mixture in the distillation
column is considered.

A3 Perfect mixing and equilibrium between vapor and liquid on all stages is assumed.

A4 The vapor phase holdup is negligible compared to the liquid phase holdup.

AS The stage pressures and the plate efficiencies are constant.

 278

Tr
Distillate
t----t'KI---....L.--f:>I<::1---- 0, Yo
7- Reflux
L

ReactorlReboiler
M, x,

Figure 2: Batch distillation column/reactor.

A6 Constant molar flows are assumed (no energy balance).

A 7 Linear tray hydraulics is considered.

AS Total condensation with no subcooling in the condenser is assumed.

A9 The chemical reaction is limited to the reactor.

AID Raoult's law for the vapor-liquid equilibrium holds.

Below i denotes the stage number and j the component number (j = 1,2 are the volatile
components Wand R2)' The following differential and algebraic equations result.

reactor/reboiler, i=l :
4
dMi/dt = L2 - V + L~jr (1)
j=l

d(M1Xl,j)/dt = L2X2,j - VYl,j + ~jr ,j =1 (2)

reaction components not distilled :

(3)
 279

column tray, i=2,N :

dM;/ dt = LiH - Li (4)
d(MiXi,i)/dt = Li+1Xi+lJ + VYi-1J - LiXi,i - VYiJ ,j = 1 (5)
condenser, i=N+l :
dMN+ddt =V - LN+l - D (6)
d(MNHXN+l,i)/dt = VYN,i - LN+lYD,i - DYD,i ,j = 1 (7)
linearized tray hydraulics:
. - L. Mi - Moi
L,- 0'+---- T
(8)
liquid-vapor equilibrium :
QiXi,i
Yi i
,
= --:--'-..:.!::-,---
1 + ((ti - 1) Xi,;
(9)

relative volatility :
Pt (T;)
(ti = f (Ti ) = Pi (T;) (10)

temperatures:
2(4)
Pi =L x;P] (T;) (ll)
;=1
On each stage the composition of component j = 2 (R 2 ) is obtained from LXi = 1. Note
that all four components were used to calculate the reactor temperature using eq. ll, but
only the two lightest components were considered in the column. The model is highly
non-linear in vapor composition Y and in temperature T. On vector form the differential
equation system to be solved can be written

dx/dt = f[x(t), u(t)] (12)

In addition there is a set of algebraic equations, equations (8)-(11)

0= g[x(t), u(t)] (13)

Eq. (12-13) constitute a set of differential-algebraic equations (DAE). The equations are
solved using the equation solver LSODE [5]. The startup conditions are total reflux and
no reaction.

3.1 Linear model

In order to investigate the controllability of a process using available tools, a linear model
is needed. Based on the non-linear model described by eq. (12) and (13) a linear model
can be developed by linearizing the equation system at· a given operating point. For
continuous processes normally only one operating point considered; that of the steady
state conditions. The linear model is then found by linearizing around this operating
point and will be valid for small deviations from the steady state. When considering
batch processes there is no such steady state; the conditions in the reactor or column are
changing with time and the model is linearized along a trajectory. A linearized model of
 280

Controlled variables (y): Manipulated variables (u):

condenser holdup MD distillate flow D
distillate composition YD reflux flow L
reactor temperature TB vapor flow V

Table 2: Controlled and manipulated variables.

the process, representing deviations from the "natural drift" along the trajectory with D,
L and V constant, can be described by the following equations:

dx/dt Ax+Bu
Y Cx (14)

Where

X [6..Xj, 6..Mj .. , 6..n3,.Y

y [6..MD' 6..YD, 6..TBf
u [6..D,6..L,6..Vf

Laplace transformation yields:

y(s) = G(s)u(s) (15)
The control problem will thus have the controlled and manipulated variables as given
in Table 2. It is then assumed that the vapor flow V can be controlled directly, even
though the real manipulated variable is the heat input to the reactor.

4 Analysis of linear model

4.1 Operating procedures
The linear model depends on the operating point. To study these variations we initially
considered four different operating procedures:

I The existing operating practice, TT = 103 0 C (one point top control, V constant)

II TB = 200 0 C (one-point bottom control, V constant)

III TB = 222 0 C (one-point bottom control, V constant)

IV TB = 228 0 C (one-point bottom control, V constant)

Temperature profiles for the four operating procedures are given in Fig. 3. For operating
procedure I, II and III the conditions are more or less constant with time, whereas pro-
cedure IV has a changing temperature profile with large variations at the beginning of
the batch but more or less stabilizing midway through the batch. For operating point I
(TT = 103 °C), the front between light and heavy component is kept high in the column
giving a loss of the heavy component R 2 • For procedure II (TB = 200 °C), the front is
low and the composition of heavy component R2 is almost negligible from tray 3 and up,
giving a very pure distillate. However, the reactor temperature is low and it is unlikely
 281

250~----~----~I:~o~)D~,e~r~m~in~12~~Dlro~c~e~d~u~re~TT~-~10~3~CT-____~~~~

~
__----IRTR----------~I-----------------
200 234
, _______ ._ _ _ _ _ _ _--=z.: ..:.1 _ _ _ _ _ _ _ _ _ _ _ _ _

5
150
------------------------~----------------------
TT 7
100 ................................................................................................................................................... .
o 2 4 6 8 10 12 14 16
Time, [hr]
250~----~----~rr~:o~)D~'e~r~a¥ti~n=2~p:r~oc~e~d~u~r~e~TB~=-~2~0~0~CT-____~____~

200tl~____~T~B~_________I~_______________

150
: 2
"~-----------------------------------------------
1000~···~··--~2-----4~--~6~--~8----~1~0----~12~--~1~4--~16
Time, [hr]

250r-----~----~IIT~:~o~oler~a~tl~·n~2D~lr~o~c£ed~u~r~e~TFB~-~2~2~2~C¥_----~----~
TB
200
•
r' - - - - - - - - - - - - - - 2
- - - - - - - - _____________ _
150 f
I
------_

3
100 ... .1~.......................................................................................................................................................... .
o 2 4 6 8 10 12 14 16
Time, [hr]
250r-----~----~~~~~~r~o~c~ed~u~r~e~TFB~=~2~2~8~C¥-----~----~
1

200
2
-----
---
150
•.•.•.•.•...................~........................... ---
o 4
....~
100~···~·~~~--~~~~~-··-···=···~··~··~··~··=·--~~-···-···- ...~
....~... ~~~~~.----~
6 8 10 12
=
. . . .. .. .14.. 16
Time, [hr]

Figure 3: Temperature profiles in column/reactor for different operating procedures.

 282

that the assumed reaction rate will be achieved. When the reactor temperature is in-
creased to TB = 222 DC for procedure III the composition of R2 in the column section
increases pushing the light/heavy component front upwards in the column. At the end
of the batch the front detracts slightly giving more light component in the bottom part.
For procedure IV, at TB = 228 DC, the front between light and heavy component is lifted
so high up in the column that it leads to a "break-through n of heavy component R2 in
the distillate and thereby causing the large variations in the profile. After the loss of R2
the light/heavy component front detracts continuously during the batch.
Of the four operating procedures, procedure III (TB = 222 DC) is the only one with
both a high reactor temperature and at the same time no loss of reactant R 2 • Procedure
IV (TB = 228 DC) gives a substantial loss of reactant R2 and is therefore not considered
further.

4.2 Linear open-loop model

To illustrate how the process behavior changes during the batch the equation system (eq.
12 and 13) is linearized at different operating points; that is at different reactor conditions
or times during a batch. (Notation: An operating point is specified as procedure-time, ego
1-8 is the conditions with operating procedure I after 8 hr reaction time.) These linear
models were found by first running a non-linear simulation of the process with control
loops implemented (level control in the condenser and temperature control of tray 1 or
of the reboiler) in order to obtain a given profile in the column/reactor. The simulations
were then stopped at the specified time, all the controller loops opened and the model
linearized numerically. (We would get the same responses, in DoYDH and DoTB from steps
in DoL and Do V if the condenser level loop was to remain closed with the distillate flow
D during the linearization. This is because L and V have a direct effect on compositions
and the effect of the level loop is a second order effect which vanishes in the linear model.)
The resulting linear model is thus an open loop description of the process at the given
time and conditions; it describes how the system responds to changes when no controllers
(or only the level controller) are implemented.

4.3 Step responses

To illustrate how the process behavior changes with conditions in the reactor we consider
step changes to the linearized models. The effect of a step in the vapor flow V on YDH
and TB (deviation from nominal value) for three different operating procedures after 8
hr is given in Figure 4. The variation of the linear model within batch III is illustrated
by Fig. 5. The responses in YDH for different reactor conditions (top part of Fig. 4) are
similar but differ in magnitude. This is because in operating point II-8, where we have a
very low reactor temperature, we have a very pure distillate. The increase in reflux will
only increase the purity marginally. Whereas in operating point 1-8, we have a distillate
which is less pure so the increase will be larger. We note from Fig. 5 that the variations
within the batch are large for the response in reactor temperature. The main reason for
the changes in the responses for this temperature (lower part of Fig. 4 and 5) is the
concentration of water in the reactor. That is, a higher water concentration gives a larger
effect.
 283

1-8

111-8

- - - -- - - - - - - - - - - - - - - - - - - - -
11-8
/'
/
10·15 L -_ _ _ _-'--_ _ _ _- ' -_ _ _ _-'--_ _ _ _- ' -_ _ _ _---'
o 0.1 0.2 0.3 0.4 0.5
Time, [hr)

2r-----.----~----~------r--------,

g, 1.5
.... ----
cO
f-
.s
a:;
1

-:.-
-.. -
,
11-8
.....
- .. -
... -.-
111-8
CO.5 1-8

Oo~~-·~~~~:·~·~==~================J
0.1 0.2 0.3 0.4 0.5
Time, [hr)

Figure 4: Step in vapor flow (6oV = 0.1) for linear mode): effect on 60YDH and 60TB for
procedure 1-8 (TT = 103°C), II-8 (TB = 200°C) and III-8 (TB = 222°C).

10.7 .------,------,------,-------,..------.

111-2_. __ -,-::':"C'- _._._ '-C_:_·-O. - . - - - . - .

10.11 '--"-'_ _ _ _-'--_ _ _ _- ' -_ _ _ _-'--_ _ _ _- ' -_ _ _ _---'

o 0.1 0.2 0.3 0.4 0.5
Time, [hr)

6r--------.--------~--------~---------r--------~

111-12
111"8 .. ··

00~~~~~~~~~~~~~::::··~-~-·:··~·:··:·~:::-:-~"~~~2::::~
0.1 0.2 0.3 0.4 0.5
Time, [hr)

Figure 5: Step in vapor flow (6oV = 0.1) for linear model: effect on 60YDH and 60TB for
procedure III-2 to III-15 (TB = 222°C).
 284

o.8r---------r---------r---------r-------~r_------~

0.7

:t: 0.6

~0.5
>-
S
8 0 .4
II 111-2
:I:
~0.3
~
<II
Cl 0.2

0.1

°0L---~--~~--------0-.L2--------~0~.3---------0~.~4--------~0.5
Time, [hrJ

Figure 6: Logarithmic transformation for linear model: Different times during batch for
procedure III (TB = 222°C).

4.4 Reducing the non-linearity for top composition

An interesting feature in Fig. 4 and 5 is that the responses in YDH to step changes have
a similar initial shape on a log-scale. This is actually a general property for distillation
[9]. The inherent nonlinearities in this variable can therefore be reduced by using a log
transformation on the distillate composition YD:

YD = -In(l-YD) (16)
which in deviation variables becomes

L\YDH = L\YDH (17)

YIJH
The responses in mole fraction of heavy component R2 in the distillate after the transfor-
mation, YDH, is given in Fig. 6 for operating points 1II-2 to III-15. These responses are of
the same order of magnitude and the non-linearity is thereby reduced. From Fig. 4 and
5 there is no obvious transformation that can be suggested to deal with the non-linear
effect for the reactor temperature. .

5 Control strategies
The varying loss of reactant R2 in the distillate and the lack of direct control of the reactor
temperature were the major problems with the existing operating practice. In the control
part of this study the following control strategies are compared:
 285

• one-point bottom control (controlling the reactor temperature directly)

• two-point control (controlling both the distillate composition and the reactor tem-
perature)

• one-point column control (controlling the temperature on a tray in the column)

The control parameters for the PI-controllers used in the simulations are given in Table
3. Note' that an integral time of Tj = O.lhr = 6min was used in all the simulations and
that the transformed variable YD was used instead of YD for two-point control.

level control: Kp = -500 and = 0.1 (MD -- D, L)

Tj

bottom control: Kp = 1.0 and = 0.1 (TB -- L)

Tj

two-point control: Kp = 0.456 and = 0.1 (YD -+ L)

Tj

Kp = -4.0 and Tj = 0.1 (TB -+ V)

column control: Kp = 1.0 and = 0.1 (Ts
rj -+ L)

Table 3: Control parameters used in the simulations.

5.1 One-point bottom control

The objective with one-point bottom control is to keep the reactor temperature constant
at the highest possible temperature as this will maximize the rate of reaction for a tem-
perature dependent reaction. The reflux flow is used as manipulated variable and the
vapor flow is kept at its maximum value (V = Vmax = 16.8 kmol/hr). However, it is very
difficult to achieve a high value of TB and at the same time avoid "break-through" of
the heavy component R 2 , in the distillate. This is illustrated in Fig. 7 which shows how
the mole fraction of R 2 , YDH, changes when the set point for the temperature controller
in the reactor increases from TB,set = 224.5 0 C to TB,.et = 225°C. An increase of 0.5 0 C
causes the mole fraction of reactant R2 to increase by a factor of 25. The loss of reac-
tant is only temporary, and YDH is reduced to ::::; 0 after about 1 hr. The break-through
is caused by the fact that when the specified temperature is above a certain maximum
value where most of the light component W is removed, then a further increase is only
possible by removing the heavy component, reactant R 2 . If the set point temperature is
specified below the maximum value, in this case::::; 224.0oC, good control of the system
(TB ::::; TB, ..t and YDH ::::; 0) is achieved. The system can, however, become unstable at
the end of the batch depending on the choice of control parameters in the PI-controller.
This due to the non-linearity in the model causing the system to respond differently to
changes at different times during the batch as illustrated in Fig. 5.
Another alternative for raising the reaction temperature, and thereby the reaction rate
for a temperature dependent reaction, is to let the set point follow a given trajectory, e.g.
a linear increase with time. Again, the maximum reactor temperature to avoid break-
through will limit the possible increase and break-through is inevitable if it is specified
too high. Fig. 8 illustrates a run when the set point follows a linear trajectory from 220°C
at t = 0.5 hr to 245°C at t = 15 hr. The loss of reactant R2 is substantial, almost 10 %
of the feed of this component. By lowering the endpoint temperature to 230°C, loss of
reactant is avoided (not shown).
 286

5.2 Two-point control

By using two-point control it may be possible to control both the top and the bottom
part of the distillation column by implementing two single control loops in the system. In
this way energy consumption can be reduced since it will no longer be necessary to keep
the vapor flow V, and thereby the temperature or amount of heating oil, at its maximum
value. In the case of the esterification process, it is desirable to control not only the reactor
tempera.ture TB but also the composition of the distillate YD, i.e. the loss of reactant R2 •
Two different control configurations are considered for the batch column:"

LV-configuration Controlling the condenser level using the distillate flow D leaving the
reflux flow L and the vapor flow V to control the distillate composition YD and the
reactor temperature TB:

MD +--+ D
YD, TB +--+ L, V

DV -configuration Controlling the condenser level using the reflux flow L leaving the
distillate flow D and the vapor flow V to control the distillate composition YD and
the reactor temperature TB :

MD +--+ L
YD,TB +--+ D, V

5.2.1 Controllability analysis of two point model

Open-loop step responses for both configurations are given in Fig. 9 and 10 for operating
point III-8 (TB = 222 DC at t = 8hr). The term "open-loop" should here be put in quotes
because we are not talking about an uncontrolled column, but assume that the condenser
level is perfectly controlled (MD +--+ D or MD +--+ L) and we consider the effect of the
remaining independent variables on the composition and reactor temperature.
From Fig. 9 it can be seen that for the LV-configuration the responses to steps in L
and V are similar but in opposite direction. For the DV-configuration the responses by a
step in D are similar as for the step in V for the LV-configuration. However, the responses
to a step in V is very small. This is a general property for distillation.
In a distillation column there are large interactions between the top and the bottom
part of the column, a change in the conditions in one end will lead to a change in the
other end as well. Because of these interactions a distillation column can be difficult or
almost impossible to control. The interactions in a system can be analyzed by various
tools (see e.g. Wolff [12]), amongst them the RGA, or Relative Gain Array. Systems with
no interactions will have an RGA-value of 1. The larger the deviation from 1, the larger
the interaction and the more difficult the process is to control. Pairing control loops on
steady-state RGA-values less than 0 should be avoided.
The magnitude ofthe I,l-element of the RGA for both the LV- and DV-configuration
is given as a function of frequency in Fig. 11 for operating procedure III-8 (TB = 222D C).
From the figure it can be seen that for the LV-configuration the RGA is very high at
low frequencies (when the system is approaching a steady state). This shows that the
interaction reduce the effect of the control input (L, V) and make control more difficult.
 287

-8 0_1 step in L -8 0_1 step in V

OX10 2.5 x 10

-0.5 2
:r: :r:
Cl -1 ~1.5
>-
~
~ -1.5 Q)
Cl
Cl
-2 0.5

-2.50 00
0.2 0.4 0.2 0.4
Time, [hr) Time, [hr)
0.1 step in L 0.1 step in V
0 2

u-0 .5 U 1.5
m
I-
m
I-
-1
.!!l .!!l
a; a;
Cl -1.5 Cl 0.5

-2 00
0 0.2 0.4 0.2 0.4
Time, [hr) Time, [hr)

Figure 9: Linear open-loop step responses for LV-configuration for operating point III-S.

-8 0.1 step in D -10 0.1 step in V

2.5x 10 2.5x 10

2 2
:r: :r:
~1.5 ~1.5
.!!l
a;
Cl
0.5

00
0.2 0.4 0.2 0.4
Time, [hr) Time, [hr)
0.1 step in D 0.1 step in V
2 0.04 (

U 1.5 Q:0.03
m-
l-
.m
I- 0.02
.!!l .!!l
a; a;
Cl 0.5 Cl 0_01

00 00
0_2 0.4 0.2 0.4
Time, [hr) Time, [hr)

Figure 10: Linear open-loop step responses for DV-configuration for operating point III-S.
(Note that the y-axis scaling is 100 times smaller for changes in V).
 288

- -____________ ~LV

10-2 10°
Frequency (radians/hr)
100r-------------.--------------r------------~r_----,

DV ~-----------

LV
-200~------------~r_-----------L~----~----~~----~
10-4 10-2 10° 102
Frequency (radians/hr)

Figure 11: RGA for LV- and DV-configuration for linear model in operating point I1I-S.

RGA for DV is generally lower at all frequencies. This difference between configurations
is the same as one would observe in a continuous distillation column.
However, the control characteristics from the RGA-plot for the LV-configuration are
not quite as bad as it may seem. For control the steady-state values are generally of
little interest (particularly in a batch process since the process will never reach such a
state), and the region of interest is around the system's closed-loop bandwidth (response
to changes), which is in the frequency range around 10 rad/hr (response time about 6
min). We note that the RGA is closer to 1 here and that the difference between the two
configurations is much less. From the high-frequency RGA, which is close to 1, we find
that for decentralized control, the loop pairing should always be to use the vapor flow V
to control the reactor temperature TB and either the reflux flow L or the distillate flow D
to control the distillate composition or the loss of reactant R 2 , YD. This is in agreement
with physical intuition.

TB <---+ V
YD <---+ L, D
MD <---+ L, D

5.2.2 Non-linear simulation of two-point model

Closed-loop simulations confirm that two-point control may be used if fast feedback con-
trol is possible. However, as in the case for one-point bottom control, we still have the
problem of specifying a reasonable set-point for the bottom temperature to avoid break-
through of reactant R2 in the distillate. An example of two-point control of the process
 289

using the LV-configuration is given in Figure 12 with the following set point for the con-
trollers: TB ••• t = 225 DC and YDH ••• t = 0.0038. (Note that we control the transformed
distillate composition YD instead of YD in order to reduce the non-linearity in the model.)
It can be seen that only a minor break-through of reactant occurs during the run. The
reactor temperature TB is kept at its set point while the distillate composition YDH is
slightly lower than its set point showing that it is difficult to achieve tight control of both
ends of the column at the same time. It should also be noticed how the vapor flow de-
creases with time which shows that energy can be saved using two-point control. Control
using DV-configuration give similar results (not shown).

5.3 One-point column control

In the existing operating practice the temperature at the top of the column was controlled.
The set point was 103 DC which gave a composition of 0.4 % of reactant R2 in the distillate.
By lowering the set point to e.g 100.1 DC the distillate would be purer, but the column
would become very sensitive to measurement noise, and this system would not work in
practice.
One alternative is to measure the composition YD and use this for feedback. However,
implementing an analyzer (or possibly an estimator based on the temperature profile) is
costly and often unreliable. A simpler alternative is to place the temperature measurement
further down in the column, e.g. a few of trays below the top tray, since this measurement
will be less ~ensitive to noise. In this investigation the temperature on tray 5 is chosen
as the new measurement to be used instead of the one on the top tray. The vapor flow
is kept fixed at its maximum value (V = Vmar = 16.8 kmoljhr). With this control
configuration (Ts ..... L) there is no direct control of the reactor temperature. However,
with an appropriate choice of set point, Ts ...t, loss of reactant R2 could easily be avoided
and one of the main causes of the operability problems thereby eliminated.
The temperature profile for one-point column control with set point Ts.••t = 130 DC
is shown in Fig. 13. The conditions are "stable" (i.e. no break-through of reactant R 2 )
throughout the batch. The reactor temperature increases towards the end and the mole
fraction of heavy component in the distillate YDH is less than or equal to 0.0001 at all
times. Also note that this control procedure with V fixed at its maximum will yield the
highest possible reactor temperature. This may be important in some cases when the
reaction is slow.

6 Reducing amount of reactant

The proposed operating procedure with one-point column control gives a lower reactor
temperature than the existing one-point top control procedure with TT = 103 DC. In
the existing procedure the amount of reactant R2 in the feed is about 20 % higher than
needed for the reaction and all the above simulations were based on this. This is done to
account for the loss of the reactant during the run. By using one-point column control
with Ts = 130 DC, loss of reactant can be avoided and the surplus of R2 is therefore
not needed. By removing the excess 20 % of the reactant from the feed (such that the
initial charge of R2 is 6.25 kmol) the obtainable reactor temperature increases by about
2DC at the beginning of the batch to about 40 DC towards the end as illustrated in Fig.
 290

250
p

-
Reactor TB 1
200 1::>::\4
IV 5
150 ~
IV r-' '"
TT 7

2 4 6 '8 10 12 14 16
Tlh,e, [hr]

0.025
0.02
:r: 0.015
o
>- 0.01
0.005
v
o 2 4 6 8 10 12 14 16
Time, [hr]

20~~----------------~----------~------~----------~

~
015
~
;;:10
o
;:

8. 5
~
2 4 6 8 10 12 14 16
Time, [hr]

20'r-----__-----,------~----_r----~------~----_r------

2 4 6 8 10 12 14 16
Time, [hr]

Figure 12: Two-point control. Temperature profile, distillate composition, vapor flow and
reflux flow for LV-configuration with set points T B ,6et = 225°C and YDH,set = 0.0038.
 291

240r----r----~--~----~--~----~--~----,

ReactorTB
220iJ

200

T5 5
120
6
1000~···~·~-~-~--~i~-~-~-~--~4~-~-~-~--~-~6~-~-~--~-~8~-~-~--~-~-~10~-~-~--~-~1~2~-~-~--~-~147-~-~-~16

Time, [hrl
Figure 13: One-point column control. Temperature profile with TB .••t = 130 °G.

290
TT=103 C
.280 .. T5=130 C
270 ._ T5=130 C (- 20% R2)
u
~ 260
~....
Q)
0..
e 250
B
....
B 240
u
'"
Q)
~
230

220)

2100~--~2----~4----~6~---8~---1~0~--~12~--~1~4--~16

Time, [hrl
Figure 14: Effect of reducing the amount of reactant R2 in the feed.
 292

14. The reason for this is the high vapor pressure of the component R2 which lowers the
temperature as given by Eq. 11. Since the temperature is considerably higher towards
the end of the batch when the excess R2 is removed, the total batch time can therefore
be reduced for a temperature dependent reaction.
In conclusion, by moving the location of the temperature lower down in the column,
we

1. Increase the reactor temperature and thus reduce the batch time

2. Avoid loss of reactant R2

3. Maintain more constant reactor conditions.

7 Comparison with conventional distillation

columns
A comparison of our column with a conventional batch distillation column, shows sig-
nificant differences in terms of control. For example, the common "open-loop" policy of
keeping a fixed product rate (D) or reflux ratio (L/D) does not work for our column
because of the chemical reaction (see also [6]). If the distillate flow D is larger than the
amount of light component W formed by the reaction, the difference must be provided
for by loss of the intermediate boiling reactant R 2 . For optimal performance we want
to remove exactly the amount of bi-product W formed. Therefore feedback from the top
is needed. In fact, our column is very similar to a conventional continuous distillation
column, but with the feed replaced by a reaction and with no stripping section.
By comparing our reaction batch column with a conventional continuous column we
find that most conclusions from conventional columns carryover. As for a continuous
column RGA(l,l) ~ 0 at steady state (low frequency) for the DV-configuration for a
pure top product column (see Fig. 11) implying that the reflux flow should be used to
control the reactor temperature [10]. However, for control the pairing must be selected
based on the RGA(l,l)-values around the bandwidth (10 rad/hr) implying that the vapor
flow should always be used to control the reactor temperature for two-point control as
was done in the simulations.

8 Conclusion
In this paper a dynamic model of a combined batch reactor/distillation process has been
developed. Based on a linearized version of the model the controllability of the process
depending on different reactor conditions and different times during a batch has been
analyzed. The responses of the industrial example has been found to change considerably
with operating point.
Controlling the reactor temperature directly using one-point bottom control, will give
a more consistent product quality. However, since the response changes with time (gain
between TB and V), a non-linear controller might be needed to avoid instability. Moreover,
because of the moving light/heavy component front in the column it is difficult to find
the right set point temperature that does not give a break-through of heavy component
 293

in the distillate. This set point temperature will therefore in practice have to be specified
low enough to ensure an acceptable performance.
Two-point control allows-both the reactor temperature and the distillate composition
to be controlled. By using two-point control energy will be saved compared with one-point
control as the vapor flow can be reduced. However, one encounters the same problems of
specifying the set point for the reactor temperature as for one-point bottom control.
The existing operating practice, controlling the temperature at the top of the column,
is poor, sensitive to noise and leads to a varying loss of reactant R2 and thereby varying
product quality. The measuring point should therefore be moved from the top tray and
further down in the column. The proposed new procedure of one-point column control,
where the temperature on tray 5 is controlled, has several advantages:

• No loss of reactant R2 (compared to controlling the top temperature)

• Need not worry about maximum attainable reactor temperature (compared to con-
trolling the reactor temperature directly by one-point bottom control)

• No interactions with other control loops (compared to two point control)

With this new operating policy addition of excess reactant R2 to the initial batch can
be avoided. Thus, the batch temperature can be increased and the batch time thereby
reduced.

NOTATION
A system matrix Ti temperature, K
B system matrix Tb boiling point, C
C system matrix TB reactor temperature, K
D distillate flow, kmol / hr TT temperature at top of column, K
G(s) transfer function u control vector
L reflux flow, kmol/hr V vapor flow, kmol / hr
Li internal liquid flow, kmol/hr x state vector
LOi initial liquid flow, kmol/ hr Xi,j molfraction of light compo (W) in liquid
Mi liquid holdup, kmol YD molfraction of light compo -(W) in distillate
MB liquid holdup in reactor, kmol YDH molfraction of heavy comp.(R 2 ) in distillate
MD liquid holdup in cond., kmol = 1- YD
MOi initial liquid holdup, kmol YD logarithmic molfraction of light compo (W)
Pi pressure on tray i, Pa in distillate = -In(1 - YD)
P'1 vapor pressure, Pa Yi,j m.)lfraction of light compo (W) in vapor
r reaction rate, kmol/ hr y measurement vector
Greek letters scripts
Qi relative volatility tray number
6. deviation from operating point j component number
T hydraulic time constant, h- 1 set set point
~j stoichiometric coefficient nominal value
 294

References:

I. Albet, J., J.M. Le Lann. X Joulia and B. Koehret: "Rigorous Simulation of Multicomponent Multisequence Batch
Reactive Distillation", Proc. COPE'9I, Barcelona. Spain, 75-80 (1991).
2. Cuille, P.E. and G.V. Reklaitis: "Dynamic Simulation of Multicomponent Batch Rectification with Chemical
Reactions", Compo Chern. Engng., 10(4), 389-398 (1986).
3. Egly, H., V. Ruby and B. Seid: "Optimum design and operation of batch rectification accompanied by chemical
reaction", Compo Chern. Engng., 3,169-174 (1979).
4. Egly, H., V. Ruby and B. Seid: "Optimization and Control of Batch Rectification Accompanied by Chemical
Reaction", Ger. Chern. Eng., 6, 220-227 (1983).
5. Hindmarsh, A.C.: "LSODE and LSOm, two new initial value ordinary differential equation solvers", SIGNUM
Newsletter, 15(4), 10-11 (1980).
6. Leversund, E.S. S. Macchietto, G. Stuart and S. Skogestad: "Optimal control and on-line operation of reactive batch
distillation", Compo Chern. Eng., Vol. 18, Suppl., S391-395 (1994) (supplernent from ESCAPE'3, Graz, July
1993).
7. Reuter, E., G. Womy and L. Jeromin: "Modeling of Multicomponent Batch Distillation Processes with Chemical
Reaction and their Control Systems", Proc. CHEMDATA'88, Gothenburg, 322-329 (1988).
8. Roat, S.D., J.J. Downs, E.F. Vogel and lE. Doss: "The integration of rigorous dynamic modeling and control
systern synthesis for distillation columns: An industrial Approach", Presented at CPC'3 (1986).
9. Skogestad, S. and M. Morari: "Understanding the dynamic behavior of distillation columnsn, Ind. & Eng. Chern.
Research, 27 (I 0), 1848-1862 (1988).
10. Shinskey, F.G: "Distillation Control", 2. ed., McGraw-Hill Inc., 1984.
II. Sorensen, E. and S. Skogestad: "Controllability analysis of a combined batch reactor/distillation process", AIChE
1991 Annual Meeting, Los Angeles, paper 140e (1991).
12. Wolff, E.A., S. Skogestad, M. Hovd and K.W. Mathisen: "A Procedure for controllability analysis", presented at
the IFAC Workshop on interactions between process design and control, Imperial College, London, Sept. 6-8,
(1992)
A Perspective on Estimation and Prediction for Batch Reactors

Mukul Agarwal

TCL, Eidgenossische Technische Hochschule, Zurich, 8092 CH

Abstract: Estimation of states and prediction of outputs for poorly known processes in general,
and batch reactors in particular, have conventionally been approached using empiricism and
experience, with apparently inadequate regard for the underlying reasons and structure. In this
work, a consistent perspective is presented that clarifies some important issues, explains the causes
behind some of the intuition-based tactics, and offers concrete guidelines for a logical approach to
the estimation problem.

Keywords: Estimation, Identification, Prediction, Model-mismatch, Extended State

1 Introduction

The science of estimation using observers and filters originated from the needs of electrical and
aeronautical applications [4,7]. In subsequent chemical-engineering applications, this science
steadily metamorphosed into an art of estimation using such techniques as the Extended Kalman
Filter and the Luenberger Observer. The more the chemical process differed in crucial aspects
from the processes that originated the science, the more the engineer became an artist using whim,
fancy, creativity, and trial-and-error to achieve a pleasing, or at least an acceptable, final result.
As in any art, success came to be validated by the final result itself, regardless of how many
unsatisfactory versions were discarded in the process or what kind of creative license was used to
deviate from the established norm. Indeed, the creative twists and deviations took on a validity
oftheir own and sneaked a place in the norm. Little surprise then that the art of estimation has its
lore, which is inundated with effect-oriented prescriptions and claims that do not always have roots
in any cause [1-3,5-8]. Some of these prescriptions and claims have evolved from empirical
observation on numerous real and simulated applications; others have been borrowed directly from
the science in spite of invalidity in process applications of the theoretical assumptions that they are
based upon. For example:
The distinction between parameters and states lies in their -dynamics.
The innovation is an indicator of filter performance. Whether estimating parameters or states,
strive to get the output residual to be a white, or at least a zero-mean, sequence
The covariance of the state errors is an indicator of the accuracy of the obtained estimates.
Inaccurately known parameters can simply be coestimated by inclusion in the extended state
vector.
Given the process and the model, the best tuning is fixed and only needs to be divined.
Filters require a stochastic model, observers a deterministic model.
 296

The tuning of a state-space filter depends solely, or mainly, on the noises in the input and
output signals.
Estimators are tuned a priori without needing any measured data, except to deduce the noise
level.
When the covariance of the estimated states indicates filter divergence, increase the noise in
the dynamic model equations.
Ifthe estimator does not work, try to somehow model the uncertainty.
Batch processes are characterized by strongly nonlinear and time-varying behavior. Batch
reactors, in particular, are moreover difficult to model and do not permit easy on-line measurement
of crucial properties such as concentration. Together these characteristics render estimation and
prediction for batch reactors specially intractable, and use of much of the lore specially treacherous.
This presentation attempts to regard the popular art of estimation from a scientific perspective,
in the hope of reinstating some science to guide the development of useful estimators and
predictors, and to discourage the propagation of artistic ones.

2 Process and Model

A simple semi-batch-reactor model assuming a single second-order exothermic reaction and a

cooling jacket that allows removal and measurement of the generated heat of reaction may take
the form:
dc
(1)
dt

q = (-M!)kc 2 (2)

where t is time, c unknown concentration of the limiting reagent, k known kinetic rate constant,
F measured input flow rate ofthe limiting reagent, (-~H) known heat of reaction, and q measured
rate of generated heat of reaction. This model corresponds to the general state-space form:

dx
dt = f(x, p, u) (3)

y = hex, p, u) (4)

where c is the conventional, possibly extended, state that comprises all the estimated variables with
both zero and non-zero dynamics, p is the conventional parameter that is known a priori, u the
measured process input, y the measured process output, and f and h known general nonlinear
functions.
 297

The inevitable model-mismatch causes deviation between the above model and the true
process. Regardless ofthe nature of the mismatch, the true process can be described as:

dx'
-;tt= fi(' • )
X,p,u + ex (5)

y' = h(x', p', u) + e'

x (6)

where x", p", and y" are the true state, the true parameter, and the true output, respectively; ex"
and fy" are the errors in the dynamic and measurement equations, respectively; and the input u and
the functions f and h are identical to those used in the model.

3 Errors ex* and ey*

The errors ex" and ey ", appearing in the process equations 5 and 6 describe the entire deviation
due to model-mismatch, and are time-dependent, in general. The mismatch could be due to
structural incompleteness in the model equations 3 and 4, due to discrepancy between the model
parameters p and the true parameters p", due to inaccuracy of the initial-state value available to
the model, due to noise in the measurements of u and y, due to unmeasured disturbances, and due
to possible approximations such as linearization that might be involved in any application of the
model equations. All these sources of mismatch are incorporated in the instantaneous deviation
represented by ex" and ey " which are unknown by definition.
At any given time, the estimation ofx using the model equations requires a prescription of the
relative instantaneous values of the unknown errors {ex". ey "}. The prescribed values, {ex' ey },
are not necessarily sought to lie closest possible to the true {ex". ')- "}. The end-effect of this
prescription l is essentially to single out, at any given time, one particular estimate from an infinite
number of possible, equally valid estimates that comprise the space bounded by all cases of extreme
prescriptions (i.e., each element of {ex' ey } being zero or very large).
Since the algorithms proposed in the literature for estimation cannot directly use prescribed
values of {fx, fy}, this prescription is invariably made in an indirect manner, and has taken a variety
offorms. The most popular or well-known form is utilized by the Kalman Filter algorithm, which
specifies {ex' ey } to be zero-mean white noises with possibly time-varying covariances.
Prescription of {ex' ey} is then realized through prescription of the covariances. In theory, these
covariances could be, and indeed for poorly modeled batch processes must be, different at each
different time. But, in practice, there is rarely enough prior reason or knowledge to prescribe each
covariance as more time-variant than one or a few step-wise constant values over the entire run.

lWhat is commonly called "tuning" of an estimator is consistently referred to in this work as "prescription of
{ex'I!y}" in order to emphasize its origin and to distinguish it from the true errors {ex *, ey *} in the dynamic and
measurement equations.
 298

In the Kalman Filter algorithm, the prescribed covariances affect the resultant instantaneous
estimate via a series of steps involving an array of intermediate matrices and vectors, each of which
has a qualitatively foreseeable effect on the resultant outcome. Prescription of {ex' ry} could
therefore, instead of being made through prescription ofthe covariances, be delegated just as easily
to prescription of any of the intermediate matrices and vectors [5]. Numerous ad-hoc methods
have been devised to do just that, e.g., covariance resetting, high-gain filters, or gain bounding.
Another popular device effects the {ex' ey } prescription by restricting the time duration over
which the model equations are deemed to be valid [1,2]. By limiting the data memory using
exponential forgetting or a moving window, these methods collapse the multidimensional ex space
to a single, albeit in general time-varying, dimension. The attendant simplification in prescribing
{ex' ry} comes at the cost of a restricted space of possible outcomes attainable by using the
reduced-dimension prescription. An even more indirect form of prescribing {ex' ey} is commonly
used by observers, through the convergence speed of the estimates.
Regardless of which indirect form of prescribing {~ey} is used, it serves, in essence, to "tune"
the attained instantaneous estimate to any location within the space comprised by all possible
prescriptions. Not surprising then is abundance in the literature of excellent estimates or
predictions obtained both in simulations and in experiments, where the employed prescription is
either not reported or simply stated without justification. In these cases, the excellent results could
readily be achieved by accordingly tuning the prescription off-line on the application run or on past
runs. In other cases, the process is so well known that either the dynamic model in equation 3, or
the measurement model in equation 4, or both, are nearly perfect so that ex", or ey", or both, are
negligibly small, and excellent results are expected with ex « ry, or ty «l'x, or all possible
prescriptions, respectively [7].
In yet other cases, the prescription is made truly on-line and a-priori using, instead of a result-
oriented tuning, some criterion that is not always justified. For example, {ex' ry} has been
prescribed to simply match the known covariances of the noises in the input and the measured
output, disregarding the possibly large contributions to {ex", ty "} from model-mismatch,
parameter discrepancy, and approximations [3,5]. Or the elements of {ex' ey} have been prescribed
corresponding to the nominal values of the respective outputs and state changes, which amounts
to setting without justification as equal all elements of the {ex' ry} corresponding to a properly
scaled model. The ordinary least-squares estimator tacitly neglects ex". In the limited-memory
algorithms, the one-dimensional prescription offorgetting factor is limited to a few favorite values,
with the lone justification of them being "typical" or "experience-based". The observers go even
a step further by completely ignoring the accuracy of the estimates or predictions while prescribing
{l'x, ey}, and basing the prescription solely on the convergence dynamics desired by the end-use,
e.g., control [8].

4 Source of Prescription

The confusion, and the silence, prevalent in the literature about the means and justification for the
prescription of {ex' ry} points perhaps to a reluctance to meet the problem head-on. A candid
perspective in this respect therefore seems in order.
 299

Since {ex". ey"} is unknown by definition, and there is no way to a priori divine it, prescription
of {~ey} has to be made a posteriori based on data from past runs or past data from the current
ran. Clearly, the {ex' ey}-prescription should be congruent with the goal of the exercise, at least
a posteriori for these past data. The proper source of {ex' 'Y }-prescription therefore derives
directly from the goal of estimation itself. Given measured data up to the current time, there are
two main, in general mutually irreconcilable, goals of estimation:
1. Predict the process output at a certain horizon in the future, or
2. Estimate the real process state at the current time,
with best possible accuracy. That these goals are mutually irreconcilable2 is evident from the fact
that, so long as fx", or ey" is not negligible, the model equations 3 and 4 and the process equations
5 and 6 process the same state value differently to give different future output values. When the
second goal is met perfectly, the current estimate ofthe state equals x"(t). Setting x(t) = x"(t) and
processing the equations 3 and 4 results in y(t+) " y"(t+) at any future time t+, unless both ex "and
ey" are negligible. The output then cannot be predicted accurately, and the first goal cannot be met
simultaneously. By similar reasoning, state estimate obtained to fulfil the first goal will have to
differ from the true current state x"(t), and consequently will not meet the second goal.
The complementary cases of estimating the current output or predicting the future state are
of secondary interest. Estimating the current process output would merely mean filtering out the
noise in its available current measurement. This is not a goal of the exercise, specially for poorly
known batch processes where contribution to ey", ofthe measurement noise is overwhelmed by
the other contributions, such as due to model-mismatch. 3 In this situation, the measured values
can be refined relatively little, and only through model-independent signal-processing techniques 4
The other complementary case, predicting the future process state, is of interest only once the
prerequisite second goal can be, and has been, met. Issues relating to prediction of the future state,
given satisfactory estimate of the current state under the second goal, is deferred to a later section.
The first goal, prediction of the future process output, is chosen whenever the output is the
sole property of interest and the state itself is of no direct import. For example, predictive control
of output or output-based supervision constitutes exactly this situation. The other goal of
estimating the current real state holds whenever the state itself is paramount for optimization of
a run or for fault diagnosis, or is to be controlled directly using, say, optimal control or PID
control. The process output is, in this case, oflittle interest, except to help deduce the state.
It is also possible to have subsets or combinations of the two goals. The first goal could refer

2The concept of reconcilability is important throughout the later discussion. A goal is considered
irreconcilable if it comprises sub-goals that cannot be met simultaneously for the given system due to mutual
contradiction.

3In case the other contributions are negligible compared to the measurement noise, e.g., when an output simply
equals a state, then estimate of the current output is readily obtained from estimate of the current state. In case the other
contributions are comparable to the measurement noise, then the output is also included as an extension to the state.
Both these cases fall under the second goal.

4Indeed, in the following discussion it will be assumed that for batch processes the noises in all measurements
of process inputs and outputs are negligible compared to other errors in {ex*, ey *}, so that for all practical purposes the
measured values (or values filtered through signal-processing techniques) can oe considered to be the true values.
 300

to only some of the process outputs, or the second goal to only some of the process states. A
combined goal could simultaneously include some of the outputs and some ofthe states.

5 States and Parameters

Despite some controversy in the literature, perhaps the most widely believed distinction between
states and parameters is based on dynamics. In this view, a boundary is drawn at some low,
arbitrary level of dynamics. Variables with faster dynamics are called states and those with slower
dynamics are deemed parameters. The boundary does not lie at zero dynamics due to slow drifts
that parameters often exhibit [5,7]. This distinction is useful for control purposes, and may even
be useful for observer-based estimation where the emphasis is on the response speeds instead of
the accuracy of estimation. In the general field of estimation, this distinction still holds for known
parameters and states. Extrapolation of the dynamics-based distinction to include the unknown,
to-be-estimated variables, as is common in the literature, is -misleading.
For estimated variables, the essential distinction between a state and a parameter stems from
the two goals of estimation discussed above. Variables that are estimated so as to serve the goal
of predicting the process output accurately cannot strive to maintain identity with a real, physical
variable, and are therefore to be thought of as "parameters" that give a best-fit output prediction.
This is in analogy to the well-known subset of our first goal, namely identification, where all the
estimated variables have zero dynamics. In identification, the emphasis is on the output; the
identified parameters can turn out to be whatever they might please, and have no true physical
value. Similarly for estimation with the first goal, regardless of whether a particular element of a
in equation 3 has significant or zero dynamics, all the estimated x variables have to sacrifice identity
with any real, physical variable in order to serve the output prediction. All x variables, in this case,
are essentially best-fit "parameters" that have no physical meaning and merely give best output
predictions.
In the complementary case of the second goal, all x variables in equation 3, regardless of
whether they have significant or zero dynamics, not only have real physical meaning but must be
measurable at least once, as discussed above. All variables therefore qualifY as "states" that have
physical meaning and can be compared with the true physical value. Thus, even a zero-dynamics
variable such as the-heat of reaction, if included in the second goal, must be independently
measurable and qualifies as a state. Of course, if some of the x variables, typically the ones with
zero dynamics, are exempt from the second goal, and might serve some reconcilable outputs in the
first goal, then these variables are regarded as parameters.
Since the conventional dynamics-based terminology of states and parameters had to be used
in the previous sections, it is consistently maintained in the remaining sections in order to avoid
ambiguity.

6 Prediction of Future State

Predictive control of the state variables or, as is customary in batch reactors, optimization of a
 301

criterion dependent on the future process states, requires accurate values of the state during or at
the end of a time horizon. In such cases, prediction of the process state at a certain horizon in the
future is an important goal. The trivial solution of simply integrating equation 3 starting with the
current-state estimate would work only when ex" is negligible.
For the more realistic case of significantly large x, a promising option is to use this goal itself
to deduce the {~ey}-prescription, analogous to the case of the first goal in the previous sections.
The prescription is consequently chosen so as to minimize the cumulative deviation between the
prediction obtained by processing each state estimate through equation 3 and the corresponding
measured value ofthe state available in the past data. The current-state estimate itself, in this case,
would have to be some bestfit value different from the true current value, and the estimated
variable would be regarded as a parameter in our terminology. A disadvantage of this option is
that it promises reliable prediction only at the end of the horizon and not at other sampling instants
within the horizon. If prediction is needed successively during the horizon, then an independent
estimator would have to be used for each time at which a prediction is needed.
An ad-hoc alternative is to first estimate the current state using the second goal as in the
previous sections, and then integrate equation 3 starting with this estimate while correcting the
integrated value at each future sampling time using some correction measure obtained from the
recent estimation results. In its simplest form, this correction measure would simply equal the
difference between the current-state estimate and its prediction based on the previous-state
estimate and on equation 3. In addition to the difference at the current time, some more most
recent differences could be used to refine the correction measure as well as to indicate its reliability.
Although ad-hoc, this alternative has the advantage of delivering all predictions during the horizon.

7 Goal Reconciliation

The first goal mentioned above can, alone or in partial combination with the second goal, lead to
an irreconcilable objective. In general, the combined goal is irreconcilable whenever the number
of states and outputs included in the goal exceeds the total number of states that are estimated.
Thus, an irreconcilable goal can be rendered reconcilable by increasing the number of estimated
states. This involves extending the state vector through inclusion of existing or new variables
appearing in the measurement equations. To illustrate this point, consider the reactor model in
equations I and 2 with fixed values of the parameters k and (-,1H). For this model, either the first
goal alone or the second goal alone is reconcilable, but both together are not. If the goal is to
predict the future q, then the current real e cannot be estimated, and vice versa. Two variables q
and e cannot together be sought, while only one state e gets estimated. One of the goals must be
abandoned, or another state must be estimated.
One way to add a state is to include the output q directly as a state, so that the model becomes:

de = -kc 2 +F (7)
dt

(8)
 302

(9)

where the measured output has been renamed as qm for distinction. In this extended model, the
combined goal of predicting the future q. and estimating the current real e is reconcilable as two
variables e and q get estimated. In keeping with our terminology introduced in the previous
section, e is here a state, and q is a parameter that merely serves to enable accurate prediction of
qm·
Another way to get the additional degree of freedom is to extend the state to include a process
parameter appearing in the measurement equation 2. Letting k be free in equations 1 and 2 leads
to the model:

de = -kc 2 +F (10)
dt

-dk = 0 (11)
dt

q = (-!ili)kc 2 (12)

Again the extended model allows the combined goal to be reconcilable. In this case, it just happens
that our terminology for state and parameter coincides -with the conventional one, since we are
not interested in the estimated value of k so long as it leads to the best possible prediction of q.
Yet another way to add a state is to create an artificial variable, such as a bias b added to the
measurement model, to give the model:

de
-kc 2 +F (13)
dt

db =0 b(O) =0 (14)
dt '

q = (-!ili)kc 2 + b (15)

This renders the combined goal reconcilable. Notice though that reconcilability would not result,
if the bias b were added to the dynamic model instead, giving the model:
de
(16)
dt
 303

db =0 b(O) =0 (17)
dt '

q = (-MI)kc 2 (18)

In this case, the added state b does not affect the output q independently of the effect of the desired
state c on the output. It therefore provides no additional degree offreedom for q and does not lead
to goal reconciliation.

8 Deduction of the Prescription

Once the goal and/or the model has been modified so that the goal is reconcilable, the prescription
of {fx, ey} can be deduced from past measurements. The strategy for prescribing {ex' ey} for each
of the above two goals is conceptually similar. The best possible setting of the indirect form of {ex'
ey}-prescription (e.g., covariances, or forgetting factor) must be sought by trial-and-error (or
equivalently using a higher-level optimization). For any assumed setting of the prescription, first
the algorithm is used over the available past data to deliver corresponding state estimates.
Then, for case of the first goal, the state estimate corresponding to each data point is translated
using the model equations 3 and 4 into a prediction ofthe process output as far into the future as
dictated by the goal. Each prediction is then compared with its corresponding measured value
from the data set, 5 and the cumulative deviation is regarded as a measure of badness of the
assumed setting of the prescription. The setting that minimizes, or reduces to a satisfactory level,
this cumulative deviation is then regarded as the {ex' ty}-prescription to be used in the next
implementation of the algorithm.
In case of the second goal, availability of some state measurements in the past data is a
prerequisite for a meaningful {ex' ty}-prescription. In other words, if the past data does not
include any measurement of the state, then {ex' ty} cannot be prescribed for the purpose of
estimating the real process state. That is, there is no way to estimate the real process state in an
application run, if the state cannot be measured, at least once, in a "tuning" run. Given such
measurement, the trial-and-error procedure follows similarly as in the other case. Those state
estimates for which a corresponding state measurement is available in the data set are compared
with that measured value, and the cumulative deviation is again regarded as a measure of the
badness of the assumed setting of the {ex' ey}-prescription.
In the case ofa partial or a combined goal, the cumulative deviation minimized by the trial-and-
error procedure stems from only those outputs and/or states that are included in the goal.

5Except for some predictions at the end for which no corresponding measured value is available.
 304

Whenever the goal includes more than one variable (output or state), the user must assign
relative weights to these variables for determining their contribution to the cumulative deviation
that dictates the {~ ey}-prescription. This requirement is obvious when the different variables are
mutually irreconcilable, but holds even when the goals are reconcilable. In the latter case, only a
{ex' ey}-prescription that changes with each sampling time could fit all the past data perfectly,
obviating the need for userspecification of relative weights. The perfect fit, however, is
undesirable, as discussed in the next section. An imperfect fit inevitably entails left-over error that
must be distributed among the different goal variables by design.
The design decision that the user must make about the desired relative accuracy of each of the
goal variables is independent of the true or suspected nature of the corresponding components of
{~il. eyil}. Still, having to specify weighting factors is undesirable in practice, because it forces an
arbitrary compromise in meeting the different goals. The sensitivity of the obtained estimates and
the achieved goals with respect to these weights depends on the left-over error. For irreconcilable
goals, unavoidably large leftover errors can render the weights so sensitive that the user effectively
ends up having to practically specify one particular outcome from a wide range of possible, equally
valid estimates. For reconcilable goals, the weight specification can be made relatively insensitive
by parameterizing the prescription so as to avoid excessive left-over error due to underfit, as
described in the next section.

9 Parameterization of the Prescription

The {ex' 'Y }-prescription deduced by trial-and-error can be used successfully in the further
application of the algorithm only if the subsequent process data has similar characteristics as the
past data that was used for the trial-and-error prescription. If not, then still this is the best that can
be done, and the goal cannot be met any better. In case the past data includes the current run, the
prescription could periodically be updated on line, and the most recent setting used for the next
implementation ofthe algorithm. Such on-line adaption of the prescription is specially useful when
prolonged unmeasured disturbances constitute a significant contribution to {ex il. eyil}.
In theory, the {ex' ey}-prescription can be different at each sampling time,6 but that would
make it overfit the past data and render it unreliable for application to new data. In practice, the
other extreme is preferred where each element of {ex' ey} is restricted to a single constant value
for the entire past data, with the associated inevitable underfit. The superior middle ground of
allowing each element to take a few step-wise constant 7 values is forgone for practical reasons.
There is no way to decide how many constant values ought to be allowed, how each constant value
ought to be ordered within a run, and how the values obtained for one run are to be applied to
another run that might not have a similar course of state changes with time.
Being forced to use a single constant value for each element of {ex' ey} involves considerable

6For the second goal of true-state estimation, the prescription can be different only at the points of successive
state measurements.

7Instead of a step-wise constant curve, a linear or higher-order curve is obviously possible, but rarely
justified as an a-priori choice.
 305

loss of freedom in fulfilling the goal. Due to the underfit caused by underparamctcrization of the
prescription, the goal might get fulfilled quite accurately in some parts of the run, but relatively
larger inaccuracies may remain in other parts of the run. A reduction in the degree ofunderfit is
possible through two kinds of modification of the model. The first kind involves making the model
inherently richer in the sense of enhancing the knowledge it embodies. This could be done, for
example, by including an additional measured variable or by modeling more accurately a parameter
that was previously set to a constant value. Process considerations and modeling limitations,
however, often rule out this option. The other kind of modification, which can be realized more
readily, entails endowing the model with further degrees of freedom for the {ex' ey}-prescription.
This reduces the degree of under-parameterization of the prescription by increasing the number of
elements in {ex' Cy}, while each element is still restricted to a single constant value.
The extra degree of freedom, or the higher-order parameterization, for {ex' ey}-prescription
is achieved by extending the model to include certain estimated variables with fixed initial values.
The sole purpose ofthese estimated variables is to increase the number of elements in {ex' ey}. No
true value is sought for them; in fact, they might not even have any physical meaning. These
extension states are strictly parameters in our terminology from a previous section. Theoretical
I

observability with respect to these states is implied by restricting their initial values to be fixed and
constant for all runs, past and future.
One way to modifY the extended model of equations 10, 11, and 12 for this purpose would be
to let another physical parameter (-tUf) be free, to give the model:

de = -kc 2 +F (19)
dt

dk
- = 0 (20)
dt
dI(-II.1-1\
....... , = 0, (-t:.H)(0) = (-t:.H)o (21)
dt

(22)

where (-tUf) 0 is fixed a priori and is constant for all runs. A somewhat equivalent effect can be
achieved by using an additive bias b in the measurement model, to give:

de
- = -ke 2 +F (23)
dt

-dk = 0 (24)
dt
 306

db =0 b(O) =0 (25)
dt '

q = (-tili)kc 2 + b (26)

Both these extension options may be inferior to the original model of equations 10, 11, and 12,
since the numerical observability of the modified model may be low.
A better way that does not deteriorate numerical observability is to modifY the extended model
of equations 10, 11, and 12 as:

dc
(27)
dt

dk
=a (28)
dt

da
- = 0 a(O) = 0 (29)
dt '

q = (-tili)kc 2 (30)

Conceptually, this gives the effect of allowing in the original model the prescription element
corresponding to k to have two values instead of one, without having to wony about how the
values are to be ordered within and between runs. A similar effect can again be sought using
instead an additive bias bas:

dc
(31)
dt

dk
- = 0 (32)
dt
db
= 0, b(O) =0 (33)
dt

(34)
 307

The above explains to some extent why the applications in the literature commonly resort to
large extended-state vectors [3,7,8]. A minimal extension of the state vector leads to goal
reconciliation, allowing them to attain good estimates of the original states as well as good single-
step predictions of the measured output. Further extension of the state vector allows the use of
single-value prescription of the {ex' ey}-elements without serious deterioration of quality due to
underfit. The extension of state should, however, be made so as to not only retain theoretical
observability (by using fixed initial values, if necessary) but also numerical observability. Including
too many variables in the extended state may compromise numerical observability and jeopardize,
due to likely overfit on the past data, the validity of the deduced prescription for application on
future data. The extension states should be regarded strictly as parameters and no connection with
true physical values should be sought for them.

10 A Look at Some Common Practices

The above perspective affords some insight into certain common practices in the estimation and
identification literature.
In identification, the state consists of several parameters, the output consists of relatively few,
or one, measured variables, and the goal is only to predict the future output. Since the number of
estimated states exceeds the number of outputs to be predicted, the goal is reconcilable. However,
many identification algorithms collapse the e.' space to a single forgetting factor, thereby losing all
those degrees of freedom, and rendering the goal irreconcilable for more than one output. The
price is paid in that the left-over errors are larger, the predictions are sensitive to the e weight
specification for the outputs, and the overall quality of prediction is poorer [6].
Most estimation algorithms yield some covariance matrix for the estimated states. Many users
tend to trust this covariance as a measure of the accuracy of the obtained estimates [7]. It is clear
from the above discussion that if the goal of estimation is to predict future outputs, then the
estimated variables are to be seen as parameters for which no true value exists. In this case, it is
meaningless to talk of a measure of the accuracy of these estimates. If the goal is to estimate the
true states, then an indication of the accuracy of obtained estimates can be taken only from the
degree of fit to past data that was achieved during the trial-and-error determination of the
prescription used. There is no independent way to check how well the obtained prescription
extrapolates to new data, except to have some state measurements in the new data. The
covariance matrix again is of little use as a measure of accuracy of the estimates in the new data.
As a corollary, it is not an indicator of filter divergence either.
Perhaps the most striking aspect of much of the literature on estimation applications is the
apparent confusion with regard to the goal of the estimation exercise and the procedure used to
prescribe {ex' ey}. The goal is often not stated explicitly, but in most cases the tacit goal is
estimation of the current true states. The associated prescription, whether obtained by trial-and-
error or by adaption, is nonetheless based on striving invariably some property of the residuals,
such as zero-meanness, whiteness, or match with theoretical covariance [3,5,7]. This may be a
result of carrying over the experience from identification problems. Having thus inadvertently
 308

borrowed the additional goal of predicting the output correctly, the applications often slip into a
two-goal situation that is irreconcilable. Failing to achieve the goals by using any prescription, the
user is then forced to extend the model to make ends meet, as outlined in a previous section. Even
having gained reconcilability and enough degrees of freedom in this way, it is perplexing how most
applications end up showing good state estimates purportedly without ever using state
measurements to determine the prescription. The reason for this might lie in the belief that the
states must be taken as unmeasurable, except for verification of shown results, and in not realizing
that it is fair, indeed indispensable, to use some state measurements to get the prescription. In fact
many applications do not even concede having used an off-line "tuning" at all.

11 Conclusion

The perspective presented above clarifies some issues that have not been clear in the estimation
literature. In any estimation exercise, it is paramount to first clearly define its goal. The next step
is to check whether the goal is reconcilable and leaves some extra degrees of freedom to allow use
of constant tuning parameters. If not, the model should be appropriately enriched or extended.
Then data must be collected to enable determination of the tuning. These data must include
measured values of all variables that appear in the stated goal, also when some of these variables
happen to be states. The data may originate from the past values gathered in the application run
itself, or in independently made runs. The tuning is then obtained by trial-and-error, higher-level
optimization, or adaption. The objective thereby is to find that tuning which, for the collected past
data, gives states estimates that best reach the stated goal on the same data. The same tuning is
then used in application hoping that the underlying characteristics of the process-model-a1gorithm-
operation combination remain comparable to those of the past-data runs. The accuracy of the
obtained state estimates or output predictions can be verified only by comparison with
corresponding additional measurements, and cannot be deduced from the covariance matrix that
the algorithm might deliver. If the accuracy is not satisfactory, one or more of the above steps will
have to be changed. The steps are therefore intermingled, in practice, and do not necessarily have
to follow the order in which they are stated above.

References

I. Biegler, L.T., DBIIliano, 1.J., and Blau, G.E. Nonlinear Parameter Estimation: A Case Study Comparison.
AICHE J., 32(1), 29-45, (1986).
2. Eykhoff, P. A Bird's Eye View on Parameter Estimation and System Identification. Automatisienmgstechnik,
36(11),413-479, (1988).
3. Goldmann, S.F. and Sargent, PWH Applications of Linear Estimation Theory to Chemical Processes: A Feasibility
Study. Chem. Eng. Sci., 26,1535-1553, (1971).
4. Jazwinski, A.H. Stochastic Processes and Filtering Theory, Academic Press, New York, (1970)
5. Liang, D.F. Exact and Approximate State Estimation Techniques for Nonlinear DynBIllic Systems. Control Dynamic
Systems, 19, 1-80, (1983).
6. Ljung, L., and Gunnarsson, S. Adaptation and Tracking in System Identification - A SutVey, Automatica, 26, 7-
21, (1990).
7. Sorenson, H.W., editor, Kalman Filtering. Theory and Application, IEEE Press, New York, (1985).
8 Zeitz, M., Nonlinear Observers. Regelungstechnik, 27 (8). 241-272, (1979), (in German)
A Comparative Study of Neural Networks and Nonlinear
Time Series Techniques for Dynamic Modeling of
Chemical Processes

A. Raich, X. Wu, H.-F. Lin, and Ali Cmar

Department of Chemical Engineering, Illinois Institute ofTechnology, Chicago, IL 60616, USA

Abstract: Neural networks and nonlinear time series models provide two paradigms for
developing input-output models for nonlinear systems. Methodology for developing neural
networks with radial basis functions (RBF) and nonlinear auto-regressive (NAR) models are
described. Dynamic input-output models for a MIMD chemical reactor system are developed
by using standard back-propagation neural networks with sigm~id functions, neural networks
with RBF and time series NAR models. The NAR models are more parsimonious and more
accurate in predictions.

Keywords: Nonlinear dynamic models, input-output models, nonlinear autoregressive models,

CSIR model, neural networks, radial basis functions.

1. Introduction

Aithough most chemical processes are nonlinear systems, traditionally linear input-output models
have been used in describing their dynamic behavior. The existence of a well developed linear control
theory enhances the use of linear models. Linear models may provide enough accuracy in the vicinity
of the linearization point, but they have limited predictive capability when the system has been
subjected to large disturbances. Recently, the interest in describing chemical processes by nonlinear
input-output models has increased significantly. This is partly due to the shortcomings of linear
 310

models in representing nonlinear processes. Advanced process monitoring and model-based control
techniques are expected to give better results when process models that are more accurate over a
wider range of operating conditions are used. Model-predictive control approaches which are
becoming popular in chemical process industries pennit direct use of nonlinear models in control
algorithms. Another important reason for the increase in nonlinear model development activities is
the availability and the popularity of new tools such as neural networks which provide automated
tools for constructing nonlinear models. Yet, several other paradigms for nonlinear model
development have been available for over two decades [4]. Consequently, it would be useful to model
some test processes using various approaches, compare their prediction accuracies, and assess their
strengths and shortcomings. Neural networks (NN) and nonlinear auto-regressive (NAR) models are
the two paradigms utilized in this study. The input-output functions of the feedforward NN utilized
are radial basis functions (RBF) and sigmoid functions. RBFs [14, IS], also called local receptive
fields, necessitate only one "hidden" layer and yield an identification problem that is linear in the
parameters. The NAR modeling approach provides a linear in the parameters identification as well.
This has a significant impact on the computational effort needed in finding the optimal values of the
parameters.
In this study, various dynamic models are developed for modeling a multivariable ethylene
oxidation reactor system. The "process data" are generated with a detailed model of the reactor
developed by material and energy balances, and kinetic data from experimental studies. The reactor
equations have both multiplicative and exponential type nonlinearities. The models developed are
used for making one-step-ahead and 5-steps-ahead predictions of reactor outputs.
The paper is structured as follows. In Section 2, NNs with Gaussian RBF are outlined. NAR
modeling methodology is presented in Section 3. The reactor system and the generation of
input-output data is described in Section 4. Predictions of both types of models for various cases are
discussed in Section S.

2. Neural Networks with Radial Basis Functions

Neural networks have been utilized to create suitable nonlinear models, especially for use in pattern
 311

recognition, sensor data processing, forecasting, process control and optimization [1, 5, 6, 8, 9, 11,
17, 19,21,22,23,24,25,26]. Sigmoid functions have been most popular as input-output functions
in NN nodes. RBFs provide alternative nonlinear input-output functions which are "locally tuned".
A single-layered NN using RBFs can approximate any nonlinear function to a desired accuracy [3].
By oVerlapping the local receptive fields, signal-to-noise ratios can be increased to provide improved
fault tolerance [14]. Control with Gaussian RBF networks has been demonstrated to be stable, with
tracking errors converging towards zero [20]. First order-lag-plus deadtime transfer functions in the
network nodes has also been discussed [24].
RBF approximation is a traditional technique for interpolation in multidimensional space. A RBF
expansion with n inputs and a scalar output generates a mapping t;: 9tD ..... 9t according to

n.
f(x) = Wo + L Wi g(lIx - Cill)
0 (1)
i=1

where x E 9t", g(o) is a function from 9t+ to 9t, II,; is the number ofRBF centers, Wi, 0 ~ i ~ II,; are

the weights or parameters, t; E 9t", 1 ~ i ~ II,; are the RBF centers, and I I denotes the Euclidian
0

norm. The equation can be implemented in a multilayered network (Figure 1) where the first layer is
the inputs, the second layer performs the nonlinear transformation and the top layer carries out the
weighted summation only. Notice that the second layer is equivalent to all hidden layers and the
nonlinear operation (with sigmoid functions) at the output layer nodes of the NN with sigmoid
functions. A frequent choice for the RBF, g ( I x-t; I ), is the Gaussian function

g(x) = exp( -/Ix - cdIIP;) (2)

where, Pi is a scalar width such as the standard deviation. Given the numerical values of the centers
Ci and of the width, Ph determination ofbest values of the weights, Wi, to fit the data is a standard

model identification problem which is linear in the parameters. If the centers and/or the width are not
predetermined and are adjustable parameters whose values are to be determined along with weights,
then the RBF network becomes equivalent to a multi-layered feedforward NN, and an identification
which is nonlinear in the parameters must be carried out.
 312

A popular algorithm for choosing

the centers and the widths is k-means
clustering, which partitions the data set
X = [IJ] into k clusters, finds their
centers so as to minimize the total
output layer
linear combination distance I . I of the x vectors from
their nearest center. Widths Pi can then
be calculated to provide sufficient
nonlinear layer
overlap of the Gaussian functions
with REF
around these centers and ensure a
smooth, continuous interpolation over
X while keeping the RBFs local enough
input layer
so that only a small portion of the
network contributes for relating an
Figure 1: Neural network structure with radial basis functions as
nonlinear functions in nodes. input vector I of X to the respective
output. This localization makes the function with the closest center t; to the input I the strongest
voice in predicting the corresponding output. The number of nodes in the nonlinear transformation
layer is set equal to the number of clusters, k. For each of the k clusters or nodes, a Gaussian width,
Pi can be found to minimize the objective function

(3)

where n is the number of columns in X (the number of input vectors used in training) and p is an
overlap parameter which assures that each cluster partially overlaps with its neighboring clusters.
With the RBF relations in each node fixed, the weights can be chosen to map the inputs to a variety
of outputs according to
k
ft(Xj) = WIO + 2: Wli . g(lIxj - cd!) (4)
i=l

where I is the number of outputs. Hence, for each output q only the weights W qi are specific to that
 313

output, while the centers and widths are common to all outputs. With the k-means clustering
algorithm to select the clusters and compute their centers, optimization of the Gaussian widths will
fix the form of g(IJ and node weights can be easily optimized to model a multivariate mapping of X
to 11 (X). The selection of the cluster members, computation of cluster means and widths are done
first, as unsupervised learning. The computation of the weights is carried out as the training of the
neural net, i. e. supervised learning.

3. Nonlinear Time Series Models

Classical model structures used in non-linear system identification have been the functional series
expansions of Volterra or Wiener which map past inputs into the present output. This moving average
type approach results in a large number of coefficients in order to characterize the process.
Input/output descriptions which expand the current output in terms of past inputs and outputs provide
parsimonious models. The non-linear autoregressive moving average with exogenous inputs
(NARMAX) model [12], the bilinear mode~ the threshold model and the Harnmerstein model belong
to this class. In general, NARMAX models consist of polynomials which include various linear and
nonlinear terms combining the inputs, outputs and past errors. Once the model structure, the
monomials to be included in the mode~ has been selected, the identification of the parameters can be
formulated as a standard least squares problem which can be solved using various well-developed
numerical techniques. The number of all candidate monomials to be included in a NARMAX model
ranges from about a hundred to several thousands for moderately nonlinear systems. Determination
of the model structure stepwise regression type of techniques become inefficient. Instead, methods
of model structure determination must be developed and included as a vital part of the identification
procedure. An orthogonal algorithm which efficiently combines structure selection and parameter
estimation for stochastic systems has been proposed by Korenberg [10] and later extended to MlMO
nonlinear stochastic systems [12].
In this paper, a special case ofNARMAX mod~ the nonlinear autoregressive (NAR) model, and
the classical Gram-Schmidt (CGS) orthogonal decomposition algorithm using Akaike Information
Criterion (AlC) are presented.
 314

Nonlinear Model Representation and CGS Orthogonal Decomposition Algorithm

A discrete time multivariable nonlinear stochastic system with m outputs and r inputs can be described
by the NARMAX model [12]

yet) = f(y(t - 1), ... , yet - n,), u(t - 1), ... , u(t - n u ), e(t - 1), ... , e(t - no» + e(t) (5)

where
YI(t») (UI(t -1») (el(t»)
yet) =( : ,u(t) = : ' e(t) = : (6)
Ym(t) ur(t - 1) em(t)

are the system output, input and noise respectively; fly, nil' ne are the maximum lags in the output,
input and noise; {e(t)} is a zero mean independent sequence; and f(') is some vector valued nonlinear
function.
A special case ofthe NARMAX model is the NAR model

yet) = f(y(t -1), ... ,y(t - ny» + e(t) (7)

which can be expanded as

Yq(t) = fq(YI(t -1), "',YI(t -ny), ... , Ym(t -1), .··,Ym(t - ny» +eq(t), q = 1, ... , m (8)
Writing fq(') as a polynomial of degree 1 yields

Yq(t) = 8~q) + t 8~~)Xil(t) t t 8~~l2Xil(t)Xi2(t)

i 1=1
+
i1=1 i 2=;1
n n

+ ... + L ... L 8~~!.. i,Xil (t) ... Xi, (t) + eq(t), q = 1, ... , m (9)
i 1 =1 i 1 =;,-1

where
n = m x n, (10)
XI(t) = Yl(t -1), X2(t) = Yl(t - 2)"", xmn.(t) = Ym(t - ny) (11)
All terms Xii (t) ... Xi, (t) in Eq. (9) are given. Hence, for each q, 1 :$ q:$ m, Eq. (9)
describes a linear regression model of the form
M
Yq(t) = LPi(t)8i + ~(t), t = 1,···, N (12)
i=1
 315

where M = 2::=1 mi with mi = mi-I· (ny ·m+i -1)/i, N is the time series datalength,p,(/) are
the monomials of degree upto I which consist of various combinations ofxlt) to xn(t) (n defined in
Equation (10», (I) is the residual, and 8, are the unknown parameters that will be estimated. pl/)
= 1 is the constant term. In matrix form Equation (12) becomes

(13)
where

Usually a model consisting of a few monomials describe the dynamic behavior of most real
processes to desired accuracy. This is a small subset of the monomials in the full model.
Consequently, the development of an algorithm that can select the significant monomial terms
efficiently is of vital importance. A method has been developed for the combined problem of structure
selection and parameter estimation [2]. The task is to select a subset P, of the full model set P and
to estimate the corresponding parameter set ea. Several least squares solution methods can be utilized,
but methods based on orthogonal decomposition ofP offer a good compromise between computation
accuracy and computation "cost".
Gram-Schmidt Orthogonal Decomposition Algorithm The classical Gram-Schmidt (CGS)
orthogonal decomposition ofP yields

aIM
a2M )

: , W = (Wl ...WM) (15)

o
o
o
o
aMi1,M

where A is an M x M unit upper triangular matrix and W is an N x M matrix with orthogonal columns
that satisfy WTW = D with the positive diagonal matrix D.
The selection of the specific monomials that will make up the model structure and the estimation
 316

of their coefficients can be combined by extending the orthogonal decomposition techniques. Let Ps
be a subset ofP withM.. columns such that 11. <M and 11.:s: N. Factorize Ps, into W,A. where W,
is an N x M.. matrix with 11., orthogonal columns and A. is an 11. x 11. unit upper triangular matrix.
The residuals can be expressed as

Denoting the inner product as <.,. >, and rearranging Equation (16) as Yq = W.9. +~, the sum
of squares of the dependent variable Yq is

M,
< Yq,Yq >= '"'
~9i2 < Wi,Wi > +< ••
::::,:::: > (17)
i=1

The reduction in the residual due to the inclusion of W, in the regression can be measured by an error
reduction ratio (li defined as the proportion of the dependent variable variance explained by W, ,

2
n._9i <Wi,Wi>
1- < Yq,Yq > (18)

The error reduction ratio can be used for extracting W, from W and consequently P, from P by
utilizing the CGS decomposition procedure. Let k denote the stages of the iterative procedure. At
the first stage (k = 1), set w/ = Pi for i = 1, ... , M and compute

(i
(i Y.
< wI' q > n(i _
(91(i)2 (i (i
< WI ,wI>
91:=1 =< (i (i
WI ,WI >
' I -
<
Y. Y.
q' q >
(19)

Select as wJ the w/ that causes the maximum error reduction ratio: w J = w/ such that {} (/ = max
{} (/ , i = 1, ... ,M). Similarly, the first element of g, is gl = g(/ .
At the kth stage, excluding the previously selected j's compute for i = 1,··· , M (i '* J)
 317

< Wl,Pi >

Oi
(i
l,k -
-
<WI, WI >'
(i
Oi k _ l ,k = <<Wk-1,
Wk-l>Pi >
Wk-1 >
(20)

1:-1 (i
(i "" (i <Wl;,Yq > (21)
WI: =Pi- LOiII: W1 , gJ: = (i (i
1=1 < WJ: ,WJ: >
Let n (/ = max(n (f , 1 ~ i ~ M (i ~ all previous))). Then wt = W (/ is selected as the kth
column ofW" together with the kth column of A" a lt =a // (/ = 1, ... , k-1) the kth element of g"
9. = g(/, and n k = n(/.
The selection procedure can continue until a prespecified residual threshold is reached. However,
since the goal is to develop a model that will be used for prediction, it is better to balance the
reduction in residual and increase in model complexity. This would reduce the influence of noise in
the data on the development of the process model. The Akaike Information Criteria (AlC) is used to
guide the termination of the modeling effort.

1 ..
AIC(k) = Nlog( N < :=:,:=: » + 2k (22)

Addition of new monomials to the model is ended when AlC is minimized. The subset model
parameter estimate 8, can be computed from A,8, = g, by backward substitution.

4. Multivariable Chemical Reactor System

A reactor model based on mass and energy balances is available for simulating the behavior of
ethylene oxidation in a nonadiabatic internal recycle reactor [18, 16]. Ethylene reacts with oxygen to
produce ethylene oxide. A competing total oxidation reaction generates CO2 and water vapor.
Furthermore, the ethylene oxide produced can dissociate to yield CO2 and H20. All three reactions
are highly exothermic. The reactor input variables that are perturbed are inlet flowrate, inlet ethylene
concentration, and inlet temperature. The output variables are outlet ethylene and ethylene oxide
concentrations and outlet temperature. Data is generated by perturbing some or all inputs by
pseudo-random binary sequences (PRBS). The two main cases studied are based on PRBS forcing
of (a) inlet ethylene concentration and total inlet flow rate, and (b) inlet ethylene concentration, total
 318

flow rate and inlet temperature. The first case has multiplicative interactions while the second case
provides exponential interactions as well. Results for the second case is reported in this
communication. Data was collected at three different PRBS pulse duration: 5, 10, or 25 s. In all
input-output models, only reactor output data are used. The length of the time series is 1000. For NN
training all 1000 values are used while for NAR models 300 data points are utilized. A second data
set has been developed for each case, in order to assess the predictive capabilities of the models
developed.

5. Nonlinear Input-output Models of the Reactor System

Dynamic Models Based on Neural Network

Dynamic models are constructed to predict the one-step-ahead and 5-steps-ahead values of reactor
outputs based on past values of outputs. The current and the most recent four previous values of all
reactor output variables are fed to the NN as the inputs and either the next or the five time steps
ahead value of the same variable is used as the output. Consequently, for a NN with generic input(s)
z(t) and generic output y(t)

y(t) = z(t + 1) = f([z(t) z(t - 1) z(t - 2) z(t - 3) z(t - 4)]) (23)

for the one-step-ahead prediction, while y(t) = z(t +5) for the 5-steps-ahead prediction.
Neural Networkwith Sigmoid Fwictions: A commercial package has been utilized for developing
the standard neural networks with sigmoid functions. Various sizes of hidden nodes have been tested
and a single hidden layer with 12 hidden nodes has provided the best prediction accuracy [13].
Neural Network with Radial Basis Functions: Training the network is done by both supervised
and unsupervised methods. Clustering using the K-means algorithm and optimization of widths is
unsupervised, dependent only on the network inputs, not the specific output to be modelled. Once
the centers and widths are determined, supervised learning of the weights for each output variable
is done. The K-means algorithm is implemented in FORTRAN to choose centers for the Gaussian
 319

functions. Davison-Fletcher-Powell method with unidimensional Coggins search [7] is used for the
optimization of Gaussian function widths and for the selection of weights for the various reactor
outputs. The network parameters (centers, widths, and weights) were saved for each of the training
cases for use with additional reactor data sets to test the predictive accuracy of networks developed.
Netwo~ks with various numbers of hidden nodes were trained and tested to find the number of
nodes yielding the smallest predictive sum of squared error and Akaike Information Criterion.

Dynamic NAR Models

The pool of monomials to be used in the NAR models included all combinations of the current and
immediate past 3 values of all three reactor output variables, resulting in 82 candidates. Models with
past 4 values made little improvement. NAR modeling was conducted with a 300 line C program,
having typical execution times on to 5 minutes on the Vax station for time series lengths of300 for
all 3 variables.

Discussion of Results

Neural Network Models. There were no clear trends in the optimal number of nodes, sum of squared
errors or AlC for the best networks for I-ahead or 5-ahead predictions, with or without noise in data
(Table I). Generally, as PRBS pulse duration increased, the number of nodes increased. For one at
a time variable prediction, while outlet ethylene concentration was nearly always the most difficult
variable to learn (minimum error was reached with higher numbers of nodes than the other two
variables), trends across time horizon or absence/presence of noise were not apparent.
The prediction plots (Figures 2-5), show that much of the prediction error is at level changes,
when a variable alternates from a fairly steady low value to a fairly steady high value. In all plots,
actua1 data is shown in solid lines, NN-RBF predictions with dotted lines and NAR predictions with
broken lines. Ethylene, ethylene oxide concentrations and temperature are denoted by y2, y3, y4,
respectively. At such abrupt changes, the NN often predicts a sharper change than actually occurs,
and then retreats to a steady value not as extreme as the real value. This tendency could be due to the
size of the window of past values used for prediction: with a smaller window than the 5 past values
 i-Ahead Prediction. S-Ahead predictions

HO HOISE .5 , NOISE HO NOISE .• 5 , 1l0ISE

PRBS Forcing Period of 5 5 PRBS forcing Period of 5 B

Training Prediction Training Prediction Training Prediction Training Prediction
best K J J J best K J ] J J

c. 0.2H 0.246 0.086 0.095 Ce 0.418 0.09 0.421 O.HI

Ceo 0.001 0.002 0.001 0.002 Ceo 0.001 0.002 0.001 0.002
T 0.004 0.00] 0.046 0.048 T 0.005
____ a _ _ _ _ _ _ _ _ y 0.004
______ 0.006 0.004
--------_.--------.-
0.251 -----.-.-----~-.--
0.1)) 0.14 4 SSE 0.425 0.445 ----.-.-----.--~--
0.418 0.448
SSE 0.2J9
n-SSE 0.05J 0.075 0.028 0.047 n-SSE 0.089 0.112 0.088 0.111
AlC 0.596 0.5H AlC 0.662 0.662

PRBS Forcing Period of 10 • PRBS forcing Period of 10 s

Training prediction Training Prediction Training Prediction Training Prediction
best K J 3 J J best K 10 10 J J
-------------------- -~---------------- -------------------- ------------------
Ce 0.122 0.193 0.122 0.19J Ce 0.238 0.381 0.325 0.478
Ceo 0.001 0.003 0.001 0.003 Ceo 0.0023 O.OOll 0.0010 0.0026
T 0.001 0.001 0.001 0.007 T 0.0101 0.0309
_________ a_. _____ . __ 0.0078 O.On~9

-------------------- ------------------
0.203 0.250 0.416 ------------------
O. JJ5 W
SSE 0.124 0.203 O.IH SSE 0.H7 I\)
n-SSE 0.014 0.091 O.OH 0.091 n-SSE 0.081 0.123 0.108 0.121 0
"IC 0.557 0.557 Ale 1.172 0.695

PRBS Forcing Period of 25 5 PRBS Forcing Period of 25 5

Training Prediction Training Prediction Training Prediction Training Prediction
best K 10 10 J J best K 1 1 5 5
-------------------- ------------------ -------------------- ------------------
Co O.OH 0.047 0.064 0.076 C. 0.211 0.218 0.216 0.24 B
Ceo 0.001 0.001 0.004 0.004 Ceo 0.0015 o.OOH O.OOH 0.0042
T 0.001 0.002 0.009 0.008 T ~_. 0.0085 0.0051 0.0017
• __________ . _ . _ • • D
_____
0.0084 _____ .x. __ ..
SSE
--------------------
0.046 0.049 0.016 0.089 SSE 0.225 0.250
---------
0.225
... ------
0.260
n-SSE 0.014 0.017 0.059 0.077 n-SSE 0.081 0.093 0.080 0.108
AIC 1. 03 6 0.608 AIC 0.649 ,0.810

n-sse = Weighted average of the squared error where the wetght is equal to l/(average of the real value for that variable)

Table 1. Sum of squared errors and Ale for best neural networks
 321

1-5IE' AHEAD PREDlcnON (DATA SET 7 wrrn NO NOISE) ~-5IE' AHEAD PREDlcnON (DATA SET 7 wrrn NO NOISE)

.i
N

,J
e
j:;,
E
"u
C . I
o .
U
c
"
"
>,
-- I
I,

.c 'J
til

.,0;---.';';;'0--.';;;00--.'""":---."":;;;---.""'0,.--"300 O.l o\----.,,,-o--.'oo;;;----;U""O,.--·"";;;;---.,,;:,o,----.l.)OO

,:: .J\
10
!
'.f>

....
..,
\
I
\
-'::
\----.,,,-O--.,OO;;;----;'"""--.200:;;;---.""'O,.---.i,]OO 9n o;---.,;::'0--.,"'oo,.---:'''''',.--'200;;;;c---.'''''0,------')OO

Time Time

Figure 2: One-step (1) and five step (R) ahead predictians. Data with no noise and PRBS forcing period of 10 s.
 322

I-Sll!!' AHEAD PREDlcnOH (DATA SlIT 7 WID! HOISE) J.STEP AHEAD PREDlcnOH (DATA SlIT 7 WITH NOISE)

'I

,..,;>-.
"
~ §
0-e
>: .-
.,c
., ::.,
-;>-. u
c
-5 0
WU

0-10 la . .. 10 '00 Ila ,<0 '60 '10 100 ala 20 <0 .. '" '00 no ..a '60 II. zoo

la'
l[US

'"'"
OJ)

-5"
'-
0

:;" ~ :.
"" e'" '.'
>,
....
...
Q. US

..
C ~

.. ..
C3 .- '.15
f- a 9.10
'"
1,';
10 .00 Ila ... '60 .80 200
,.75 0
20

Time Time

Figure J: One-step (L) and five step (R) ahead predictions. Data with no noise and PRBS forcing period of lOs. Lighter
solid lines show prediction with neW'al networlcs with sigmoid functions.
 323

3-'

J ~
(
) f:
r-.. fI

')!.
r- ~
.~
e
;; :u

8
~
~
_•..
.~
. .··v:
i-
Iii
1-' V\ .,.;..
\::
\:.:.:.:
..
~

10 '0 100 130 200 250 300

3-'
N ')!.
>,

eE" .-----.
c .2

gc
0

~
"uc 8
0 ~
u ~
"c Iii
"
]>
Iii
10 '0 100 1'0 200 250 300

3-'
')!.
'f.
.2
~
il tlui'~
U
~ :u
8
~
~ 1

;ll -r-C:::::===.I. _I .'-

1.!l F .. .

10 '0 100 1'0 200 250 300

Time

Figure 4: One-step ahead predictioos of ethylene concentration. Data with no noise and PRBS forcing period of 5. 10. 255.
 324

(V
3..5

J\L (\
( {
(I
J. .'-~ r.
~

."-. .... ....--- I·...... ~.

1J V "'-_J \ \.... \

10 SO 100 ISO 200 250 300

3..5
/\
\
f\J (\
3~y . . ,. _
..
/
N
» -. 11 ;: t
<:
.~
0 '>!. ': ... :

!::
"
.R
<:
8<:
~ ~

~
u0 8
"<: i .")
.......
);
"
;>, ill
-=
Iil
I.S .~
~ \ \.... \I

10 so 100 ISO 200 250 300

3..5

2.5

1-'

10~--~SO~--~100~--~I~SO~--~200=---~25~0----~300

Time

Figure 5: One-step ahead predictions of ethylene concentration. Data with no noise. 1% and 2% noise and PRBS forcing
period of} 0 s.
 325

used fuirly equally in clustering, more recent values would have more impact on prediction, perhaps
enabling better "steady" predictions. The predictions based on neural networks with sigmoid functions
(NN-S) are also shown in Figure 3 (thin solid lines). The predictions with NN-S are better for I-ahead
predictions of ethylene and ethylene oxide concentrations, but worse for 5-ahead predictions
compared with NNRBF predictions. The period of the forcing functions of reactor inputs have a large
effect on the accuracy ofNN prediction (Figure 4). As the PRBS period increased, the fit and the
predictions improved. As expected, in general excessive noise in data degraded the prediction
accuracy (Figure 5).
l-Ahead Predictions However, with a 'small"
TRAININC ------- .......... . PREDICTION . ...... .
NOISE' 0.5 a 0.5 1 2
amount of noise, NNs
# Hodes 15 15 15 15

Ce 0.1223 0.1224 0.0406 0.0621 0.1931 0.1927 0.1347 0.1101

trained with noisy data not
Ceo 0.0006 0.0007 0.0027 0.0018 0.0033 0.0033 0.0020 0.0021
T 0.0010 0.0013 0.0021 0.0079 0.0067 0.0069 0.0055 0.0394 only yielded errors of the
SSE 0.1240 0.1245 0.0496 0.0719 0.2031 0.2030 0.1423 0.1517
n-SSE 0.0336 0.0337 0.0547 0.0416 0.0907 0.0908 0.0588 0.0561 same magnitude as those
AIC 0.5569 0.5572 1.2995 1.2729

5-Ahead Predictions
trained without noise, but
TRAININC PREDICTIOU
NOISE' 0.5 0.5 1 the NNs with noise made
I Nodes 10 10 10 10
predictions, in some cases,
Ce 0.2380 0.3247 0.3253 0.2517 0.3813 0.478 0.4779 0.3999
0.0023 0.003 0.0030 0.0041 0.0033 0.0026 0.0026 0.0025
Ceo
T 0.0101 0.0078 0.0084 0.0172 0.0309 0.0059 0.0066 0.0565 with smaller total errors
SSE
n-SSE
0.2505 0.3355 0.3368 0.2731 0.4157 D.U65 0.4873 0.4589
0.0827 0.1083 0.1089 0.1151 0.1234 0.1266 0.1281 0.1133
than the corresponding
AIC 1.1750 0.6946 0.6955 1.2280
NNs without noise,
sometimes with only half as
much error (Table II).
Table 2. Effect of noise on NN selection. SSE for data with PRBS period of 10 s.
There is no clear trend in
improvement of prediction accuracy as a function of the number of nodes used. For the case
considered in Figure 6, three nodes yielded the smallest SSE, the next best was 15 nodes. A reason
for the poor performance of the NN with respect to NAR may be due to clustering data from
multivariable systems The clusters may be "nonrepresentative". A feedforward NN with sigmoid
functions was developed by utilizing a commercial package. After various adjustments of the data,
and long training times (over 2 days with a PC/286 with a math coprocessor) much better fit was
obtained. This may indicate the limitations of clustering when a multivariable NN is being developed.
Additional work is currently being conducted on this issue. Norma1ization of the data used in training
the NN with RBF did not yield any appreciable improvement.
 I-Ahead Prediction 5-Ahead Prediction

without Noise with 0.5 % Noise without Noise with 0.5% Noise

Variables Models Prediction Model Prediction Variables Models Prediction Model Prediction
Terms Errors Terms Errors Terms Errors Terms Errors

PRBS Input forcing period: 5 s. PRBS Input forcing period: 5 s.

Ce 3 2.67E-2 3 2.67E-2 Ce 3 3.67E-5 3 3.90E-1

CeO 5 6.87E-5 6 7.53E-5 CeO 5 1.48E-3 6 1.6OE-3
T 6 1.34E-4 9 4.28E-3 T 6 4.12E-3 9 8.IOE-3

SSE 2.69 3.12 SSE 5.6IE-1 3.998+1

AIC -1.398+3 -1.33E+3 AlC -1.868+3 -5.698+2

PRBS Input forcing period: lOs. PRBS Input forcing period: lOs.

Ce 2 1.678-2 2 1.67E-2 Ce 2 3.098-1 2 2.82E-1 c.>

I\J
CeO 4 3.078-5 4 3.27E-5 CeO 4 I.IIE-3 4 9.98E-4 0>
T 4 6.388-5 7 2.34E-3 T 4 2.87E-3 7 7.67E-3

SSE 1.68 1.91 558 3.12E+1 3.008+1

AIC -1.548+3 -1.498+3 AIC -6.598+2 -6.658+2

PRBS Input forcing period: 25 s. PRBS Input forcing period: 25 s.

Ce 3 8.88E-3 3 8.9IE-3 Ce 3 1.61E-1 3· 1.59E-1

CeO 4 2.5IE-5 4 2.628-5 CeO 4 7.24E-4 4 7.238-4
T 5 3.49E-5 9 4.35E-4 T 5 1.46E-3 9 1.89E-3

SS8 8.94E-1 9.36E-1 SS8 1.63E+ I 1.621!+1

AIC -1.728+3 -1.701!+3 AIC -8.491!+2 -8.441!+2

Table 3_ Sum of squared errors in prediction with best NAR models

 327

Effect of Number of Neural Network Nodes on Prediction

c
.12 0.4
...
';
c..,
u
c 0.35
8..,
"C
'x 0.3
0
..,c
..,
>.. 0.25
.c
Iii
....0
c 0.2 :0
.2
.~ ~.
:-~
..,
"C

t.t: 0.15
"C
..,'" Nodes
2 3 5
-
10 15
... 0
....~ 0.1 0
50 100 150 200 250 300

Time

Figure 6. One-step ahead ethylene concentration predictions. Data with DO Doise and PRBS forcing period of 10 s.

Nonlinear Auto-Regressive Model: For all cases considered, NAR models provided more accurate
predictions than NN models (Table III). Since predicted values are used in predicting multi-step ahead
values"the accuracy is affected by the prediction horizon (Figures 2-3). Noise in data necessitates
more monomials, and 3 to 9
Monomial Terms AIC SSE AIC SSE Terms monomials are enough to
Number Ce Ce Ce CeO CeO CeO

I y,(1<-I) -1247.6 7.22 -2876.9 0.0503 y,(k-I)

minimiz.e Ale. The effect of each
2 y,(k-2)y.'(k-1) -1458.9 3.75 -3374. I 0.0109 y,(1<-2)y,(1<-1)
3 y,(k-I) -1456.3 3.73 -3378. I 0.0106 y,'(k-2)y,(k-2)
monomial added to the NAR
4 -3384.3 0.0103 y,Ik-10y,(k-2)
5 -3380.6 0.0102 y,'(k-I)Y,Ik-1) model is tabulated for one case in
6
Table IV. The last terms in each
column show the effect of the
Table 4. SSE after adding each monomial-PRBS input period 10 s.
 328

next candidate monomial which has not been included in the model. The best NAR models for reactor
data generated by PRBS forcing period of lOs and with no noise or with white measurement noise
superimposed are, respectively

Y2(k) = 1.6113Y2(k - 1) - O.0062y,(k -1)2Y2(k - 2)

Y3(k) = 1.6952Y3(k -1) - O.0715Y3(k - 2)Y4(k - 2)
+ O.0005Y2(k - I?Y4(k - 1) - O.0039Y2(k - I)Y2(k - 2); (24)
Y4(k) = 1.8003Y4(k -1) - O.7548Y4(k - 2)
- O.0005y,(k - l)y,(k - 2)2 + O.0260Y2(k - I)Y3(k - 2)

Y2(k) = 1.6085Y2(k -1) - O.0061Y2(k - 2)Y4(k _1)2

Y3(k) = l.7402Y3(k - 1) - O.0736Y3(k - 2)Y4(k -1)
+ O.0005Y2(k - 2)2Y4(k - 2) - O.0044Y2(k - I)Y2(k - 2); (25)
Y4(k) = 1.7456Y4(k -1) - O.0070Y2(k - 1)3 - O.I053Y4(k - I?
+ O.0821Y2(k - 1?Y3(k - 2) + O.0029Y4(k - 1)2Y4(k - 2)
- O.1508Y2(k - 2)Y3(k - 2) + O.0030Y2(k - 1)2Y4(k - 2) .

The best models for reactor data with no noise and PRBS forcing periods of 5 and 25 s are,
respectively
Y2(k) = 1.5557Y2(k -1) - O.0064Y2(k - 2)Y4(k _1)2 + O.0002Y4(k _1)3
Y3(k) = 1.7402Y3(k -1) - O.0718Y3(k - 2)Y4(k - 2) + O.0355Y2(k -1)
- O.OI39Y2(k - 2) + O.0003Y2(k - 2)3 - O.0003Y2(k - I)Y4(k - 2)2
y,(k) = 1.2795Y4(k -1) - 4.9542Y3(k - 2)3 + 414.9254Y3(k _1)2 (26)
- O.0031Y4(k - 1)3 - O.0410Y2(k - 2)Y3(k - 2)Y4(k -1)
+ 35.6044Y3(k - I)Y3(k - 2)Y4(k - 1) - 40.1064Y3(k - 1)2Y4(k -1)
+ O.0565YY2(k - 2)2Y3(k - 2) - 362.9179Y3(k -1)Y3(k - 2)
 329

Y2(k) = 1.6514Y2(k -1) - O.6306Y2(k - 2) - O.0008Y2(k - 2)2Y4(k -1)

Y3(k) = 1.7662Y3(k -1) - O.7761Y3(k - 2) + O.0111Y2(k -1) - O.OOOlY2(k - 2)Y4(k - 2)2
y,(k) = O.6039y,(k -1) + O.3873y,(k - 2) + O.1299Y2(k -1) (27)
- O.0059Y2(k -1)Y2(k - 2)y,(k - 2) + O.0047Y2(k - 2)3
+ O.1053Y2(k -1?Y3(k -1) - O.2363Y2(k - 1)Y3(k - 2)Y4(k - 1)
+ 2.0335Y2(k - 1)Y3(k - 2) + O.1319Y2(k - 2)Y3(k - 1)

6. Conclusions

Nonlinear time series modeling techniques and neural network techniques offer methods that can be
implemented easily for modeling processes with severe nonlinearities. In this study, the NAR models
have provided more accurate predictions than NN models with radial basis functions. These results
are certainly not conclusive evidence to draw general conclusions. However, they indicate that other
nonlinear modeling paradigms can be used as easily and may provide as good models as the NN
approach. Both approaches have strong points: NN models may be trained directly for multistep
ahead predictions, while NAR models are more parsimonious and the functional relationships can
provide physical insight. The availability of general purpose software and the capability to capture
nonlinear relations have made NN a popular paradigm. This popularity has fueled the interest in other
nonlinear modeling paradigms. We are hopeful that future studies will provide powerful nonlinear
modeling methods as well as guidelines and heuristics in selecting the most appropriate paradigm for
specific types of modeling problems.
 330

References
I. Bhat, N. and T. J McAvoy (1990): Use of Neural Networks for Dynamic Modeling and Control of Chemical
Process Systems, Computers Chem. Engng 14 (4/5) 573
2. Chen S., S. A. Billings and W. Lou (1989): Orthogonal least squares methods and its application to non-linear
system identification, Int. 1. Ctrl., 50 (5), 1873-1896
3. Cybenko, G. (1989): Approximations by SupCl]JOsitions of a Sigmoidal Function, Math. Cont. Signal & Systems,
2 303-314
4. Haber R. and H. Unbehauen (1990): Structure Identification of Nonlinear Dynamic Systems - A survey on
Input/Output Approaches, Automatica, 26, 651-677
5. Haesloop, D. and B. Holt (1990): A Neural Network Structure for System Identification, Proc. Ameri. Cntrl Conf.
2460
6. Hernandez, E. and Y. Arlam (1990): Neural Network ModelIing and an Extended DMC Algorithm to Control
Nonlinear Systems, Proc. Amen. Cntrl Coof. 2454
7. Hinunelblau, D. M. (1972): Applied Nonlinear Programming, McGraw-HilI, New York.
8. Holcomb, T. and M. Morari (1990): Analysis of Neural ControlIers, AlChE Annual Meeting. Paper No. 16a.
9. Hoskins, 1. C. and D. M Himmelblau (1988): Artificial Neural Network Models of Knowledge Representation in
Chemical Engineering Comput. Chem. Engng 12, 881
10. Korenberg M. 1. (1985): Orthogonal Identification of Nonlinear Difference Equation Models, Models, midwest
Symp. on Circuits and Systems, Louisville, KY
II. Leonard, 1. A. and M Kramer (1990): ClassiJYing Process Behavior with Neural Networks: Strategies for Improved
Training and Generalization, Proc. Amen. Cntrl Coof. 2478
12. Leontaritis I. 1. and S. A. Billings (1985): Input-output parametric models for nonlinear systems, Int. J. Ctrl., 41,
303-344
13. Lin, Han-Fei (1992): Approximate Dynamic Models with Back-Propagation Neural Networks, Project Report,
I1Iinois Institute of Technology
14. Moody and Darken (1988): learning with Localized Receptive Fields, Research Report YALEUIDCSIRR-649,
Yale Computer Science Department, New Haven. Connecticut
15. Niranjan M. and F. Fallside (1988): Neural Networks and Radial Basis Functions in Classifying Static Speech
Patterns, Report No. CUEDIF-INFENGITR 22. University Engineering Department, Cambridge, England
16. Ozgulsen F., R. A. Adomaitis and A. Cinar (1991): Chem. Eng. Sci.• in press
17. PoUard, 1. F., D. B. Garrison, M R Broussard and K Y. San (1990): Process Identification using Neural Networks,
AlChE Annual Meeting. Paper No. 96a
18. Rigopoulos, K (1990): Selectivity and Yield Improvement by Forced Periodic Oscillations: Ethylene Oxidation
Reaction, Ph D. Thesis, Illinois Institute ofTechnology, Cbicago,IL
19. Roat. S. and C. F. Moore (1990): Application of Neural Networks and Statistical Process Control to Model
Predictive Control Schemes for Chemical Process IndustIy. AlChE Annual Meeting. Paper No. 16b.
20. Sanner and Siotine (1991): Gaussian Neural Networks for Direct Adaptive Control, Proc. Amer. Control conC.•
2153
21. Ungar. L. H, B. A PowelJ and S. N. Kamens (1990): Adaptive Networks for Fault Diagnosis and Process Control,
Computers Chem. Engng 14 (4/5) 561
22. Venkatasubramanian, V., R Vaidyanathan and Y. Yamamato (1990): Process Fault Detection and Diagnosis Using
Neural Networks: I. Steady State Processes. Compu!. Chem. Engng 14.699
23. Whiteley, 1. R. and 1. F. Davis (1990): Backpropagation Neural Networks for Qualitative InteJpretation of Process
Data, AlChE Annual Meeting. Paper No. 96d
24. Willis, M. 1., G. A. Montague, A. 1. Morris, andM. T. Tham (1991): Artificial Neural Networks:-APanacea to
Modelling Problems?, Proc. Amen. Cntrl Coof. 2337
25. Y80, S. C. and E. Zafiriou (J 990): Control System Sensor Failure Detection via Network of Local Receptive Fields.
Proc. Ameri. Cntd Coof. 2472
26. Ydstie, B. E. (J 990): Forecasting and Control Using Adaptive Connectionist Networks, Computers Chem. Engng
14 (4/5) 583
Systems of Differential-Algebraic Equations

R. W. H. Sargent

bnperial College of Science, Technology and Medicine, Centre for Process Systems
Engineering, London, SW7 2BY, UK

Abstract: The paper gives a definition and a general description of the properties of systems of
differential-algebraic equations. It includes a discussion of index and regularity, giving simple
illustrative examples. It goes on to describe methods for index-reduction and numerical methods
for solution of high-index problems.

Keywords: Differential-algebraic equations, high-index problems, ordinary differential equations,

regularity, index-reduction, numerical solution, applications

1. Introduction - What is the Problem?

Most chemical engineering degree courses give a good grounding in the theory of differential
equations, and in numerical methods for solving them. However, dynamic models for most process
systems consist of mixed systems of differential and algebraic equations and these sometimes have
unexpected properties. It is the purpose of this talk to describe the properties of such systems, and
methods for their numerical solution.

1.1 Systems of Ordinary Differential Equations (ODEs)

Let us start by recalling the general approach to the numerical solution of ordinary differential
equations (ODEs) of the form:

x (t) = f(t, x (t», where t € R, X (t) € R nand f: R x R n - R n (1.1)

For numerical solution we seek to generate a sequence {Xn}, n = 0, 1, 2, ... which

approximates the true solution, Xn " x (t n ) at a sequence of times {x n }, n = 0, 1,2... using
 332

equation (1.1) in the form:

(1.1 a)

whereXk '" x(tk)'

Linear multistep methods for the numerical solution of ordinary differential equations make
use of formulae of the form:

(1.2)

where hk = tk - t k-l' Yk is a scalar, and If> k-I a function of past values x k" xb k' = k -1, k -2, ....
Of course, the value of Yk and the form of the function If> k-I depend on the particular formula.

The method is called explicit ifYk=O, yielding explicitly "Ie = <ilk-I' and implicit ifYk ¢ 0, in
which case (1.1 a) and (1,2) have to be solved to obtain xk An explicit method involve less
computation per step, but the step-length is limited by stability considerations, and if these dictate
a very small step-length it is worth using an implicit method, allowing a larger step. In this case,
an explicit method is often used to give an initial prediction ofxk' followed by iterative solution
of(l.la) and (1.2) to yield a corrected value. Usually Newton's method is used for the iteration,
or rather its simplified version in which the Jacobian matrix fx. (t k' x k) is held fixed and only
re-evaluated every few steps.
For initial-value problems, we require the solution for t H Ogiven initial values x (to) = XO.
Of course, in applying (1.2) the past consists only ofxO, Xo, so (1.2) can only contain two
parameters, and hence approximates x (tl ) only to first order. As we generate further points in the
sequence, we have more information at our disposal and can use a formula giving a higher order
of approximation. Multistep methods in use today have reached a high degree of refinement,
incorporating automatic error estimation and approximate optimization to choose the order of
approximation of the formula, the step-length, and the frequency of re-evaluation of the Jacobian
matrix.

Runge-Kutta methods make use of several evaluations ofX(t) from (1.1) per step, but the
same general principles apply. For fuller details of both kinds of method, the reader is referred to
the excellent treatise by Butcher [3].
 333

1.2 Systems of Differential-Algebraic Equations (DAEs)

To illustrate the use of such standard methods, let us consider the following simple example:

Example 1 - Stirred tank: batch reactor.

We consider a perfectly stirred tank: in which the simple endothermic reactions:

take place. The reactor is heated by a coil in which steam condenses (and the condensate is
removed as it is formed via a steam-trap). The instantaneous steam flow-rate is Fs and the'
condensation temperature Ts. The contents of the reactor are at temperature T, with molar
concentrations a, b, c of components A, B, e respectively, and total volume V. Both reactions are
first order, with rate expressions
(1.3)
where R is the gas constant and klO' E 1, k20' and E2 are given constants.
If A, B, e represent the total amounts of the corresponding components present in the reactor,
we have immediately:
A=Va, B=Vb, e=vc. (1.4)
The dynamic molar balances on each component are then given by:

(1.5)

while the energy balance yields

(1.6)

where!!. Hs (Ts ) is the latent heat of condensation of the steam at T s ' S is the heat-transfer
surface area of the coil, U the overall heat-transfer coefficient (assumed constant), and H is the
total heat content of the material in the reactor, given by:

H = hA A + hB B + he C (1.7)
 334

where hA, h B, and he are partial molar enthalpies.

We also have

(I.8)

where vA, vB' Vc, are partial molar volumes, and these partial molar quantities are given by

(1.9)

This example shows how mixed differential-algebraic systems naturally arise in modelling the
dynamics of chemical processes. Material and energy balances usually give rise to first-order
differential equations in time, while volume relations, chemical kinetic expressions and physical
property relations give rise to algebraic relations between the instantaneous values of the variables.
To use a standard ODE integration method to solve these equations, we need to convert the
equations to the form of (1.1), which implies eliminating the variables whose derivatives do not
appear in the equations (henceforth called the "algebraic variables"), leaving relations between the
others (called the "differential variables").
We could carry out this elimination symbolically, but it is equivalent to do it numerically at
each step, using the following algorithm:
Given A, B, C, H:
1. Solve (1.4), (1.7), (1.8), (1.9) for T, V, a, b, c.
2. Compute kl, k2 from (1.3) and Fs from (1.6)
3. Compute A, :8, C, Ii from (1.5) and (1.6).
For the initial conditions, we would be given the initial contents of the reactor, hence T(O),
V(O), a (0), b (0), c(O), from which A, B, C, H are easily calculated using (1.4), (1.7) and (1.9),
and remaining variables as in steps 2 and 3 above.
The approach used in this example can be generalized to deal with any differential-algebraic
system in the semi-explicit form:
x( t)= f (t, x (t), y (t», (1.10)
0= g (t, x (t), y (t», (1.11)
where x (t) E Rn represents the differential variables, y (t) E Rm the algebraic variables, f R x Rn
x Rm - Rn, g: R x Rn x Rm - Rm, and (1.11) can be solved for y (t) in terms ·oft and x (t).
 335

It will not in general be possible to solve (1.11) analytically, but (as in the above example) it
can be solved numerically. If g (t, x, y) is nonlinear in y we shall need an iterative solution, and if
Newton's method is used, we shall need to evaluate the Jacobian matrix 8y (t, x, y).
Alternatively, we note that since (1.11) holds at every time t, then its total derivative with
respect to time must also be zero, yielding
o= gt + 8x. +8y. y, (1.12)
where for simplicity, we have dropped the arguments, which are the values oft, x (t), Y(t) on the
solution trajectory. If the Jacobian matrix 8y of y in (1.12) is nonsingular, the equation can be
solved for y (t), and we can eliminate * (t) using (1.10) to yield symbolically:
y = - [8yr 1 [gt + 8x f], (1.13)
though of course we would normally carry out these operations numerically. Equations (1.10) and
(1.13) now give an ODE system in (x, y), to which standard integration methods can be applied.
The most general form of differential-algebraic system is:
f(t,*(t), x (t), y(t» = 0, (1.14)
where again t € R, X (t) € Rn, y (t) € Rm,
but nowf: Rx Rn xRnx Rm _ Rn+m,
and in general f(·) can be nonlinear in all its arguments.
However, if (1.14) can be solved for * (t), y (t), given t and x (t), we can again use a standard
ODE integration method, with the y (t) variables being obtained as a by-product of the solution
for * (t).
In all cases considered above, we have assumed that the equations can be solved for the
variables in question, but unfortunately this is often not the case, as illustrated by next example:

Example 2 - Continuous Stirred-Tank: Reactor.

We again consider the system described in Example 1, but this time there is a continuous feed,
with volumetric flow-rate Fl, temperature r> and composition aD, bO, cO, and continuous product
withdrawal with flow-rate P, with of course temperature and composition identical to those of the
tank: contents. The describing equations are now:
k, = klO exp (-EIIRT), k2 = k20 exp (-E2IRT) (1.15)
A=Va, B=Vb, C=Vc, H=Vh (1.16)

. °°
A=F a -Pa- Vkl a (1.17)

. °b°-Pb+Vk,a+Vk2 b
B=F (1.18)
 336

C=FO c °-Pc + Vk2 (1.19)

Ii = F °h °-Ph + US (Ts - T) (1.20)
FsAHs (Ts) =US(Ts - T) (1.21)
vA A +vBB +vC C =V (1.22)
hAa + hBb+ hCc = h (1.23)
vA = vA (T, a, b, c), hA = hA (T, a, b, c), .... (1.24)
If the feed~, ~, aO, bO, c~ and product flow (P) are given as functions of time, there is no
difficulty in evaluating all the remaining variables in terms of A, B, C, H, using the scheme given
in Example 1.
However, ifP is adjusted to maintain the hold-up V constant at a given value, it is impossible
to determine P or the derivatives A, S, C, Ii from the above equations. This is evident from the
fact that P, an algebraic variable, appears only in the equations which determine the derivatives.
Nevertheless, it is clear on physical grounds that the system is well defined!
Closer examination reveals that A, B, C, H are related through (1.22), so their derivatives
must also be related, and the required equation is obtained by differentiating (1.22). Substitution
of A, S, C from (1.17), (1.18), (1.19) will then yield on algebraic equation, which can be solved
for P. The values of the remaining variables can then be obtained as before.
This example shows that locally unique solutions can exist, even if the Jacobian matrix for the
system is singular, and also that further information is implicit in the requirement that the algebraic
equations must be satisfied at every instant of time - hence differentiating them provides further
relevant equations.
The next example shows that a single differentiation may not be enough.
Example 3 - The Pendulum
We consider a pendulum consisting of a small bob suspended by a string of negligible weight.
We shall use Cartesian coordinates, as indicated in Figure I, and assume that the string is of unit
length and the bob of unit weight (Le. mg = I).
The describing equations are:
x =u, y =v (1.25)
v=l-zy (1.26)
(1.27)
Here, we have 4 differential variables (x, y, u, v) and one algebraic variable (z), which again
appears only in the equations defining the derivatives.
 337

y mg =1
x =horizontal distance from fulcrum u =horizontal velocity

y =vertical distance from the fulcrum v =vertical velocity

z =tension of the string

Figure 1. The Pendulwn

Again, it is clear that u and v are related, and by differentiating (1.27) we obtain:
xu+yv=O (1.28)
This still does not allow us to determine z, but again u and v are related, as shown by
differentiating (1.28): x u +X u + Y v + yv = 0,
and substitution from (1.25) and (1.26) yields:
z = (y + u 2 + yZ)/<1- x 2) (1.29)
Thus, two successive differentiations of (1.27) were required to obtain all the necessary
information for solution.

2. Properties of DAE Systems

Let us examine the structure and properties ofDAE systems a little more closely.
From now on, it will be more convenient not to identify separately the algebraic variables,
and we shall consider the general form:
f(t, x(t), *(t» =0 (2.1)
with t E R, X (t) E Rn
We shall also restrict ourselves to the case where fO is continuous in x (t), *(t), though not
necessarily in t, since this covers most cases of practical interest.
 338

We shall need the following formal definitions:

Definition 1 A solution of (2.1) is a function x(t) defined and continuous on [to, tf ], and
satisfying [2.1] almost everywhere on [to, tf ],.
Definition 2 A system (2.1) is regular if:
a) It has at least one solution.
b) A solution through any point (t, x) is unique.
Definition 3 The index of system (2.1) is the smallest non-negative integer m such that the
system:

f (t, x (t), x (t» = 0,

(2.2)
dr
- f (t, x (t), x (t» = 0, r = 1, 2, m, )
dt r

defines x(t) as a locally unique function of t, x (t),

Obviously, m will in general vary as t, x (t) vary, so the index is a local property.
The existence of a finite index at a point (t, x) is a necessary and suffcient condition for the
existence ofa unique solution trajectory through (t, x), and this solution can be extended so long
as the index remains bounded. Thus, a system is regular on any domain of (t, x) on which a
bounded index exists.
Of course, the index may fail to exist at a given point because derivatives of f (.) of sufficienty
high order do not exist at this point. However, such points will in general form a set of measure
zero, defining submanifolds in the full space of variables. Pathological functions for which
derivatives of a certain order are undefined over a full-dimensional region are unlikely to arise
from modeling physical systems.
It also follows from the definition that the index is invariant under nonlinear algebraic one-
to-one transformations of the variables or equations.
The following examples further illustrate some of the implications of the above:
Example 4
Consider the system:
x- 2; + 1 = 0 (2.3)
y-2z 2 +1=0 (2.4)
x 2 +r-1 =0 (2.5)
 339

We note that, z occurs only in the equation defining y so the index must be greater than one.
Differentiating (2.5) and substituting from (2.3) and (2.4) we obtain:
x (21 - 1) + Y(2 z2 - I) = 0 (2.6)
This determines z in terms ofx and y (though not uniquely), but we must differentiate again to
obtain an equation fori: :
(2y 2 _ 1 )2 + (2 z2 - 1) 2 + 4xy (2 i - I ) 4 Yz i:= 0 (2.7)

Now, it can be shown that neither y=0 or z = 0 is consistent with (2.5), (2.6) and (2.7), so
(2.7) uniquely defines i:, and the index is two for all x, y, z. It follows that the system is regular.
However, if we set x = ±l / {i, it follows from (2.5) that y = ±l / {i and from (2.6) that
z = ±l/ {i, and these are consistent with (2.3), (2.4) and (2.7) almost everywhere. Hence, there

is an infinity of solutions with x, y and z switching independently between + 1/ {i and -1/ {i

arbitrarily often!
Nevertheless, there is no contradiction; none of these "solutions" are a solution in the sense
of Definition 1, since they are not continuous.
Of course, there may be situations in which we are interested in such discontinuous solutions,
but we must realize that they fall outside the scope of the standard theory.

Examle 5
Consider the system:

xI = x I + x2 - y (2.8)

)(2 =x 1 + x 2 - Y (2.9)

x I =2 x 2 +u =0 (2.10)
where u(t) is a piece-wise constant control function (i.e. u (t) = 0 almost everywhere).
Differentiating (2.10) and substituting from (2.8) and (2.9):
3xl-x2-y=0, a.e. (2.11)
Now suppose that u (t) = 1
u (t) = 2
and x 1 =1 at t= t 1-
Then, at t = t; we have x2 = -1 from (2.10) and y = 4 from (2.11).

But, what are the values of XI, x2, y at t 1 ? From (2.10), it is clear that XI or x or both must
have a jump at t 1, but nothing more can be deduced from the above, and the solution for t2 ~ t 1
is not uniquely defined.

Nevertheless, the system satisfies all the conditions for (2.1), and clearly has an index of two
 340

everywhere, so it is regular. Again, however, there is no contradiction, because a solution in the

sense of Definition 1 cannot exist at times where u jumps.
However, we are often interested in solving optimal control problems for which jumps in the
control are allowed, and hence in solutions which may be discontinuous at such points.
In a real physical situation, we would in fact have infonnation on the behaviour of physical
quantities in the presence of discontinuities in the controls. For example, ifwe have a tank of
liquid, and the control u represents the rate of addition of hot feed, there cannot be a jump in the
temperature of the tank contents. On the other hand, if u represents an instantaneous addition of
a finite quantity of feed (an "impulse" of feed), there will be a corresponding jump in the
temperature of the contents, (obtained by energy balance).
The most important point to note is that specification of behaviour at points of discontinuity
is a part of the model, and is not implicit in the DAE system itself. We obtain the complete solution
by piecing together segments on which the DAE system does have a solution in the sense of
Definition 1, using these "junction conditions· for the purpose.
Example 6
Consider the system:
x+x2 +yz -1 =0 (2.12)
=0 (2.13)
=0 (2.14)
Again, we must differentiate (2.14) and substitute from (2.12) and (2.13) to obtain an equation
forz:
x (1- x2 - yz) + y (xz - xy) = O.
Thus, using (2.14):
xy(y-z)+yx(z-y)=O,
and the equation is satisfied identically!
Obviously, this will remain true for further differentiations, so we can obtain no more
information not already implicit in the original fonnulation (2.12) - (2.14). However this is not
enough to determine z(t), or even z (t), so the index is not defined.
In fact, we are free to choose any arbitrary function z (t), and any initial values for x and y
consistent with (2.14), and x (t), y (t) will then be uniquely defined.
This example shows the dangers of assuming that a system has a finite index, and hence
inferring that it is regular.
 341

Having given warnings of possible pitfalls, it may now be helpful to discuss the structure and
properties oflinear, constant-coefficient DAE systems, which have the general form:
A. x (t) + B.x (t) = c (t), (2.15)
where A and Bare n x n matrices and x(t), x(t), c(t) are n-vectors.
The analysis of such systems dates back to Kronecker, and an excellent detailed analysis of
the general case will be found in Gantmacher [5]. Here, we will treat only the regular case, where
regularity of the system in the sense of Definition 2 coincides with regularity of the matrix pencil
[A + AB], where Ais a scalar.
Such a pencil is said to be regular if det IA + ABI is not identically zero for all A; otherwise
it is said to be singular. It then follows that for a regular pencil, IA + ABI is nonsingular except
when A coincides with one of the n roots of the equation:
det IA + ABI = O. (2.16)
For a regular pencillA + ABI, there exist nonsingular n x n matrices P and Q such that

o o

P [A + A B] Q = + A (2.17)

where

1 ...
Ir = , Nr = (2.18)

and If' Nr , r = 0, 1, .. m, are ~ x ~ matrices.

The index of nil potency of the pencil is max nr
r

Now, to relate this to the system (2.15), let us define:

x (t) = Q. z(t) c(t) = P .c(t) (2.19)
Then from (2.15), (2.17), (2.18) and (2.19):
 I
342

Zo + BOZO = Co
(2.20)
Nr Z + Zr = e;:, r = 1, 2, ... m

and for a typical N r :

I
. (i+ll (il
Z + zr = :c....r(il' i = 1, ... (nr .... 1)
(2.21)

from which we deduce:

(2.22)

From (2.20) and (2.22) we see that the system completely decomposes into (m+l) non-
interacting subsystems. The first of these (r = 0) is a standard linear constant-coefficient ODE,
explicit in Zo ' whose general solution involves no arbitrary constants. The other m systems contain
no arbitrary constants, and the 2T are obtained from the RHS vector and successive differentiations
of it.
Thus, to obtain Zr we need fie differentiations ofe;: and it is clear that the index of the DAE
system (as given in Definition 3) is equal to the index of nilpotency of the matrix pencil [A + AB].
This analysis clearly shows that the system index is a property of the system equations, and
independent of the specification of the boundary conditions, but caution is necessary here, as the
next example shows:
Example 7. Flow through Tanks.
We consider the flow of a binary mixture through a sequence of n perfectly stirred tanks, as

FigureZ
 343

illustrated in Figure 2. The molar concentrations of the components leaving tank i are llj , bj , with
volumetric flow-rate Fj while the feed to the first tank has constant flow-rate FO and time-varying
composition ao, bOo For simplicity, we assume that all the tanks have the same volume, V.
The describing equations are:

d3.j
Vdt- = F·1- 1~:, - 1-F·~:
I,
1,2, ... n, (2.23)

dbi
Vdt- = F·1- lb·1- 1-F·b·
1 1
1, 2, ... n, (2.24)

i = 1, 2, ... n, (2.25)

where va' \j, are partial molar volumes of the two components, here assumed constant for
simplicity.
As in Example 2, equation (2.25) relates the two differential variables, so we differentiate it
to yield:

Then substituting from (2.23) and (2.24) and using (2.25) yields:
i = 1, 2, ... n.
Since all flows are equal to FO' which is constant, we can transform the time variable:
t=FOtlV, (2.26)
whereupon (2.23) and (2.24) can be written:

ai = ai _1 ai
}
-
1,2, ... n (2.27)
hi = bi _1 - bi

were ~ denotes dllj/dt etc., and we see that the system decomposes into two independent sub-
systems.
Now, if we specified the feed composition ao (t), bo (t), t~ to ' and the initial concentrations
 344

in the tanks: ~ (to), bi (to)' i = 1,2, ... n, (2.27) represents a constant-coefficient ODE system
(hence of index zero), which can be integrated from t = to .
On the other hand, if we require the feed composition to be controlled so that the outlet
concentrations ~ (t), b n (t) follow a specified trajectory for t ~ to then An (t), bn (t), are
thereby specified, and we must use (2.27) to compute ao -1, bn -1 .The same then applies to tank
(n -1), and so on recursively, until ao(t), bO(t) are computed.
In this second case, we were not free to choose the initial concentrations in the tanks, and the
solution was obtained by successive differentiations instead of integration, showing that the index
is (n+1).
Thus, the index of the system is crucially dependent on the type of boundary condition
specified!
The paradox is resolved by noting that the "system" just referred to is the physical system, and
the corresponding DAE system is different for the two cases. In the first case the, differential
variables are IIj (t), bi( t), i = 1, 2, ... n, ao< t), bO< t ) and are given driving functions, while in
the second case the variables are and the specifications a: ( t), b: ( t) are given driving functions.
From the above properties of the linear constant-coefficient system, in which regularity and
index of the DAE system coincide with regularity and index of the matrix pencil [A + AB], one
is led to suspect that the same relations might hold in the nonlinear case for the local linearization
of the DAE system. The matrix pencil in question would then be [fx + A . fx] , but unfortunately
no such correspondence exists, even in the linear time-varying case (except if the system has index
one or zero), as illustrated by the following examples:
Example 8. - Regularity and Index
a) Consider the system:

t 2.x+ t'y-y=t (2.28)

tx + Y+ x =0 (2.29)

We have: det Ifx + Afxl = det lt2

+A
'
t-A
1
=A
2
 345

Hence, the matrix pencil is regular everywhere. However, multiplying (2.29) by t and
subtracting (2.28), we obtain:
~ +y=-t
Differentiating and substituting for y in the LHS of (2.29) then yields:
tx + y+ x = -1
This contradicts (2.29), so the system is inconsistent and hence has no solution.
b) For the system of example 6 we have

I = 2Ax AZ
det Ifx + Afxl + det A(Y-Z) I+Ax AY1 = 2 (y-z)
-~
2Ax 2Ay

Thus, the pencil is regular, except along the line y = z,

Moreover, we have the factorization:

1 0, °
r
IP1X, 0, 1I2IP1'
-IP1XZ
2x, 2y, -X(II+71Y [ fx + Afx] 0, 1/2IP1Y, 1I2y =x
0, 0, 1I2IPI xy, IP2' °
+ A

where IPI = (y - z) (y/x + xl';) , IP2 = (z - zy) / z IPI';

This factorization is well defined, and the factors are nonsingular, ify "# Z and x, yare non-zero,
so that the index of nilpotency is three under these conditions.
However, we saw in Example 6 that the DAE index is not defined, and there is an infinity of
solutions through each point, so the system is not regular in the sense of Definition 2.
c) Consider the system:
*-ty=t (2.30)
x-ty=O (2.31)
 346

We have:det Ifx + ).J~ = det ~11 -t

-At
I = 0 aliI...
Hence, the the pencil is singular everywhere, and there is no factorization of the form of
(2.17), so the index of nilpotency is not defined.
On the other hand, differentiating (2.31) and substituting for x from (2.30) yields
x = ty
y = ty + t.
+

Whence y = t and x = t2. Thus, the DAE index is two, and we have a unique solution.

3. Reformulation of High-index Problems

The complicated properties ofDAE systems with index greater than one (commonly referred to
as "high index" systems) and the difficulties associated with their solution has led some engineers
[9] to assert that high index models are in some sense ill-posed, and hence avoidable by "proper"
process modelling.
Certainly systems for which the index is undefined over a full-dimensional region (such as that
in Example 6) are usually functionally singular, indicating redundant equations and an
under-determined system. However, we have seen that a system is regular and well-behaved so
long as a bounded index exists, and as we shall see later, such a system is reducible to an index-one
system, so there is nothing "improper" about such systems.
The natural way of modelling the system may well be the high-index form, and the
mathematical reduction procedure may destroy structure and lead to a system which is not easily
interpreted in physical terms.
It is however always useful to investigate how the high index arises. In general, it is because
something is assumed to respond instantaneously, so this focusses on underlying assumptions, for
which the relative advantages and disadvantages can then be assessed.
For example, high index in Example 2 arises because we assumed that the product flow-rate
can be instantaneously adjusted to maintain the hold-up V exacty constant. In practice, this may
be achieved by overflow, or more often by adjusting a control valve in the product pipe to maintain
the level constant.
We could instead model this controller as a proportional-integral-derivative (PID) controller
 347

p - p* = Kp (V - V*) + KI I + KD ~ (V - V*)!
dI .. (3.1)
-=v-v
dt
where V· is the desired (set-point) valve of V, and p. arises from setting up the controller - for
example:
At t = to: P (to) = p., V (to) = V·, I (to) = 0 (3.2)
Then, ifwe use no derivative action (KD = 0) we find that the expanded model now has index
one.
This is certainly a more realistic model, not significantly more complicated than the original,
though we do have to choose appropriate values for the parameters Kp, KI. Note, however that
if we add derivative action we would again have an index two model!
Example 9
A further example is afforded by modelling the dynamics of the steam heating-coil of the
above reactor, replacing the instantaneous steady-state model of equation (l.21). The equations
are (see [10]):
Vs dE. = Fs - L (3.3)
dt

(3.4)

Ps = R Ts Ps = PexpCfrrs). (3.5)
where the inlet steam is superheated vapour at temperature Tso, assumed to obey the ideal gas law
and a simple vapour-pressure equation, with constant specific heat Cs and latent heat M~. The
volume of the coil is Vs and the condensate, removed as it is formed via a steam-trap, has a
flow-rate L.
L appears only in the differential equations, so (3.5) must be differentiated, showing that this
sub-system has index two.
Again, the high index arises from an idealized control system, for the steam-trap is in reality
a level controller. However, to model it more realistically we should have to introduce a
condensate phase with its own balance equations, as well as the controller relations, and one might
then feel that this is more complicated than the differentiation of(3.5).
Alternatively, we could look for further simplification, for example by neglecting the variation
of vapour hold-up (not the hold-up itselt), setting dp/dt = O. Then, L = Fs and the system has
index one.
 348

Thus, we have two contrasting approaches. The first is to identifY instantaneous responses and
then model their dynamics more accurately. The second is to identify an algebraic variable which
appears only in differential equations, and to convert one of these to an algebraic equation by
neglecting the derivative term.
Another technique for obtaining lower index models, particularly useful in considering
mechanical systems, is to look for invariants of the motion (e.g. energy, momentum or angular
momentum), then transform the variables to include these invariants as independent variables. The
pendulum considered in Example 3 nicely illustrates this:
Example 10 - The Pendulum revisited.
If we use radial rather than Cartesian coordinates, we can make direct use of the fact that the
pendulum is of constant length, hence avoiding the need for the algebraic relation (1.27).
Using the same assumptions as before, the system can now be modelled in three terms of only
variables: the angle to the vertical (8 ), the velocity of the bob (V), and again the tension in the
string (z), yielding:

e=v (3.6)
e = -sin8 (3.7)
z =.,2 + cos 8 (3.8)
This is clearly an index one system!
It would in fact seem that we have a counter-example to our assertion in Section 2 that the
index is invariant under a one-to-one algebraic transformation, since the transformation from
Cartesian to radial coordinates is just such a transformation:

x = r sin 8, y = cos 8 )
(3.9)
u = V sin q>, v = V cos q>

t = V cos (q> - 8), r 8 = V sin (q> - 8)

V = cos q> - Zr cos (q> - 8) , V <p = Zr sin (q> - 8) (3.10)

r = ±1
 349

However, direct application of(3.9) to (1.25), (1.28), (1.27) yields the system:
This still has an index of three, and two differentiations are required to reduce it to (3.6) - (3.8).
Whilst such reformulations are useful, and occasionally instructive, it is obvious that we need
more systematic methods of solving high-index systems, and we turn to this problem in the next
section.

4. The Solution of DAE Systems

Gear and Petzold [7] were the first to propose a systematic method of dealing with high-index
systems, and we have in fact used their method in solving the problems in Examples 2-6.
The method involves successive stages of algebraic manipulation and differentiation, starting
from the general system (2.1), and at each stage the system is reduced to the form:

(4.1)

where is ( xp Yr) a partition ofX .

The algorithm can be formally stated as follows:
Gear-Petzold Index-Reduction Algorithm
O. Set: r= 0, Yo = x x =0 '}
fOO=f(o) XoO= 0 .

1. Solve frCt, x, Yr ) for as many derivatives as possible, and eliminate these

}
equations to yield: x'r+l = Xr +1 (t, x, Yr+l) (402)
o = 11r+1Ct,x)
where (x'r+l, Yr+l) is a partition of Yr
2. Form xr+l = (xr +l, x'r+l,)
Substitute xr+1 = Xr+l(t, x, Yr+l) into XrCt, x, Yr)and
append Xr+l(t, X, Yr+l) to form Xr+l (t, x, Yr+l)
3. IfYr+l = 0, STOP.
4. Differentiate 11r+1 (t,x) with respect to t, then substitute xr+l = Xr+l (t, x, Yr+l) to
form fr+l (t, x, Yr+l)
5. Set r: = r+1 and return to step 1.
 350

On tennination, we have the explicit O.D.E. system:

x=X (t, x). (4.3)
Of course, we need appropriate initial conditions for this system, and these need to satisfY the
algebraic equations 11r+I (t,x), r = 0, 1,2, ... generated by the algorithm.
We illustrate this algorithm by applying it to the system of Example 7:
Example 11. Gear-Petzold Index Reduction.

}
Given: ~=~-I-~ I = I,2, ... n,
(4.4)
~ = an*(t)
r=0 The system is already in solved from for all the derivatives in the equations, so we
have Yt = ao, xl = x'I = (aI, a2, ... , An)·
Differentiating the algebraic equation and substituting for derivatives yields
. .*
~ = ~ - 1 - ~ = an
r= 1 The form of the system is unchanged, with Yz = YJ, x2 = xI' Differentiation
and substitution yield:
~ - 2 - 2~ - 1 + ~ = an
.*
r=n Differentiation and substitution yields

(4.5)

r=n+I Equation (4.5) is in solved form, so we append it to the ODEs in (4.4) to yield
the final system, and since Yr+1 = 0 we stop.
In this case, the algebraic equations are of the general form:

~ (-1)
i:O i( r)i ~-I = (an*)(f)' r=O, I, ... n, (4.6)

and are sufficient to detennine uniquely all ~(tO)' i = 0, 1, ... n.

We note that the system of ODEs contains only the (n+ 1) - derivative (a *in+ 1) , so that
n
changing an* by an arbitrary polynomial of order n:
 351

would not change the ODEs. Of course, choosing initial conditions consistent with (4.6) would
then ensure that the coefficients Co> cl, ... cn were all zero, but in practical computation rounding
and truncation errors may cause non-zero ci, and an error in Cn would be multiplied by iu, causing
rapid error growth.
Brenan et aI [2] discuss this problem of "drift" in satisfYing the algebraic equations TJr +1 (t, x),
r = 0, 1, 2, ... and report on various proposals in the literature for stabilizing system (4.3). The
simplest and most effective proposal was made by Bachmann et aI. [1], who noted that the
algebraic equations in (4.2) should be retained instead of differential equations. Since by
construction these have maximum rank, they determine a corresponding subset y of the elements
ofx in terms of the remainder, z.
The equations defining the derivatives of these y-variables in (4.3) can then be dropped,
leaving a semi-explicit index-one system of the form:

i:(t) = X(t,x) )
(4.7)
o = G(t,x)

where G (t, x) represents the union of the algebraic equations TJr+ 1(t, x), r = 0, 1, ....
Consistent initial conditions are obtained simply by assigning values for z (to)'
In fact, Bachmann et aI. described their algorithm only for linear systems, simply stating that
"the algorithm can also be used for nonlinear systems" without giving details. Chung and
Westerberg [4] describe a very similar conceptual algorithm for index reduction, though they then
propose solution of the reduced index-one problem, rather than retaining algebraic equations.
Although the proposal of Bachmann et al deals satisfactorily with the stability problem, and
the problem of identitying the means of specitying consistent initial conditions, there is still the
fundamental weakness that step 1 of the algorithm requires symbolic algebraic manipulation.
One consequence of this is that only a limited class of systems can be treated. Since one
differentiation occurs (in step 4) on each iteration, and an ODE system is finally generated, it is
clear that the index of the system is equal to the number ofiterations.
However, since (4.3) was generated symbolically, the reduction is true over the whole domain
of definition of f (t, X, ic), implying that the index must be constant over this domain. We cannot
therefore treat systems whose index varies over the domain.
More generally, the explicit solution of the form (4.2) cannot always be generated
 352

symbolically, and in any case symbolic manipulation is a significant task which is not easily
automated.
These problems are circumvented by the numerical algorithm put forward very recently by
Pantelides et aI. [11]. Here, the index-reduction procedure is carried out numerically during the
integration, each time that the Jacobian matrix is re-evaluated.
Thus, given fx (t, x , x) we start by obtaining the factorization:

(4.8)

where ~ is a truncated unit upper-triangular matrix, Po is a permutation matrix and QO is either

orthogonal or generated by Gaussian elimination using row and column interchanges.
We then define:

(4.9)

where it should be noted that f 1 should not be interpreted as df 1/ dt since QO may itself be a
nonlinear function of (t, x, x).
Now, ~ E R rl , where rl is the rank offx ' and if this remains constant over a sub-domain
of, (t, X, x) then (4.9) defines a function g10 E R n -rl with ~ = 0 over the sub-domain, and
hence glO may be written gl (t, x), independent of X. In general, the rank r, and hence the
dimension of g'O, may vary from point to point but from the continuous differentiability
assumption this will happen only on a set of measure zero.
If fx is nonsingular, there are no null rows in (4.8), and we have immediately:

which can be used in a Newton iteration to solve (2.l) for x(t), recalling that fi is now unit
upper-triangular.
Otherwise, we form and factorize:
 353

Again g; = 0, and with the assumption of continuous differentiability g2( t, x) is independent

of x almost everywhere.
Again, if r; is nonsingular we have

and we can use this to solve (2.1).

The recursion can he continued with the general form:

(4.10)

with (g (f) )r + 1 independent of x.

Eventually, when r =m (the index) we shall havef;; nonsingular and:

fro ox = - fm. (4.11)

As before, the algebraic equations (g (f)) r d = 0, r = 0,1, .. ffi, can be collected to form the system:
G(t,x) = 0 (4.12)
The Jacobian matrix Gx of this system can be factorized to yield:

(4.13)

where Gy is unit upper-triangular and (y, z) is a partition ofx. Given z, we can then use:

Gy· oy = -Q.G. (4.14)

As in the Bachmann algorithm, we use the integration formula (cf(1.2):

z = yh.i: + <p (4.15)

 354

to detennine new values of z.

The algorithm is then:
Pantelides - Sargent Algorithm
Given an estimate ofx and x:
o (0) 0
1. Set r =0, fx = fx ' (gx) = '"
2. Compute Qp Pp r;+I, [fr+l,--<g(f)r+l] using (4-10)
3. If(g(f) r+l = 0, go to step 6.
4. Add (g(l)) r+l to G and(g~r~r+l to Gx
5. Set r:=r+1 and return to step 2.
6. Compute x := x + ox using (4.11)
7. Factorize Gx using (4.13)
8. Compute y := y + fJy using (4.14)
9. Compute z from (4.15)
10. Repeat from step 1 to convergence.

Again, as in the Bachmann case, consistent initial conditions are obtained by specifying only zeta).
In fact, this can be generalized to specifying the same number of algebraic initial conditions:

J(to, x(to), x(to)) = 0, (4. 16)

such that the Jacobian matrix:

(4.17)

is non-singular. Equations (4.16) then replace equations (4.15) in step 9 when the algorithm is
used to evaluate initial conditions.
As in standard ODE practice, the Jacobians f;, Gx need not be recomputed at each
iteration, nor even at each integration step, and we note that the integration formula can be explicit
or implicit.
As described above, it seems as if a numerical rank determination is implied in the factorization
in (4.10). However, in practice the factorization can be discontinued as soon as the remaining
elements fall below a suitable threshold value. Additional differentiations caused by a finite
 355

threshold do not invalidate the algorithm, and are even beneficial in dealing with the
ill-conditioning. For the same reason, changes in the rank. or index cause no difficulties.
The algorithm can be implemented using automatic differentiation for the requisite
differentiations of f (-), and otherwise only standard numerical linear algebra is involved. Of
course, sparse matrix techniques can be used to deal with large-scale problems, and in principle
the algorithm will deal with any systems with bounded index. In practice, the index may be limited
by the limitations of the automatic differentiation package in generating successively higher-order
derivatives.
As shown by Pantelides et aI. [10], high index problems are common in chemical engineering,
and multistage countercurrent systems can in some circumstances give rise to very high index
systems. It is not therefore surprising that there has been a search for direct solution methods
which do not involve the successive differentiations required in index-reduction algorithms.
The earliest such technique was proposed by Gear [6]. He noted that the general implicit linear
multistep integration formula (1.2) can be solved for xkin terms ofxk and past data, and the result
substituted into (2.1) to yield a set of n equations in the n unknowns xk. The solution could then
be used in conjunction with standard ODE techniques for optimal choice of order and step-length
for the formula in question.
If the corrector formula is solved by Newton's method, we require nonsingularity of the
corresponding Jacobian matrix [fx + yhk' fx] , and this also guarantees uniqueness of the solution
generated. However we note that there will be such a solution for any initial condition, whereas
we have seen that true solutions must satisfy a set of algebraic equations. Clearly, we must start
with consistent initial values, but the question then arises of whether the generated solution
remains consistent. It has in fact been shown (see [6]) that for systems of index zero or one, and
semi-explicit systems of index two, the order of accuracy of the solution is maintained so long as
the initial conditions are consistent to the order of accuracy of the integrator formula, and at each
step the equations are solved to the same order. However, there are counter-examples for higher
index systems, and even for general index-two systems.
It will have been noted that the underlying Jacobian matrix is of the same form as the matrix

pencil arising from the local linearization of the equations, so that regularity of this pencil ensures
non singularity of the Jacobian for all but a finite set of values ofhk- Unfortunately, as we have
seen, regularity of the pencil has no connection with regularity of the DAB system, so we cannot
conclude that the Jacobian will be nonsingular if the DAB system is regular. Moreover, since by
 356

definition the matrix fx is singular for all DAE systems with index greater than zero, the condition
number for [fx + yhk' fx] tends to infinity as hk - O. In fact, it can be shown that the condition
number is O[h-fi ], where m is the index. This is unfortunate, since we rely on reducing hk to
achieve the specified integration accuracy, but the solution then becomes increasingly sensitive to
rounding errors.
This problem was studied by Bachmann et aI.[I], and the table below is taken from their
numerical results for the solution of the system (cf Example 7)

ilj=~-1 - ~, i=l, 2, ... n }

(4.18)
au = l-exp( -t12), O~t~ I.

h 1.0 0.1 0.01 0.001 0.0001

n
4 0.958 0.371 E - 2 0.378 E - 3 0.164 E - 4 0.438
8 0.154 E +2 0.441E-3 0.412E+l 0.368 E + 9 0.326 E+17
12 0.246 E +3 0.141E+ll 0.168E+1O 0.990 E + 22 *
16 0.394 E +4 0.423 E + 17 0.170 E +19 * *
* mdlcates fllliure With a "smgular' matnx.
The table gives the error in ao at t = 1 for the integration of this system using the implicit
Euler method, starting with the analytically derived consistent initial values at t = O.
Even for n = 4, the accuracy attainable is limited by the above ill-conditioning, and the system is
violently unstable for higher values of n. Of course, the results could be improved by use of a
higher-order fonnula, but it is clear that the process is intrinsically unstable.
It seems that the higher derivative infonnation is essential to provide reliable predictions, but
usually we are interested in behaviour over an extended period and this prediction problem would
be avoided by a global approximation over the whole period of interest, as is the case for
two-point boundary value problems. Thus, we might expect satisfactory results from use of
collocation or finite-element techniques. However, there seems to be no published work on an
analysis of these techniques as applied to high-index DAEs.
An approach in the same spirit has recently been proposed by Jarvis and Pantelides [8], in
which the integration over a fixed time interval is converted into an optimal control problem.
Here, the system is taken as described by (1.14):
 357

f(t, *(t), x(t), y(t)) =0 (4.19)

and at each step we linearize the system and carry out a factorization of the matrix [fx, fy] as in
(4.8):

(4.20)

where again Un is unit upper-triangular, and the vector (x, y) has been partitioned into [zl' z2].
The linearized equation is then:

(4.21)

Of course, for consistency we must have f2 '" 0 within rounding errors, and then the iteration:
Un tiz 1 = - fl (4.22)
until II fl II S E effectively solves a subset of(4.19) for zl' given zi

(4.23)

Thus, ifz 2 (t) is a known function, we can use (4.23) to integrate over the time interval of
interest, say [t o,t f]. Hence, we treat z2 (t) as a control function, and choose it to minimize:

(4.24)

subject to satisfying (4.23) at each t.

This problem can be solved by choosing a suitable parametrization ofz2 (t):
z2(t)=1jI (t, p), (4.25)
where p is the set of parameters, which converts the optimal control problem into a nonlinear
programming problem in p.
Again of course the factorization (4.20) need not be performed at every step, as in standard
ODE practice. In fact, we still have the problem of consistent initialization, since as we have seen,
 358

the initial x (to)' y(tO) must satisfY a set of algebraic equations, but these can be correctly
established by using the Pantelides-Sargent algorithm described earlier. The advantage of the
optimal control approach is that this need only be done at the initial point, or at subsequent points
at which discontinuities occur.
A number of high-index problems have been successfully solved using this approach, including
the pendulum problem (Example 3) and versions of the canonical linear problem described below
with index up to 12.
The analysis of this last problem is instructive in indicating the requirements and limitations
of these various methods:
Example 12
The canonical linear system (2.21) yields the system:

Xi = xi.1 i = I, 2, ... ,(n - 1), } (4.26)

xI (t) = cos t

The advantage of using cos t as the driving function is that the solution is bounded by ± I and
the analytical solution is immediately available:
Forn=6:
x I = cost =1- t2 12! + t4/4! + 0 (t 6)
x2=sint=t+t 3/3!-t 5/S!+0(t 7 )
x 3 = - cos t = -I + t 2/2! - t 4/4! + 0 (t 6)
x4 = sint=t-t 3 13! +t 5/S! +0(t7)
x 5 = cost = I _t 2/2! + t4/4! + 0 (t 6 )
x 6 = - sin t = -t + t 3/3! - t 5/S! + 0 (t 7)
The partition given by the factorization is
z I = [XI, x2' x 3, x4' xS]
As an illustration, we use the simple parametrization:
x
x6(t) = 6, all t.
In this case, there are no degrees of freedom in the initial conditions, so all initial values other than
x I are unknown parameters:
x2(O) = x2, x3(O) = x3, x4(O) = x4, xS(O) = Xs
Thus, the optimal control problem is:
 359

. If
nunj
X h (1:) - COS I
1: dt
o

subject to the differential equations in (4.26).

Again, for given parameters we can solve the system analytically:
xl= 1 + x2 t +x3 t 2 /2! + x4 t 3 /3! + x5 t4/4! +x6 t 5/5!
- - -2 -3-4
x2= x 2 + x 3 t + x4 t 12!+x5 t /3! + x6 t 14!
- - - 2 - 3
x3=x3+x4t+x5t 12!+x6t /3!
- - - 2
x4 = x4 + x 5 t + x6 t 12!
x5 = x5 + x6 t
x6 = x6
Comparing these expressions with the analytical solution, we ought to have x6 = o.
To determine this value from evaluation off 0 in a local integration formula would require
the determination of {xl(t) - cos t} to an accuracy of at least O[h 5]. However this error
propagates, and to obtain x 6 = 0 from the optimal control problem requires evaluation of the
integral only to O[tf 5].
Of course, in either case we rapidly lose accuracy in the variables representing higher
derivatives, and only methods which directly use higher derivative information, like the
index-reduction methods, can obtain the requisite accuracy in all the variables.

5. Conclusions

Lumped-parameter dynamic models for chemical process systems usually give rise to a system of
DAEs, and the same is true of distributed-parameter systems solved by discretization of the space
variables. These systems often have an index of two or three, and higher index systems can arise,
particularly in modelling behaviour under abnormal conditions (eg. during start-up or under
potentially hazardous conditions).
Although physical insights and reformulation can often be useful in reducing the index, this
is likely to be the preserve of expensive specialists, and as the size of systems being studied grows,
this approach becomes more and more time-consuming, and less and less effective. It is therefore
important to have reliable and accurate general-purpose numerical methods for solving such
 360

systems, which can be very large-scale, involving possibly hundreds of thousands of variables.
This paper has attempted to describe the essential nature of the problem and some of the
difficulties which arise, and to review the current state of the art in techniques for solving these
problems.
In order to capture all facets of the behaviour, as represented by the model, there seems to be
no alternative but to use a method which makes use of the full system defining the index (2.2),
such as the index-reduction methods, and we are just beginning to see the emergence of effective
numerical algorithms in this area, such as the Pantelides-Sargent algorithm.
Particularly in large-scale systems there is often only a small proportion of the variables which
are of detailed interest, and if their behaviour does not depend on higher derivative information
it may be acceptable to use methods which use only the DAB system itself (2.1). Some analysis
is required to establish the possibilities, but this is provided by using a full algorithm, at the initial
point, which is in any case necessary for consistent initialization.
In this area, we pointed out the intrinsic instability of Gear's original method for higher index
systems, and argued that no method based on a purely local approximation is likely to be effective.
This leaves the field to global approximation methods based on collocation, finite elements, or the
hybrid-optimal control approach of Jarvis and Pantelides. We can look forward to more results
in this area, particularly in optimal control applications and other applications giving rise to
distributed boundary value problems.

References
I. Bachmann, R, L. Brull, T. MIZigiod and U. Pallaske, "On Methods for Reducing the Index of Differential-Algebraic
Equations", Compul. Chem. Engng., 14, pp 1271-1273 (1990)
2. Brenan., K.E., S.L. Campbell and L.R Petzold, "Numerical Solution of Initial-Value Problems in
Differential-Algebraic Equations", North Holland, New York (1989)
3. Butcher, J.C., "The Numerical Ana1ysis ofOrdiruuy Differential Equations", John Wiley & Sons, (Chichester, 1987)
4. ChlDlg, Y., and A W. Westerberg, "A Proposed Numerical Algorithm for Solving Nonlinear Index Problems", Ind.
Eng. Chem. Res., 29, pp 1234-1239, (1990)
5. Gantmacher, FR, "Applications of the Theory of Matrices", Interscience, (New York, 1959)
6. Gear, C.W., " The Simultaneous Numerical Solution of Differential-Algebraic Equations", IEEE Trans. Circuit
Theory, CT -18, pp 89-95 (1971)
7. Gear, C.W., and LR Petzold, "ODE Methods for the Solution of Differential-Algebraic Systems", SIAM J.
Numer. Anal. 21, pp 716-728 (1984)
8. JaIVis, RB., and C.C. Pantelides, "A Differentiation-free Algorithm for Solving High-Index DAE Systems", Paper
146g, AIChE Annual Meeting, Miami Beach, 1-6 November, 1992
9. Marquardt, W., "Dynamic Process Simulation - Recent Progress and Future Challenges", in Y. Ar1am and W.H.
Ray (Eds), "Chemical Process Control-CPC IV", pp 131-180, AIChE (New York 1991)
10. Pantelides, C.C., D.M. Gritsis, K.R Morison and RW.H. Sargent, "The Mathematical Modelling ofTransient
Systems Using Differential-Algebraic Equations", Compul. Chem. Engng., 12, pp 449-454 (1988)
II. Pantelides, C.C., R W.H. Sargent and V.S. Vassiliadis, "Optimal Control of Multistage Systems Described by
Differential- Algebraic Equations", Paper 146h, AIChE Annual Meeting, Miami Beach, 1-6 November, 1992
Features of Discrete Event Simulation

Steven M. Clark, Girish S. Joglekar

BalCh Process Technologies, Inc., P. O. Box 2001, W. Lafayette, IN 47906, USA

Abstract: The two most important characteristics of batch and semicontinuous processes
which demand special methodolop from the simulation standpoint are the continuous
overall changes with time as well as the discrete changes in the state of the process at
specific points in time. This paper discusses general purpose combined discrete/continuous
simulation methodology with focus on its application to batch processes.

The modeling of continuous overall change involves the solution of simultaneous

differential/algebraic equations, while the modeling of discrete changes in the process state
requires algorithms to detect the discrete changes and implement the actions associated
with each discrete change. The workings of the time advance mechanism which marches
the process in time are discussed with the help of a simple batch process.

Keywords: Simulation, batchlsemicontinuous

1. Introduction

A major portion of the research and development activity over the past 25 years in the area
of the simulation of chemical processes has been targeted to steady state processes. As a
result, steady state process engineering has benefited significantly from the use of
simulation based decision suppon tools to achieve reduced capital costs and improved
control systems. Employing simulation for steady state process engineering has reached
such a level of confidence and maturity that today any quantitative process decision
without its use would be inconceivable.
 362

The batchlsemicontinuous processes significantly lag behind the steady state processes
in the availability of versatile simulation based decision support tools. The complex nature
of this mode of operation poses challenging problems from the perspective of efficient
solution methodology and information handling. The key features of batchlsemicontinuous
processes are as follows.

Batchlsemicontinuous processes are inherently dynamic in nature, that is the state of the
process changes with time. The state of a process is defined as a set of variables which
are adequate to provide the necessary description of the process at a given time. Some
variables, which describe the state, change continuously with time. For example, the
concentration of species in process vessels may change constantly due to reactions, the
level of material in vessels changes as material is withdrawn or added, the flowrate and
conditions of a vapor stream (temperature, pressure and composition) may change due to
the changes in the bulk conditions. Alternatively, the values of some variables may change
instantaneously. These discrete changes may be introduced on the start and finish of
operations and by specific operating decisions. For example, a processing vessel may be
assigned to a new operation after completing an operation, a reaction step may be
terminated when the concentration of a particular species reaches the desired value, or an
operator may be released after completing an assignment.

In a typical batchlsemicontinuous process several products are made concurrently over

a period of time, each according to a unique recipe. A recipe describes the details of the
operations associated with the manufacture of that product, for example which ingredients
to mix, the conditions for ending a particular process step, the pieces of equipment suitable
for an operation.

The key operating decisions which influence the overall performance of a process are
concerned with the assignment of equipment to an operation and the assignment of
materials and shared resources as and when needed during the course of the operation. The
materials, resources and equipment are shared by the entire process and typically have
limited availability. For example, only one operator may be available to manage 10
reactors in a process, or an intermediate storage tank may be allowed to feed material to a
maximum of two downstream mixing tanks at a time.
 363

Therefore, in addition to the ability to model the dynamics of physicaVchemical

transformations, the design of a simulator for batch processes must include the ability to
implement operating decisions.

Several simulators are available for modeling the dynamics of chemical processes [1].
These range from programs written in a procedural languages, such as FORTRAN, for
performing specific simulations to general purpose continuous simulation languages. Also,
various solution techniques have been employed to solve the underlying differential
equations, such as the sequential modular approach using Runge-Kutta, or the equation-
oriented approach using implicit integrators. However, none of these simulators have been
designed to handle discrete changes in the process state, and in some cases to handle
simultaneous algebraic equations. These shoncomings make them unsuitable for
applications in batch process simulation.

Several discrete simulators are now available, such as SLAM [9], GPSS [10], for
applications predominantly in the discrete manufacturing sector. Some of these tools have
been successfully used in simulating batch processes [5]. These simulators also provide an
executive for the combined discrete/continuous simulation. However, since these
simulators are merely general purpose simulation languages, they require that the user
program the necessary process dynamics models and the logic associated with the
processing of discrete changes. Since these simulators were designed mainly for discrete
manufacturing systems, the algorithms for solving differential equations are not only very
inefficient, but also unable to solve algebraic equations. Therefore, even though they
provide the basic solution methodology, the discrete simulators have enjoyed very limited
success in modeling batchlsemicontinuous processes.

The design of a general purpose process modeling system for batchlsemicontinuous

processes has received considerable attention over the past few years, and has resulted in
the development of simulators such as DISCO [6], UNIBATCH [4], BATCHES [3] and the
gPROMS [1].

This paper presents the methodology which is central to a combined discrete/continuous

simulator. The discussion of the methodology is based on its implementation in the
 364

BATCHES simulator. However, the concepts are applicable to any combined

diSCrete/continuous simulator. The special requirements of a general purpose batch process
simulator have been discussed in another paper [2].

2. The Time Advance Mechanism

Central to a combined discrete/continuous simulator is an algorithm, the time advance

mechanism, which marches the process being modeled in time. The four key components
of the time advance mechanism are:

1. Manipulation of the event calendar

2. Solution of the differential/algebraic system of equations

3. Detection and processing of discrete changes

4. Implementation of operating decisions

The role played by each component in the time advance mechanism will be discussed in
this section.

2.1 The Event Calendar

The events during a simulation represent discrete changes to the state of the process. The
events are of two types: time events or state events.

A time event is that discrete change whose time of occurrence is known a priori. For
example, if the recipe for a particular step requires the contents of a vessel to be mixed for
1.0 hour, then the end-of-mixing event can be scheduled 1.0 hour after the start of mixing.
Of course, the state of the process will usually change when an operation ends.

A state event occurs when a linear combination of state variables crosses a value, the
threshold, in a certain direction. An example of a state event is given in Figure 1. The
exact time of occurrence of a state event is not known a priori. For example, suppose the
recipe for a particular step requires the contents of a vessel to be heated to 400 K.
 365

Therefore. the time at which the contents of the vessel performing that step reach 400 K
cannot be predicted when the heating is initiated. As a result, during simulation the time
advance mechanism must constantly check for the occurrence of state events.

I
}:a;)'j
i=l

_ .. -- -- - -_ .... ----- --- --- - - - - - - - - - - - - - Threshold

I
I
I
I

Time~

Figure 1. An example of a state event

The event calendar is a list of time events ordered on their scheduled time of
occurrence. Associated with each event is an event code and additional descriptors which
determine the set of actions. called the event logic, which are implemented when that event
occurs. In the example given above. the end-of-mixing event may initiate a heating step in
the vessel. or may initiate the addition of another ingredient to the vesseL

The event calendar is always in a state of transition. The events are removed from the
calendar when they occur. and new events are added to the calendar as they are schedul¢
by the event logic and other components of the time advance mechanism. Several time
events on the event calendar may have the same time of occurrence.

2.2 Solution of DifferentiallAlgebraic Equations

The continuous overall change in the state of the system can be represented by a system of
non-linear differential/algebraic equations. the state equations. of the following form:

F ( y, Y. t ) = 0
 366

where, y is the vector of dependent variables, the state variables, and y = dy/dt. The initial
conditions, y(O) and y(O) are given. In the BATCHES simulator, the state vector comprises
only those variables which could potentially change with time. For example, if the
material is heated using a fluid in a jacketed vessel and there are no material inputs and
outputs, and no phase change, the composition and the total amount do not change. Only
the temperature, volume, enthalpy and heat duty change with time. Therefore, there is no
need to integrate the individual species balance equations nor the total mass equation.

The state equations are solved using a suitable integrator. The BATCHES simulator
uses the DASSL integrator [8]. The integrator uses backward difference formulas (BDF)
with a variable step predictor-corrector algorithm to solve the initial value problem. The
integrator is robust and has been extensively tested for both stiff and non-stiff system of
equations.

2.3 Detection and Processing of Discrete Changes

As described earlier, the exact time of Occurrence of a state event is not known a priori.
Therefore, after each integration step the time advance mechanism invokes an algorithm to
detect whether a state event occurred during that integration step. To detect the occurrence
of a state event the values of the desired linear combination of state variables before and
after the integration step are compared with the threshold. For example, suppose a heating
step is ended when the temperature of a vessel becomes higher than 400 K. If the
temperature before. an integration step is less than 400 K and after that step it is greater
than or equal to 400 K, then a state event is detected during that step. The direction for
crossing in this case is 'crossing the threshold of 400 K from below'. Thus, if the
temperature before an integration step is higher than 400 K and after that step it is lower
than 400 K then a state event as described above is not detected.

The next step after detecting a state event is to determine its exact time of occurrence,
the state event time. The state event time is the root of the equation:

y. - thr. =0
 367

where, y. is the linear combination of variables for which the state event was detected, and
thr. is the threshold. The upper and lower bounds on the root are well defined, namely,
the time before and after the current integration step. The DASSL integrator maintains a
polynomial for each state variable as part of its predictor-corrector algorithm. As a result,
a Newton interpolation which has a second order convergence can be used for determining
the state event time [7]. During a given integration step more than one state event may be
detected. As shown in Figure 2, variables Yl and Y2 crossed thresholds thrl and thr2'
respectively, during the same integration step. When multiple state events are detected, the
state event time for each event is determined and the event(s) with the smallest state event
time is(are) selected, 1e, in this case. The simulation time is reset to that value and the
state variables are intexpolated to match the new simulation time.

- - - - - - - - - - - - - thrl
I I
I I
I I
I I

Figure 2. Example of multiple state events in one time step

 368

The processing of an event consists of implementing the actions. associated with the
specified event logic. For example, ending a filling step may result in releasing the
transfer line used during filling and advancing the vessel to process the next step. In
BATCHES a library of elementary actions is provided, such as shutting off a transfer line,
opening a valve, releasing a utility. An event logic consists of a combination of these
elementary actions. Also, customized user written logic can be incorporated into
BATCHES to implement complex heuristics which cannot be conveniently modeled with
the existing event logic library.

2.4 Implementation of Operating Decisions

The operations associated with a recipe perform the necessary physical and chemical
transformations to produce the desired end products from raw materials. The operations
can be broadly divided into two categories: those operations which are initiated to pull the
upstream material, and those operations which are initiated because the upstream material
is pushed downstream. The operations in the first category must be independently initiated
while those in the second category are initiated by the operations upstream when they
become ready to send the material downstream. In BATCHES you can specify processing
sequences which define the order in which the 'pull' type operations are independently
initiated. The time advance mechanism reviews the processing sequences at each event to
check whether a piece of equipment could be assigned to initiate an operation. Also,
BATCHES uses a prioritized flrSt-in-flrSt-out (FIFO) queue discipline to process' the
requests for. assigning equipment generated by the operations which push material
downstream. The queues also are reviewed at each event by the time advance mechanism
to determine whether any requests could be fulfilled. The processing sequences and
priorities used in the queue management are specified by the user. Hence, by suitably
manipulating the processing sequences and the priorities the user can influence the
assignment of equipment to operations and the movement of material in the process. The
priorities are also used for resolving the competition for shared resources.
 369

3. Process Variability

The operating conditions in a typical batch process are characterized by random

fluctuations. For example, the duration of an operation may not be the same for every
instance" but instead vary within a range, or a piece of equipment may breakdown
randomly forcing delays due to repairs. The random fluctuations significantly affect the
overall performance of a process and their effects should be included in the decision
making process.

I
I
Reactor 3 I tI::!
I
I x+a
Reactor 2 I I
x I
X- a

6
I
Reactor 1
, , I
I ,
I
Separator I
I
y I
I
I1 __________________________

o o
Time-+ Time-+

(a) (b)

Figure 3. Effect of fluctuations in reactor cycle time on the makespan

Suppose a process consists of three reactors and a separator. Each reactor batch is
processed in the separator which produces the fmal product. Also, suppose that the
successive batches initiated on the reactors are offset by a fixed duration. If one assumes
no process variability, with the reactor batch cycle time of x and the separator batch cycle
time of y, the Gantt chart for processing three reactor batches is shown in Figure 3a. With
no process variability, all the operations are well synchronized resulting in the total elapsed
time of (X + 3Y) between the initiation of the first reactor batch and the completion of the
last separator batch. However, in reality the reactor batch cycle time may not be constant
for each batch. Suppose the reactor batch cycle times are normally distributed with the
 370

mean of x and the standard deviation of 0', and suppose the cycle times of the three reactor
batches are (x - 0), (X + 0'), and x. As shown in Figure 3b, the reactors and the separator
are no more well synchronized due to the variability. The shorter frrst batch results in some
separator idle time, while the longer second batch introduces delays in the third reactor
batch. Therefore, due to the interactions through the separator the effective time to
complete the third reactor batch is (X + 0') instead of x. The total elapsed time between the
initiation of the first reactor batch and the completion of the last separator batch is (X + 3y
+ 0').

The simple example discussed above illustrates the effect of process variability on its
performance. In reality, most of the operations in a batch process exhibit variability.
Furthermore, the interactions betweep operations are quite complex. Therefore, a simulation
study involving random fluctuations in process variables requires careful statistical analysis.
Typically, the variability is taken into account during the decision making through the use
of confidence intervals or through hypothesis testing [5]. In general, the computation of
confidence intervals or the testing of hypothesis is based on results obtained from
experiments with the simulation model. The most common technique used to generate
different conditions for simulation experiments is changing the initial seeds of the random-
number streams used for sampling the parameters. Also, the simulation run length, data
truncation to reduce the bias introduced by the initial transients and the number of
experiments are some of the important factors which must be considered for reliable
statistical analysis.

To represent process variability, the design of a simulator for batch processes must
provide the capability to sample the appropriate model descriptors and to collect the
necessary data for statistical analysis. The average, minimum, maximum and standard
deviation for a variables are commonly required for statistical analysis.

4. Combined Discrete/Continuous Simulation Methodology

In this section the combined discrete!continuous simulation methodology is illustrated using

a simple batch process.
 371

MIX 1

FILTER

MIX 2
Figure 4. Process flow diagram of a batch process

End Fill-A End or

in MIX_I Simulation

Integrate I -i
tOO
(a) Event calendar at time 0.0.

End Fill-A End or

inMIX_2 Simulation

Integrate
MDU.MIX_2
! (b) Event calendar at time 1.0
Equations End or
Simulation

! Integrate
(c) Event calendar at time 2.0
--~
End mixing End or
in MIX_I Simulation

6!5
(d) Event calendar at time 3.5

End mixing End mixing End or

in MIX_l inMIX_2 Simulation

I
6.0
I
6.5 9!0
--~
(d) Event calendar at time 6.0

Figure 5. Changes in the event calendar with the progression of time

 372

Consider a process which consists of three pieces of equipment; two mixing tanks,
named MIX_I and MIX_2, and a filter, named FLTR1. The process flowsheet is shown in
Figure 4. Transfer lines 1 and 2 are available for filling raw materials A and B,
respectively, into the mixing tanks. Transfer line 3 feeds material to the f1lter from either
mix tank A or B. Transfer lines 4 and 5 are used for removing material from the two filter
outputs.

One product, named PR_1, consisting of two operations, MIX and FIL1ER is
processed in this facility. The recipe of the two operations is given in Table 1.

Table 1. Recipe of MIX and FILTER steps

MIX FILTER
1. Fill 20 kg Raw Material A 1. Filter the contents of either
in 1.0 hour (FilL-A) MIX 1 or MIX 2 at the rate of
2. Fill Raw material B at 40 kg/hr 60 kgim. 90% Of the material
until the vessel becomes full coming in leaves as a waste
(FILL-B) stream. Stop fIltering when
3. Mix for 3 ± 0.2 hours (STIR) 20 kg is accumulated in FLTR1.
4. Filter the contents (FEED-FLTR) 2. Clean FLTR1 in 0.5 hour.

The MIX step can be performed on either MIX_lor MIX_2, while the FILTER step
can be performed on FLTR1. A processing sequence to initiate 2 mixing tanks batches is
specifIed. Suppose an event to stop the simulation run at 100 hr is specified.

At the beginning of the simulation the event calendar has two events, 'START
SIMULATION' at time 0.0, and 'STOP SIMULATION' at time 100.0. The 'START
SIMULATION' event forces the review of the processing sequence which in turn start the
FILL-A elementary step in MIX_1. Since there is only one transfer line available to
transfer raw material A, the MIX step cannot be initiated MIX_2. Also, the FIL1ER step
cannot be initiated because none of the mixers is ready yet to send material downstream.
Since the FILL-A elementary step is completed in 1 hr, a time event is scheduled at time
1.0. The filling step in MIX_A entails solving the differential equations for the species
mass balance and the total amount equations. Figure Sa shows the event calendar after all
the events at 0.0 are processed. Since FILL-A requires the solution of differential
 373

equations, the simulator starts marching in time by integrating the state equations. The goal
of the simulator is to advance the process up to the next time event time. Since there are
no state event conditions active during Fill-A step, the integration halts at time 1.0 because
of the time event.

At time 1.0, the Fill-A elementary step is completed and FILL-B is staned in MIX_I.
Also, since transfer line 1 is released after the completion of FILL-A in MIX_I, the MIX
step is initiated in MIX_2. A new event to end Fill-A in MIX_2 is scheduled at time 2.0.
Figure 5b shows the event calendar after all the events at 1.0 are processed. Fill-B ends
based on a state event (MIX_I becoming full), hence there is no event to end FILL-B in
MIX 1 on the event calendar. The FILTER still cannot be initiated. The new set Of
equations to be integrated consist df the species mass balance and total mass equations for
both MIX_I and MIX_2. Since, there is one active state event condition to end Fill-B step
in MIX_I, the integrator checks for a state event after each integration step. The next
known event is at time 2.0 when Fill-A is completed in MIX_2.

Suppose no state event is detected and the integration halts at time 2.0. After
completing Fill-A, MIX_2 has to wait because Fill-B is still in progress in MIX_I and
there is only one transfer line available for transferring raw material B. Therefore, after
processing the events at time 2.0, there is only one event on the calendar as shown in
Figure 5c. The state vector consists of the species balance and total mass equations for
FlLL-B in MIX 1.

Suppose a state event is detected at time 3.5 because MIX_I became full. Transfer line
2 is released and MIX_I is advanced to the STIR elementary step which has no state
equations. The duration of STIR varies randomly between 2.8 and 3.2 hr. Suppose for the
fIrst batch the duration is 2.9 hr. An event is scheduled at time 6.4 to mark the end of the
STIR elementary step in MIX_I. Also, Fill-B is initiated in MIX_2. Figure 5d shows the
event calendar after all the events at 3.5 are processed. The state vector consists of the
species balance and total mass equations for Fill-B in MIX.J. One state event condition is
active, namely MIX_ 2 becoming full. The goal of the simulator is to advance the process
to 6.4. However, a state event is detected at 6.0, and the integration is halted.
 374

At 6.0, MIX_2 is advanced to the STIR elementary step which has no state equations.
Suppose the duration of the STIR elementary step for the second batch is 3.1 .hr. As a
result, a time event is scheduled at 9.1 to mark the end of the STIR elementary step in
MIX_2. Figure 5e shows the event calendar after all the events at 6.0 are processed.
Since both MIX_I and MIX_ 2 are processing the STIR elementary step there are no state
equations to be integrated and the time is advanced to 6.4 hr.

At 6.4, the FEED-FLlR elementary step is staned in MIX_I, and FIL1ER step in
FLlR 1. The state vector consists of the species mass balance and total mass equations for
MIX_I, and species balance and total mass equations for FLlR1. Two state events active
when integration is resumed, one to mark the end of the FEED-FLlR elementary step
when MIX_I becomes empty, and one to stop filtering when the total mass in FLlRl
reaches 300.0 kg. No new time events are added to the event calendar.

Suppose MIX_l becomes empty at 8.4 hr. As a result the flow into FLlRl is stopped,
and MIX_I is released. Since there are no more batches to be made, MIX_1 remains idle
for the rest of the simulation. 12 kg is accumulated in the FLlRl. Mter processing all
events at 8.4, the time is advanced to 9.1 since there are state equations to be integrated.

At 9.1, the FEED-FLlR elementary step is started in MIX_2, and the FIL1ER step is
resumed in FLlR 1. The state vector consists of the species mass balance and total mass
equations for MIX_2, and species balance and total mass equations for FLlRl. Two state
events are active when integration is resumed, and there is only event on the event
calendar, namely, end simulation at 100.0.

At 10.433, the FIL1ER step is ended because 20 kg is accumulated in FLlR1.

Therefore, the flow is stopped with 40 kg still left in MIX_2. Cleaning is initiated in
FLlRl, and an event is scheduled at 10.933 hr to mark the end of cleaning of FLlRi.
Since there are no equations to be integrated the time is advanced to 10.933 hr.

At 10.933, FLlRl is reassigned to the FIL1ER step and the filtration of the rest of the
material in MIX_2 is resumed. At 11.6 hr MIX_2 becomes empty and the filtration is
halted with 4.0 kg left in FLlRl. Since there are no more batches to be made and no more
material to be processed the time is advanced to 100.0 hr and the simulation is ended.
 375

The example given above illustrates how the time advance mechanism works in a
combined discrete/continuous simulator.

5. Conclusions

The combined discrete/continuous simulation methodology discussed in this paper is used

in several simulators for modeling discrete manufacturing 'systems as well as batch and
semicontinuous processes. The time advance mechanism uses the event calendar to
coordinate the execution of the following key blocks: Solution of the differential/algebraic
system of equations, Detection and processing of discrete changes, and Initiating operations
in equipment.

6. References

1. Barton, P.I .• and Pantelides, C.C.: The Modelling and Simulation of Combined Discrete/Continuous
Processes. International Symposium of Process Systems Engineering. Montebello, Canada, August 1991
2. Clark, S., and Joglekar. G.S.: Simulation Software for Batch Process Engineering. NATO Advanced
Study Institute. This volume, p. 376
3. Clark. S., and Kuriyan. K.: BATCHES - Simulation Software for Managing Semicontinuous and Batch
Processes. AIChE National Meeting. Houston. April 1989
4. Czulek, AJ.: An Experimental Simulator for Batch Chemical Processes. Comp. Chern. Eng. 12,253 -
259 (1988)
5. Felder. R., Mcleod, G., and Modlin, R.: Simu1aIion for the Capacity Planning of Specialty Chemicals
Production. Chern. Eng. Prog. 6, 41 - 61 (1985)
6. Helsgaun. K.: DISCO - a SIMULA Based Language for Continuous Combined and Discrete
Simulation. Simulation. I, July 1980
7. Joglekar, G.S., Reklaitis, G.V.: A Simulator for BalCh and Semicontinuous Processes. Compo Chern.
Eng. 8, 315 - 327 (1984)
8. Petzold, L.R.: A description of DASSL : A differentiaVAlgebraic system solver. IMACS World
Congress. Montreal, Canada, August 1982
9. Pritsker, A.A.B.: IntrOduction to Simulation and SLAM II. Systems Publishing CotpOration 1986
10. Schreiber, T.: Simulation Using GPSS. John Wiley 1974
Simulation Software for Batch Process Engineering

Steven M. Clark, Girish S. Joglekar

Batch Process Technologies, Inc., P. O. Box 2001, W. Lafayette, IN 47906, USA

Abstract: Simulation is ideal for understanding the complex interactions in a

batchlsemicontinuous process. Typically, multiple products and intermediates are made in a batch
process, each according to its given recipe. Heuristics are often employed for sequencing of
operations in equipment, assignment of resources and control of inventories. A decision support
tool which can integrate both design and operating features is necessary to accurately model batch
processes.
The details of the BATCHES simulator, designed specifically for the needs of batch
processes, are discussed. The simulator provides a unique approach for representing a process in
a modular fashion which is data driven. The library of process dynamics models can be used to
control the level of detail in a simulation model. Its integrated graphical and database user interface
facilitates model building and analysis of results.

Keywords: Simulation, batchlsernicontinuous

1. Introduction

A general purpose simulation tool for batchlsernicontinuous processes must be able to meet their
special requirements from the computational standpoint as well as from the model representation
and analysis standpoint. The combined discrete/continuous simulation methodology necessary to
model the process dynamics, accompanied with discrete changes in the process state is discussed
by Clark [I]. Apart from the special computational requirements, the large amount of information
necessary to build a simulation model of batch processes requires innovative modeling constructs
to ease the data input and make process representation flexible, modular and intuitively clear. Also,
the simulation results must be presented in a way so as to allow the analysis of the time dependent
 377

behavior of the process, and provide infonnation about the overall perfonnance measures.
Over the past few years, several modeling systems have been reported in the literature,
gPROMS [I], UNIBATCH [4], BOSS [5], which incorporate a combined discrete/continuous
simulation methodology necessary for a general purpose simulator for batchlsemicontinuous
process engineering. However, none of these systems adequately address the special requirements
of batch processes from the process representation, data management and analysis standpoint.
These needs must be fulfilled for a wider acceptance and sustained use of a tool for batch process
engineering.
In this paper, the modeling constructs and the data input and analysis capabilities provided
by the BATCHES simulator are discussed.

2. Process Equipment and Recipes

In a typical batch process, several products are manufactured in the given set of equipment items.
Each product is made according to a specific recipe. A recipe describes the series of operations
which are perfonned in the manufacture of a product. Each operation, in turn, consists of a series
of elementary processing steps perfonned in the piece of equipment assigned to that operation.
A process equipment network is given in Figure I, while the description of a recipe is given in
Figure 2.

Raw material.
RMA

Product
REAC2 Stream

Figure 1. Example of a process equipment network

 378

• Recycle storage operation:

1. Fill raw material 'RMB' at 1000.0 kglhr until the tank becomes full
2. Allow recycle of material from the separator and withdrawal of material by
downstream reactors

• Reactor operation:
1. Transfer 100.0 kg of raw material 'RMA' in one hour. Use operator
'REACTOR-OP' during filling.
2. Preheat the contents of the reactor for one hour using 'STEAM'. Use operator
'REACTOR-OP' during heating.
3. Allow the contents to react for one hour.
4. Transfer 100.0 kg of material from the recycle storage tank in one hour. Use
operator 'REACTOR-OP' during filling.
5. Let the reaction continue for one more hour.
6. Cool the contents and let the material age for one hour.
7. Transfer the contents into the intermediate storage tank in one hour.

• Intermediate Storage Operation:

1. Allow transfer of material into the tank from the reactors and withdrawal of
material by the downstream separator

• Separator operation:
1. Separate continuously the contents of the storage tank into a product and a
recycle stream. Send the recycle stream to the recycle storage tank.

Figure 2. Description of operations in a recipe

In a batch process, operations and equipment items have a many to many relationship, ·that
is, several operations may be performed in a given piece of equipment, and several pieces of
equipment may be suitable to perform a given operation. This represents a significant departure
from the steady state processes where an operation and a piece of equipment have a one to one
relationship.
The recipe in Figure 2 further shows that, during the course of an operation, a series of
physical and chemical changes may take place in the assigned piece of equipment. For example,
during the RXN operation, the first two elementary steps represent physical changes, followed by
a chemical change, and so on. As a result, the mathematical equations which describe the process
dynamics are different for each elementary step. This also is markedly different from a steady state
operation in which a unit involves one physical/chemical change. An operation in a batch process
represents a series of 'unit operations' which are performed in the assigned piece of equipment.
The BATCHES simulator provides two constructs, the equipment network and the recipe
 379

network, to model the many to many relationship between equipment items and operations, and
to model the recipes of various products

2.1 Process Equipment Network

A process equipment network represents the physical layout and connectivity of the equipment
items in the given process. The equipment parameters describe the physical characteristics of each
equipment item such as volume, or heat transfer area. Also, any physical connectivity constraints
are specified in the equipment network through the use of transfer lines. For example, some pieces
of equipment from a stage may be connectible to only a few pieces of equipment from another
stage. Similarly, to transfer material between two stages, a manifold may be available which allows
only one active transfer at a time, or from a storage tank only one active transfers may be allowed
at any given time.

2.2 Recipe Network

Each product in a batch process is represented by a recipe network. In a recipe network, an

operation is represented by the BATCHES construct task, while an elementary processing step
is represented by the construct subtask. Figure 3 shows the recipe network for the recipe described
in Figure 2. The appropriate choice oftask and subtask parameters allows the user to represent
the recipe details. Thus, the basic building block for recipe networks is a subtask.

Subtask Models: The most important subtask descriptor is the model used to represent the
corresponding elementary processing step. A library of 31 subtask models is provided with
BATCHES. Each subtask model in the library represents. specific physicaJJchemical
transformations. For example, filling, emptying, reaction under adiabatic conditions, continuous
separation of material and so forth. The models range from a simple time delay model to a
complex batch reactor model. The pertinent material and energy balances associated with the
underlying transformations are described as a set of simultaneous differentiaJJalgebraic equations
(DAE), the model equations. The formulation of model equations is based on certain assumptions.
For example, the following assumptions are made in one of the models, named '=FILLING': there
is only one active phase in the bulk, the subtask has no material outputs, each material input is at
constant conditions, namely, constant temperature, pressure and composition of species. Based
 380

FILL+EM~TY

Figure 3. Example of a recipe network

on these assumptions the '=FILLlNG' model is described by the following equations:

tty. elM I Itj

M _-:J_ + ".i - = ~ ~ ltjpj Fip j = I,n
dt dt i=1 p=1

elM I It;
- = : E :EFip
dt i=1 p=1

dE I Itj
_ . =:E :EEip
dt i=1 p=1

E =MH.(T'p,xj)
dP
-=0
dt

9pV=M

where,
 381

n number of components V effective volume of equipment

x mass fraction 1t number of phases
E total enthalpy p density
F mass fiowrate e volume fraction
H enthalpy per unit mass I subtask input index
I number of subtask inputs component index
M total mass liquid phase
P Pressure (pa) p phase index
T Temperature (K)
Thus, when a piece of equipment executes a subtask which uses this model, the generalized
set of equations given above is solved for the specific conditions in that piece of equipment at that
time. Since the number of species associated with each recipe could be different, the number of
equations with each instance of a model could be different. Furthermore, since the species
associated with each recipe will be different, the mass fraction variables with each instance of a
model could be associated with different species. The subtask models in the library provide the
modularity necessary for a general purpose simulator. If the models in the library do not meet the
requirements ofa particular elementary processing step, new models can be added to the library.
The other important subtask parameters are: subtask duration, state event description,
operator and utility requirements. Typically, the available resources are shared by the entire
process, and the instantaneous rate of consumption of a resource cannot exceed a specified value
at any time. For example, the maximum steam consumption rate may be constrained by the design
capacity of the boiler, or only one operator of a particular type may be available per shift.
Individual operations compete for the resources required for successfully executing the elementary
processing steps in that operation.
The other building blocks for a recipe network are flow lines, raw materials, and infinite sinks.

Flow Lines: Material is transferred from one piece of equipment to another during specific
elementary processing steps. The transfer of material is represented by a flow line connecting the
appropriate subtasks. The flow parameters describe the amount of material being transferred and
the flow characteristics such as flowrate, and flow profiles.

Raw Materials: A raw material is an ingredient which is available whenever a subtask needs it,
 382

and is characterized by its temperature, pressure, and composition. A raw material is represented
by a pentagon shown in Figure 3 and is identified by a unique name.

Infinite Sinks: An infinite sink represents the material leaving a process. An infinite sink is
represented by a triangle shown in Figure 3, and is identified by the subtask and the output index
connected to it. The simulator continuously monitors the cumulative amount and the average
composition of the material withdrawn from the process.

2.3 Link Between Equipment and Recipe Networks

For every task in a recipe an ordered list of equipment items suitable to process that task is
specified. At a given time, if there is a choice, the piece of equipment best suited to perform the
task is selected. A particular piece of equipment suitable for several tasks appears on the lists
associated with the corresponding tasks. The suitable equipment lists provide the link between the
equipment and recipe networks.

2.4 Executing an Operation in Equipment

After a piece of equipment is assigned to perform an operation, all subtasks in that task are
implemented according to the recipe.
Typically, before the execution of an elementary step certain conditions are checked. For
example, a filling step may start only when sufficient amount of upstream material, a transfer line,
and an operator are available. A heating step may entail checking the availability of a particular
resource. The execution of a subtask begins only when the following conditions are satisfied:
• The required amount of input material is available upstream
• Equipment items are available downstream for every output
• A transfer line is available for each material transfer
• Operators and utilities are available in the specified amounts

2.5 Advantages ofthe Two-Network Representation

Representing a batch process as a set of two network types provides several advantages. First of
all, it is a natural representation of how a batch process operates, because there is a natural
dichotomy between the physical equipment which merely are 'sites', and the standard operating
 383

procedure for manufacturing each product using the sites. As a result, a two-network model
becomes easy to understand at a conceptual level as well as at the implementation level.
The two-network representations provide an efficient mechanism for building simulation
models of multiple manufacturing facilities. For example, very often the same products are made
in different physical locations which may have different equipment specifications and
configurations. To model such facilities one equipment network can be constructed for each
facility, with one common set of recipe networks for the products made in them. To model a
specific facility the appropriate combination of equipment network and recipe networks can be
selected. Similarly, in a given facility several products can be made over an extended period of
time, such as one year. However, for a simulation study, the time horizon may be much shorter,
for example one day or one week, during which the user may want to consider only a few recipes.
In such cases, to create a process model the appropriate subset of recipe networks and the
appropriate equipment network can be selected.

3. Decision and Control Logic

During the operation of a batch process decisions are constantly made which affect its overall
performance, for example, the sequence of making products, assignment of equipment to
operations, assignment of resources to elementary steps, and transfer of material between
equipment items. Also, actions which help in synchronizing operations are implemented based on
the state of the process at that time. In a recipe network, various task and subtask parameters
allow the user to select the suitable decision and control logic options.

Processing Sequences: For a simulation run, sequences for initiating the operations are specified.
The identity of the operation to be initiated and a stopping criteria are specified for each entry in
the sequence. For example, make 5 reactor batches of recipe A, followed by sufficient reactor
batches of recipe B to produce 100 kg of reaction mixture and so on. Typically, those operations
which mark the beginning of the processing of material associated with a product are
independently initiated through processing sequences. The processing sequences merely specity
the desired sequence. The actual start and end times of operations, determined by the simulator
as it marches the process in time, are some of the key performance indicators.
 384

Equipment and Resource Assignment: The requests for assigning equipment items to initiate
tasks are handled by First In First Out (FIFO) queues which are ordered by user specified
priorities. Similarly, the requests for assigning operators and utilities are handled by FIFO queues.
The decision to assign a piece of equipment to initiate a task is governed by the logic option
associated with that task. For example, one of the options may assign a suitable piece of
equipment as soon as one becomes available and then searches for upstream material, downstream
equipment to send material to and so on. Another option may not assign a piece of equipment if
upstream material is not available which prevents unnecessary reservation of equipment. Thus, by
selecting an appropriate option one can accurately represent how assignments are done in the
actual process.

Selection of Upstream Equipment: The selection of upstream equipment items which provide
the input material is governed by a logic flag. The following options are available to select
upstream equipment and start the material flow:
• Wait until the required amount becomes available in a single piece of equipment upstream
• Wait until the required amount becomes available cumulatively in one or more pieces of
equipment upstream
• Start the transfer of material as soon as some material becomes available upstream and keep
searching for additional material until the amount requirement is satisfied

Conditional Branching: Often, during an operation a different set of elementary steps is executed
based on quality control considerations or the state of the material at that particular time. For
example, in the RXN recipe shown in Figure 2, 80% of the batches may require additional
processing of 1 hour after the aging step to bring the material within allowed specifications. This
is modeled by specifying conditional branching information with the AGING subtask.

User Defined Logic: Certain complex decision logic implemented in a process may be outside the
scope of the existing logic options available through various parameter choices. In that case, the
user can incorporate customized decision logic into BATCHES. The simulator provides a set of
utility subroutines which can access the process status. By retrieving the appropriate information,
the user can formulate and implement any desired set of actions triggered by the process status.
 385

4. Shortcomings of Discrete Simulators

The discrete simulators such as SLAM, GPSS, in principle, provide the basic discrete/continuous
simulation methodology necessary for modeling batch processes. However, these simulators are
merely general purpose simulation languages, and therefore require considerable expertise and
effort for simulating batch processes. In spite of their availability for more than 15 years, only
relatively few chemical process applications have been reported in the literature. In general, the
discrete simulators have not been accepted as effective tools for simulating batch processes
because of several limitations. However, it must be noted that because these simulators have been
written in procedural languages like FORTRAN and provide a mechanism to incorporate user
written code there is no inherent limit on adapting them to satisfY a particular need, provided the
user has the luxury of time and the expertise required for developing the customized code.

Process Representation: In discrete simulators, the manufacturing 'activities' are performed on

'entities' by 'servers'. The servers and activities would be equivalent to process equipment and
elementary processing steps in a recipe, respectively. An entity is a widget which can be very
loosely compared to a batch of material. A widget keeps its identity and its movement can be
tracked in the process. Also, when a server completes an activity, modeled as a time delay, the
entity is released and the server can be assigned to process another entity. In a batch process,
during a task several elementary steps are performed in the assigned piece of equipment. Also, not
all elementary steps can be modeled as pure time delays. For example, a step can end based on a
state event, or it can end based on interactions with other pieces of equipment. During a task,
material from several upstream steps may be transferred into a piece of equipment and several
downstream steps may withdraw material from it. Thus, a batch of material may constantly change
its identity. Also, whenever there is a transfer of material between elementary steps, pieces of
equipment must be available simultaneously from several processing stages. For example, in order
to filter material from a feed tank and store the mother liquor in a storage tank, three pieces of
equipment ('servers'), one from each stage, must be available simultaneously.
All parallel servers in a discrete process are assumed to be identical and suitable for an
activity. The parallel pieces of equipment in a batch process are seldom identical, thus resulting
in equipment dependent cycle times for processing steps and also equipment dependent batch
sizes. Also, some pieces of equipment in a stage are often not suitable for certain tasks because
 386

of safety and corrosivity considerations. Additionally, there may be constraints due to connectivity
on transfer of material between certain pairs of equipment items in parallel stages.

Push and Pull: The key mechanism for the movement of entities assumed in the discrete
simulators is 'push entities downstream', that is, whenever a server finishes an activity it releases
the entity and the entity waits in a queue for the assignment of a server to perform the next
activity. However, in batch processes, two mechanisms for the movement of material are
prevalent, namely 'push' and 'pull'. In the 'pull' mechanism, an elementary processing step searches
upstream for the required material and transfers ('pulls') the required amount which may be a
fraction of an upstream batch. For example, a large batch of a catalyst may be prepared which is
consumed by the downstream reactors in small amounts.

Process Dynamics: The discrete simulators use an explicit integration algorithm such as
Runge-Kutta for solving the system of differential equations describing the processing steps. The
explicit methods cannot solve a system of DAEs very effectively and therefore the differential
equations must be explicit as given below:

y=F(y,t}

Also, the explicit methods are not recommended for solving stiff equations. Since even a
simple dynamic process model such as '=FILLING' illustrated in Section 2.1 requires the solution
ofDAEs, the discrete simulators have a severe limitation. The problem is compounded by the fact
that the more complex models such as reactor and evaporation have implicit DAEs and are
generally stiff.
In discrete simulators, the differential equations and the state events must be defined prior to
a simulation run. Applied to the simulation of multiproduct batch processes that would translate
into defining the equations for all feasible combinations of subtasks and equipment items, and the
associated state event conditions, along with the logic to set the values and derivatives of the
variables which would be inactive at a given time, prior to a simulation run. While not impossible,
this is an overwhelming requirement for the team generating and maintaining the model. Also,
defining all of the possible combinations in the state vector and checking the state event conditions
at the end of every integration step would result in tremendous computational overhead
considering that at any time at the most one combination could be 'active' per equipment item. The
 387

modularity of BATCHES eliminates all the programming as well as computational overheads

mentioned above.

Hard-wired Model: In discrete simulators, since the event logic, dynamic process models and
operating characteristics are implemented through user written code, the models tend to be very
problem specific. Therefore, any 'what if ... ' which is outside the scope of the existing code can be
studied only after suitably changing the code. Since the main objective of a simulation study is to
evaluate various alternatives, the prospect of having to change the software code in order to
evaluate the impact of a change restricts its use to experts. Since BATCHES is data driven and
modular, the changes to the process can be made and evaluated easily.

Simulation Output: The output report from discrete simulators is restricted to utilization and
queue statistics about the entities processed. In addition to this information, the BATCHES
simulator provides mass balances and a detailed breakdown of cycle times. The former is necessary
for the determining the production rates of useful as well as side or waste stream, while the latter
is crucial in pinpointing the bottlenecks. The BATCHES simulation output is discussed in the next
section.

5. Simulation Input and Output

A simulation project involves many complex activities, and requires an understanding of systems
and information flow. Activities on a project include collecting data, building models, executing
simulations, generating alternatives, analyzing outputs, and presenting results. Making and
implementing recommendations based on the results need to be a part of a simulation project. The
BAtches SHell, BASH, is a software package which creates an environment that supports all of
these activities. To provide this support, BASH has been integrated with BATCHES and contains
database and graphic capabilities.

5.1 BASH Overview

The BATCHES simulation models are data driven. To build a simulation model, one must specify
appropriate values of various parameters associated with the modeling constructs. BASH provides
interactive graphical network builders and forms to build and alter simulation models. Models built
 388

in this fashion are cataloged and easily recalled.

BASH provides the capability to store, retrieve, and organize simulation generated outputs.
The use of a database makes possible the flexible presentation of project results for comparison
and validation purposes.
BASH has a complete system for providing both graphical and tabular outputs. It segregates
formats from the data to be displayed and allows general or stylized formats to be built
independently. Over the years, formats used on various projects become available for reuse.
An animation is a dynamic presentation of the changes in the state of a system or model over
time. A BASH animation is presented on a facility model. Icons are used to represent the elements
of the system. During a simulation run, traces are collected for user specified events. The
animation rules translate an event trace into actions on the screen.
The architecture of BASH is shown in Figure 4. The outer ring in Figure 4 indicates the user
interfaces to BASH. The BASH language gives the user access to the BASH modules for network
building, schedule building, building simulation specifications, control building, data entry, format
preparation, and facility, rule and icon building. In addition, the BASH language allows the user
to specify directly operations such as data analysis, report generation, graphics generation, and
animation.

Figure 4. BASH architecture

 389

5.2 Simulation Output

The BATCHES simulator generates textual summary reports and time series data for analyzing
the performance of a simulation run.
The summary report consists of mass balance summary, and equipment and resource
utilization statistics. Additionally, the cycle time summary for each piece of equipment, along with
a breakdown of the time lost due to waiting for material and resources, is reported. An example
of the cycle time and wait time summary for a piece of equipment, which is suitable to perform
task {PR 1, RXN}, is given below.

..... -•..•...•.•••......•....•..••..•..•..• _....

* BATCH CYCLE TIME AND WAITING TIME STATISTICS *
TIME IN (hr)

**** EOUIPMENT NAME : REAC 1

(P. T) BATCHES TOT-PROC-TM AV-CYCLE-TM TOT-ACTIV-TM TOT-WAIT-TM
PR 1
RXN 183.0 22.88 112.0 71. 00

DURING SUBTASK TOTAL TIME SPENT WAITING FOR ACTIVE

UPSTR-EOP DONNSTR-EOP OPS+UTILS SIC-CHAIN TIME
FILL-RMA O. O. 54.00 O. 8.000
PREHEAT O. O. 6.000 O. 8.000
REACT-1 O. O. O. O. 32.00
FILL-RHB 3.000 o. O. O. 16.00
REACT-2 O. O. O. O. 32.00
AGE o. O. 8.000 O. 8.000
EMPTY O. o. O. O. 8.000

((TOTAL)) 3.000 O. 68.00 O. 112.0

BATCHES TOT-PROC-TM TOT-ACTIV-TH TOT-WAIT-TM

TOTAL FOR REAC 1 8 183.0 112.0 71.00

First, the name of the piece of equipment is printed. Next, the name of the task and the
number of batches completed, the total processing time, average cycle time (Total processing
time/number of completed batches), the total active time and the total waiting time for the
\

corresponding task are printed. The total processing time is the sum of the total active time and
the total waiting time. This is followed by a detailed breakdown of the waiting and active times
for each subtask. The time spent waiting for upstream material andlor transfer line (UPSTR-EQP),
the downstream equipment (DOWNSTR-EQP), operators andlor utilities (OPS+UTIL). Column
S/C-CHAIN denotes the time spent waiting for either an upstream subtask to send information to
 390

proceed or a downstream subtask to initiate the flow of material. The last column denotes the time
spent in actively processing the subtask.
The cycle time statistics are critical in identifYing bottlenecks. For example, ifthe waiting time
for operators/utilities is significant, then increasing their availability may resolve the problem, or
if waiting time for downstream equipment is significant, adding a piece of equipment in the
appropriate stage may resolve the problem.

5.1 Time Series Output

During a simulation run, one can collect time series data for the selected variables, for example,
the amount of material in a specific piece of equipment, the process inventory, the cumulative
production, utility usage and so on. The time series data are useful in characterizing the dynamic
behavior of a process. For example, when do the peaks occur, how saturated a particular variable
is, is there any periodicity etc. Also, by comparing the time series data for the desired variables one
can measure the amount of interaction or establish correlations between them. For example, are
equipment items demanding a particular utility at the same time, are the product inventories related
to the unavailability of operators and so on. Most importantly, by comparing time series data from
various simulation runs one can study the effects of changes on the dynamic behavior of a process.
Each simulation run is one 'what if .. ' defined by a particular combination of values of model
descriptors, a 'scenario'. The time series data for a simulation run are stored in the BASH database
under the scenario name.
Typically, the first step in the analysis of time series data is to present the data in a suitable
graphical form such as plots, Gantt charts, pie charts and so on. BASH provides a wide variety
of presentation graphics options including the ones mentioned above [2]. Examples of a plot and
a Gantt chart are shown in Figures 5 and 6. For detailed quantitative analysis, BASH has
functionalities to compute summaries for characterizing the time series data such as minimum,
maximum, average, standard deviation, frequency distributions and so on. Also, the database
provides additional data manipulation capabilities such as extracting data within a specific time
window, or extracting data occurrences within a specific range of values. After extracting data
which satisfY certain characteristics one can either present them graphically or compute summaries.
Thus, very detailed information required for analysis can be easily derived. For example, the total
 391

Jlass fractions in REACTOR during /"REACT/

1.0

0.8

0.6

0.4

0.2

o.oJILC-=:~--r---:==::=:::;====i=~~;;:'
0.0 200.0 400.0 600.0 800.0 1000.0 1200.0 1400.0 1600.0
TIME (min)
-+- O-XYLENE.1
-+- M-XYLENE.1
-+- P-XYLENE.l
-)to BENZENE.1
TOLUENE.1

Figure 5. Example of an X-Y plot showing concentration profiles in a piece of equipment during a
subtask

time when the process inventory was higher than a specific value, or the time spent by a piece of
equipment in processing a particular task and so on.
The BASH database facilitates comparison ofresuIts from various simulations. The data from
multiple simulation runs can be presented in multiple windows, or alternatively the desired
variables from multiple simulation runs can be simultaneously displayed on one graphical output.

6. Conclusions

The BATCHES simulator, designed for the special requirements of batchlsemicontinuous

processes, provides a unique approach for representing a process in a modular fashion which is
 392

lASt

SEPARATOR

REACTOR2

REACTOR!

RECSTN2

RECSTN!

0.0 50.0 100.0 150.0 200.0 250.0 300.0 350.0 400.0

TNOW
o CR UDE-GRADE
~ FINISHED_GRADE
m HIGH_GRADE
REFINED_GRADE
XYLENEJEED

Figure 6. Example of a Gantt chart based on recipe in equipment. (Equipment names are shown in the left
hand column)

data driven. Its integrated graphical and database user interface facilitates model building and
analysis of results.

7. References
1. Barton, P.I., and Pantelides, C. c.: The Modeling and Simulation of Combined Discrete/Continuous
Processes. International Symposium of Process Systems Engineering, Montebello, Canada, August 1991
2. BASH User's Manual. Batch Process Technologies, Inc., West Lafayette, IN, (1992)
3. Clark S., and Joglekar, G.S.- Features of Discrete Event Simulation. NATO Advanced Study Institute,
This volume, p. 361
4. Czulek N.: An Experimental Simulator for Batch Chemical Processes. Compo Chem. Eng. 12,253- 259
(1988)
5. Jog\ekar, G.S., Reklaitis, G.V.: A Simulator for Batch and Semicontinuous Processes. Compo Chem.
Eng. 8, 315 -327 (1984)
The Role of Parallel and Distributed Computing Methods in
Process Systems Engineering*

Joseph F. Pekny

School of Chemical Engineering, Purdue University. West Lafayette IN 47907, USA

Abstract: Parallel computers are becoming available that offer revolutionary increases in
capability. These capability increases promise to shatter notions of intractability and make a
substantial difference in how scientists and engineers fonnulate and solve problems. This
tutorial will explore some process systems areas where a parallel computing approach and
advances in computer technology promise a substantial impact. The tutorial will also
include a summary of trends in underlying technologies, current industrial applications, and
practical difficulties. An important underlying theme of the tutorial is that researchers must
begin to devise flexible algorithms and algorithm design environments to help close the
widening gulf between software and hardware capability. In particular, since parallel and dis-
tributed computers complicate algorithm design and implementation, advanced software
engineering methods must be used to reduce the burden.

Keywords: parallel computing, distributed computing, special purpose algorithms, process

engineering, algorithm design

1. Introduction
Computers are a critical enabling technology for the Chemical Process Industries. Indeed,
no modern chemical plant could function or could have been designed without the benefit of
the computer. The last thirty years have seen a 10,000 fold increase in the performance per
unit cost of computers [41]. Each order of magnitude increase in capability has brought
about a qualitative and quantitative improvement in process efficiency. This improvement
manifests itself in better energy, equipment, and manpower utilization as well as in improved
safety, inventory management, and flexibility. Commensurate with the important role of
computers, the study of process systems engineering has intensified. Advances in methodol-
ogy, algorithms, and software as well as ever expanding applications has been made possible
by the rapid advances in computer technology. In this tutorial, we will explore the impact

• excepted in part from J. P. Pekny, "Parallel Computing Methods in Chemical Engineering", course notes,
ChE 697P, School of Chemical Engineering, Purdue University, West Lafayette, IN 47907
 394

that high performance computing will have on the direction of process systems engineering
in the near future. The tutorial is divided into three parts: (i) advances in computer technol-
ogy and basic concepts, (ii) designing parallel algorithms, and (iii) using optimal computer
architectures through distributed computing.

2. Bask Concepts and Advances in Computer Technology

Parallel computing is a dominant theme of high performance computing in this decade. One
operational definition of parallel computing is the application of spatially distributed data
transformation elements (processors) to cooperatively execute a computational task. The
underlying goal of all parallel computing research is to solve currently intractable problems.
In science and engineering at large, there are a number of applications that make nearly
unlimited demands on computer capability. For example,
• protein, polymer conformations and design
• global climate prediction
• artificial intelligence
• image, voice recognition
• computer generated speech
• quantum chromodynamics - calculate mass of elementary particles
• computational chemistry - quantum mechanics
• astronomical calculations (e.g. galaxy formation)
• turbulence - prediction and control
• seismology and oil exploration
• fractals and chaos
• wind tunnel simulations
• gene sequencing
Within Chemical Engineering a number of areas demand dramatic increases in computer
capability. For example,
• Simulations for Molecular Description of ChE Processes
• Process Design, Control, Scheduling, and Management
• Transport Phenomena, Kinetics (Fluid flow simulation, Combustion, Chemical vapor
deposition, Reactor profiles)
• Boundary Element Methods (Suspension simUlations, Protein dynamics)
• Molecular Dynamics (Protein solvation, Macromolecular chain conformations)

• Monte Carlo Simulation (Thermodynamics, Reaction pathways,

Polymerization/pyrolysis reactions)
 395

• Expert Systems and AI (Fault diagnosis, Operator assistance)

In the 1960s, computational research was sometimes considered the weakest contributor of
the theoretical, experimental, and computational triad underlying science and engineering.
However continued dramatic improvements in computer capability will result in an increas-
ing reliance on computational methods. Computational research presents exciting opportuni-
ties since rarely in the history of science and engineering have the underlying tools improved
so dramatically in such a short period of time.
The hardware performance increases to be realized over the next decade are funda-
mentally different than the hardware performance increases realized over the last twenty
years which largely came about without any effort on the part of applications researchers. In
particular, the boundaries separating applications researchers, computer scientists, and com-
puter architects is blurring since the design of high performance computers and support tools
is largely dependent on specific applications and the choice of algorithms is dictated by
hardware considerations. The users of high performance computing can still largely be
shielded from the complexities of parallel computing by suitably packaging the algorithms
and applications software.

2.1. Trends in Computer Technology

Based on projected improvements in circuit packing densities, manufacturing technology,
and architecture, there is every reason to expect computer capability gains at the same or an
accelerated rate for the foreseeable future [7]. In fact, [7] projects that high performance
computers could achieve 10 12 operations per second by 1995 and 10 14 operations per second
by 2000 or roughly a factor of 100 and 10,000 times faster than the highest performance
computers today, respectively. Supporting these processing rates will be main memories of
10 to 100 gigawords and disk storage up to a terabyte [42]. Mirroring the advances in paral-
lel supercomputing, desktop, laptop, and palm-top computers should increase in capability to
the point that these classes of computers will be as or more powerful than current supercom-
puters with differentiation occurring' as to the types of peripherals that are supported, i.e.
input devices, graphics, disk storage, and I/O capability [42].
Because computer circuitry is approaching fundamental limits in speed, the high per-
formance computers of the future will certainly possess hundreds to many thousands of pro-
cessors. As flexible computer manufacturing techniques are perfected, cheap and high per-
formance special purpose architectures will proliferate for common calculations. Already
manufacturers cheaply produce special purpose control hardware and high performance spe-
cial purpose machines exist for digital signal processing [7]. Within the process industries
we can expect to see further special purpose hardware arise to perform such computations as
expensive thermodynamic property calculations, sophisticated control algorithms, and sparse
matrix computations for simulation and design calculations. There is every reason to expect
that the most useful and widely needed algorithms will be compiled into silicon especially
since the process industries are expected to be among the top five consumers of high perfor-
mance computing [7].
 396

The fundamental improvements in digital electronics technology will impact each of

the components of computer architecture and hence application capabilities. Below, a short
projection is given for capabilities that can be expected in the near future for each of the
major architecture components.

2.1.1. Processing Elements

Speed improvements in computer processing elements are limited to a constant factor, how-
ever costs for a given unit of capability will continue to fall dramatically for the foreseeable
future. Fabricators hope to make 1,000 Million Instruction Per Second (MIPS) processors
available within the next few years (Sun Sparc, Digital Alpha, mM Power Series, HP Preci-
sion Architecture) for use within engineering workstations [21]. Supercomputer vendors
have entered into agreements to build machines based on hundreds or thousands of these pro-
cessors which have substantial support for parallel processing built within them. Indeed,
these processors will use a substantial amount of parallelism internally to achieve the pro-
jected speeds (instruction pipelining, multiple instruction launches, 64-bit words, etc.). Spe-
cial purpose processing elements have been and are being developed for important opera-
tions from linear algebra (matrix inversion, multiplication), multimedia presentation (ray
tracing, animation with high resolution graphics/audio), and data transformation (compres-
sion, encryption, Fast Fourier Transformation, convolution). With the advent of effective sil-
icon compilers, computational researchers will have the option of implementing part or all of
an algorithm in hardware for maximum speed. There are technologies on the horizon such as
optical and quantum well devices that promise to dramatically improve (one or more orders
of magnitude over conventional technology) the sequential speed of processing elements but
they are probably at least ten years away from having a practical impact. Thus the prolifera-
tion of parallel processing is virtually assured into the next century. When new implementa-
tion technologies are practical, they will not usher in a new era of sequential computing since
they can also be combined into parallel architectures for maximum performance. The
experience gained with parallel algorithms and hardware will ultimately guarantee this
result.

2.1.2. Volatile Memory

The memory hierarchy within computers will continue to be stratified with the introduction
of additional caching layers between fast processors and the most abundant volatile memory
[15]. The management of this hierarchy in a parallel environment continues to be the focus
of intense research interest. There is also a trend to make some memory elements active in
the sense that they perform simple processing activities in addition to passively storing data.
This is just another manifestation of parallel processing. Passive memory capacity will con-
tinue order of magnitude improvements in capacity but not access speed over this decade.
Techniques for guaranteeing the reliability and integrity of large amounts of passive memory
will become increasingly important as computing systems utilize gigabytes and terabytes of
memory capacity.
 397

2.1.3. Long Distance Interconnection

Bandwidth and latency are two fundamental measurements of interconnection technology.
The latency of a computer network is defined to be the amount of time required for a small
message to be sent between two computers. Network bandwidth is defined as the rate at
which a large message can be transmitted between two computers. Actual values for latency
and bandwidth depend on the geo~raphic location of the computers, network hardware and
software technology, and the amount of network traffic. Latency is limited by the speed of
light but there is no fundamental limit to bandwidth. The next five years will see dramatic
improvements in bandwidth (as much as a factor of 10,(00) as fiber optic technology
matures. Advanced interconnection technology will allow cooperative parallel computing
over regional, national, and international distances. Quite likely specialized supercomputing
centers will arise with dedicated hardware for performing certain tasks. The availability of
high bandwidth interconnection technology will allow automatic utilization of these centers
during the course of routine calculations.

2.1.4. Permanent Storage

The next few years will be characterized by a proliferation of alternative permanent storage
strategies: CD-ROM, Read-Write Optical Disk, etc. In addition, basic magnetic storage
technology will continue to make dramatic improvements in cost per bit of storage, access
times to data, and bandwidth through new combinations of conventional disk technology
such as Redundant Arrays of Inexpensive Disks (RAID) [36]. Bandwidth and reliability
increases will primarily be a result of data parallelism as researchers and vendors perfect the
use of interleaved disk drives. By the end of the decade, disk drive capacity per unit cost
should increase several orders of magnitude (factor of 1,(00). The decreasing cost of
integrated circuits will make disk controllers and interface electronics ever more sophisti-
cated which is just another example of parallelism. The general trend is order of magnitude
improvements in storage capacity at declining cost and increased accessibility of data.

2.1.5. Sensor Technology

Dramatic cost reductions in all forms of digital electronics will continue to promote the
introduction of lower cost sensors within all industries. As a result this decade will see an
unprecedented increase in the volume of data made available concerning the manufacture,
distribution, and consumption of products. Thus there will be a great demand to transform
this data into useful knowledge. Parallel computers will playa critical role in transforming
this data into knowledge via increasingly sophisticated models. High speed computer inter-
connection networks will dramatically decrease the cost of instantaneously transporting large
amounts of data.

2.1.6. Interface Technology

Interface technology serves the dual purpose of facilitating user convenience and inducing
intuition in applications. Most improvements in interface technology will take place in the
area of audio-graphics. High resolution, three dimensional, color graphics depictions of
 398

complicated simulations and calculations will soon become commonplace. During the latter
part of this decade voice and handwriting recognition as a means of user input should
become commonplace.

2.1.7. Overall Component Technology Conclusion

The only barrier to continued dramatic advances in computer technology lie in the ability to
sustain performance increases for processing elements. Parallel computing should circum-
vent this limitation. Thus the overall trend in computing through the 1990's should be sus-
tained order of magnitude improvement in all hardware aspects. A continuing barrier to a
commensurate increase in the usefulness of computers is the inability to improve algorithm
design and software implementation advances at the same rate. A significant amount of
research effort will have to be devoted to parallel algorithm design and software implementa-
tion. The interaction of algorithms with architecture will force applications researchers, such
as chemical engineers, to become more familiar with all aspects of parallel computing,
software engineering, and algorithm design.

2.2. Trends in Software Engineering

The development of special purpose computer hardware mitigates the difficulty of develop-
ing parallel computing software since manufacturers will hardcode algorithms in silicon.
However, the single greatest barrier to fully exploiting incipient computer technology lies in
the development of reliable and effective software. The process industries will benefit from
computer science research in the area of automatic parallelizing compilers, parallel algorithm
libraries, and high level parallel languages but current research indicates that effective paral-
lel algorithms require applications expertise as well as an understanding of parallel comput-
ing issues [17,33]. Thus the process industries will have to sponsor research into the
development of parallel algorithms for industry specific problems that are computationally
intensive [26,34]. Even putting parallel processing issues aside, the development of high
quality software will remain a hurdle to taking full advantage of this capability. Indeed,
several studies have shown that software development productivity gains lag far behind gains
in hardware since the 1960s [12]. Computer scientists have introduced notions such as
object-oriented programming which promotes large scale reuse of code, ease of debugging,
and integration of applications and Computer Aided Software Engineering (CASE) tools that
reduce development, debugging, and maintenance effort but much research remains before
computers achieve the potential offered by the hardware alone.

2.3. Progress in Computer Networking

So far we have only outlined the expected advances in stand-alone computing capability.
However, perhaps the greatest potential for revolutionizing computer-aided process opera-
tions over the remainder of the decade lies in combining this promised capability with per-
vasiveness and high speed networks. Pervasiveness of computing technology will be made
possible by the fact that 100 million operations per second computers with megabytes of
memory will become available for only a few dollars and they will only occupy a single chip
 399

involving a few square inches [42]. Such digital technology will allow manufacturers to
imbue even passive process equipment with a certain level of intelligence. Thus we will
have pipes that not only always sense fiowrates, temperatures, and pressures but also warn
when they might rupture, tanks that always keep track of their contents, and equipment that
suggests how more economic use may be obtained or suggest to field technicians how it
might be serviced. At the very least, pervasive digital technology will make an extraordinary
amount of information available about processes and the component equipment. Some
researchers even suggest that hundreds of embedded computers will exist in an area the size
of an average room [42]. Wireless networks and high speed fiber optic networks will enable
the enormous amount of information to be instantaneously transported to any location in the
world at very little cost [41].
By 1995, limited deployment of one gigabit/second fiber optic networks will occur
and by early next century gigabit/second fiber networks and 0.25 to 10 million bit/second
wireless networks promise to be as common as copper wire based technology [41]. In fact,
limited gigabit/second fiber networks are now becoming available in experimental testbeds
sponsored by the U.S. government [19,41]. Global deployment of such high speed networks
will offer a multitude of opportunities and challenges to the process industries. For example,
the intimate linking of geographically distributed plant information systems with high speed
networks will make close coordination of plants a competitive necessity. An order originat-
ing at one point on the globe will be scheduled at the most appropriate plant based on tran-
sportation costs, plant loadings, raw material inventories, quality specifications, etc. A single
order for a specialty chemical could instantaneously spawn an entire web of orders for inter-
mediates and raw materials not just within a single company but across many suppliers and
vendors. Processes and inventories will be controlled with great efficiency if scheduling and
management systems can be constructed to take advantage of this information. A field tech-
nician servicing a piece of process equipment could call up schematics, illustrations, a com-
plete service record, and operating history as well as consult with experts or expert systems
all from a lap-top or palm-top computer. High bandwidth networks will allow expert consul-
tants to view relevant process information, consult with operators and. engineers, view pro-
cess equipment, and exchange graphics all without leaving the office. The abundance of
information will offer greatly increased opportunities for improvements in on-line process
monitoring, diagnosis, and control. The research challenge is clearly to develop algorithms
that can profitably exploit this information. Certainly a large amount of information is
already gathered about processes, although most engineers will attest that it is put to little
use. Digital pervasiveness will ensure that at least an order of magnitude more of informa-
tion will become available about processes but greatly increased computer capabilities and
high speed networks promise that this information can be put to productive use.
From a modeling and simulation perspective, high speed networks offer the oppor-
tunity for large scale distributed computing whereby demanding calculations are allocated
among computers located around the globe. In a distributed environment, heterogeneous
computing becomes possible. Different portions of a calculation can be transported to the
most appropriate special purpose architecture and then reassembled. In such an environ-
ment, aggregate computer power is the sum of the machines available on the network. The
 400

research challenges are clear: Which problems are amenable to distributed solution? How
should problems be partitioned? How will distributed algorithms be designed and imple-
mented in a cost efficient manner? The key challenge is to learn to deal with network
latency which arises due to the finite speed of light. Latencies prevent computational
processes from communicating with arbitrarily high frequencies, however, large bandwidths
allow them to exchange enormous quantities of information when they do communicate.
Distributed computing algorithms must be designed to accommodate this reality.
Having offered an overview of emerging computer technology, we will now discuss
basic concepts necessary for designing and assessing parallel algorithms.

2.4. Terminology and Basic Concepts

Parallel computing can be studied from many different perspectives including fabrication of
hardware, architecture, system software design and implementation, and algorithm design.
This section will be confined to those concepts that are of immediate importance to design-
ing effective algorithms for process systems applications. We begin with simple metrics for
measuring parallel algorithm performance and then discuss those architectural attributes that
control the nature of parallel algorithms.

2.4.1. Granularity
The concept of granularity is used qualitatively to describe the amount of time between inter-
processor communication. A computational task is said to be of coarse granularity when this
time is large and of fine granularity when this time is small. Granularity provides a broad
guide as to the type of parallel computer architecture with may be effectively utilized. As a
rule of thumb, coarse granularity tasks are very easy to implement on any parallel computer.
Some practical problems result in coarse granularity tasks, e.g. decision problems that arise
from design, optimization, and control applications.

2.4.2. Speedup and Efficiency

Speedup and efficiency are the most common metrics used to rate parallel algorithm perfor-
mance. For the simplest of algorithms, speedup and efficiency can be computed theoreti-
cally, but for most practical applications speedup is measured experimentally. The most
stringent definition of speedup is
Speedup = Timefor 1 processor. most efficient algorithm
Time for n processor algorithm
In practice, speedup is usually reported as
Speedup = Time for 1 processor, using n processor algorithm
Time for n processor algorithm
The definition for efficiency uses the notion of speedup

Efficiency = Speedup x 100%

Number of Processors
The goal of the parallel algorithm designer is to achieve 100% efficiency and a speedup
 401

equal to the number of processors. This is usually an unrealistic goal. A more achievable
goal is to develop an algorithm with efficiency bounded away from zero with an increasing
number of processors. Many papers use the second definition of speedup which often gives a
faulty impression of an algorithm's quality. The first definition of speedup is consistent with
the goal of using parallel computing to solve intractable problems, while the "practical"
definition is not necessarily consistent. To see that this is so, consider the speedup that is
possible with a sorting algorithm that generates all possible permutations of n items and
saves a permutation that is in sorted order. Obviously, such" an algorithm is inefficient but
since speedup and efficiency are relative measures, a parallel algorithm judged using the
second definition of speedup could look attractive. Problem size is often a central issue.
Consider matrix addition. If the number of processors is larger than the number of elements
in the matrix then efficiency will suffer. As long as the number of elements exceeds the
number of processors then processor efficiency may be acceptable (depending on the com-
puter architecture). Early research in parallel computing was centered on determining the
limits to speedup and efficiency.

2.4.3. Amdahl's Law

In the late 1960s, Gene Amdahl argued that parallel computing performance was fundamen-
tally limited. His arguments can be concisely stated using the following:

S(n)!> ~!>]
f+
P
where f is the fraction on an algorithm that is inherently sequential and S(n) is the speedup
achievable with n processors. This expression of Amdahl's law simply says that, even if the
execution time of the parallel portion of an algorithm is made to vanish, speedup is limited
by the inherently sequential portion of the algorithm. Amdahl's law is often cited as the rea-
son why parallel computing will not be effective. In fact there is some controversy over how
to measure f, but by any measure, f is" very small for many engineering and scientific applica-
tions. Late"r, we will discuss the concept of hierarchical parallelism as a way to counter
Amdahl's law. The essential premise behind hierarchical parallelism is that there is no truly
sequentially calculation in the sense that parallelism is possible at some level. Indeed, the
fraction of an algorithm that is inherently sequential largely depends on the type of computer
architecture which is available.

2.4.4. Computer Architecture

A central issue in parallel computing is the organization of the multiple functional units and
their mode of interaction and communication. Such architectural issues are of importance to
algorithm designers since they impact the type of algorithms which can be effectively imple-
mented. Parallel computing architectures may be classified along a number of different attri-
butes. Probably the simplest classification scheme is due to Flynn [10] who designated com-
puters as SISD (Single Instruction, Single Data), SIMD (Single Instruction, Multiple Data),
or MIMD (Multiple Instruction, Multiple Data). SISD computers were the conventional
design from the start of the computer era to the 1990s in which one processor performs all
 402

computational tasks on a single stream of data. SIMD computers are embodied as an array
of processors which all perform the same instruction on a different streams of data, e.g. for a
parallel matrix addition, each processor adds different elements together. MIMD computers
represent the most flexible parallel computer architecture in which multiple processors act
independently on different data streams. Unfortunately, MIMD machine performance is
hampered by communication and coordination issues which can impact algorithm efficiency.

2.4.4.1. Topology of Parallel Computing Systems

Because placement of data and the ability to communicate it in a time critical fashion is cru-
cial to the success of parallel algorithms, application researchers must be aware of the rela-
tionship among hardware components. Parallel computing vendors employ three baSic stra-
tegies in the construction of systems: (1) use large numbers of cheap and relatively weak
processors, e.g. pre-CMS Connection Machine Computers, (2) use dozens of powerful off-
the-shelf microprocessors, e.g. Intel Hypercube which consists of 16 to 512 Intel i860 pro-
cessors, and (3) use small number of expensive and very powerful processors, e.g. Cray
YMP/832 which possesses eight 64-bit vector processors. Architectures relying on large
numbers of weak processors for high performance demand high quality parallel algorithms
with a very small sequential character. To date, only very regular computations have been
suitable for massively parallel computers such as the Connection Machine. Vendors have
had trouble making money with strategy two in recent years since by the time a parallel com-
puter was designed and built microprocessor speed improved resulting in the need to
redesign the architecture or market a machine barely competitive with the latest generation
of microprocessor. As technological limits are reached with microprocessors this should
become less of a problem, although these limits may be somewhat circumvented by increas-
ing the amount of parallelism available in microprocessors. Supercomputer vendors such as
Cray have been very profitable using strategy three, however in the Fall of 1990 they com-
mitted their research efforts to developing a next generation of supercomputers using hun-
dreds and perhaps thousands of powerful processors.

communication bus

Example: Sun Microsystem Workstation

Figure l. Traditional (Von Neumann) sequential computer architecture.

 403

For purposes of algorithm design, the most important architectural features are the
relationship of processors to memory and each other. Figure 1 illustrates the traditional
sequential computer architecture. A defining feature of this architecture is the communica-
tion bus over which instructions and data are moved among the processor, memory, and
input/output devices. The rate at which the bus transmits data is a controlling factor in deter-
mining overall computer performance. Figure 2 shows a parallel extension of the traditional
architecture that places multiple.-processors and memory modules along the communication
bus.

communication
bus

Examples: Alliant FX series, Sequent Symmetry

Figure 2. Extension of traditional architecture to multiple processors.

The bus architecture can support 0(10) processors until bus contention degrades system
efficiency. Cache memory located on a processor can be used to reduce bus contention by
reducing bandwidth requirements. However, the coherency of data in various processor
caches is a complicating factor in computer and algorithm design.

Examples: University of Illinois Cedar Project

Figure 3. Hybrid architecture avoids hus contention.
 404

Even though the bus architecture can support only a small number of processors, it can typi-
cally be incorporated into hybrid architectures. Processors communicate through the com-
mon memories. Some mechanism must exist for arbitrating collisions among processors try-
ing to access a given memory. A memory hot spot is said to exist when several processors
attempt to access a single memory. Algorithms should be designed to avoid these hot Spots
by controlling the way in which data is accessed and by distributing data across different
memory modules. Figure 3 illustrates how the bus architecture may be combined in a hybrid
pattern to support a larger number of processors. Data frequently used by a processor must
be stored in a local memory while infrequently used data may "be stored in a remote memory.
In general, the placement of data in a parallel computer controls algorithm effectiveness.
Figure 4 depicts a crossbar architecture which tends to minimize contention by providing
many paths between processors and memories. Unfortunately the cost of the interconnection
matrix in a crossbar architecture scales as the product of the number of memories times the
number of processors. Because of this scaling, the crossbar becomes prohibitively expensive
for large numbers of processors and memories. The crossbar can be embedded in hybrid
architectures.

Example: Cray YMP

Figure 4. Crossbar Architecture.

There has been an enormous amount of research concerning interconnection strategies whose
performance is nearly as good as a crossbar but at substantially less cost. We will briefly
examine a few of the best such interconnection strategies. Probably the most widely investi-
gated alternative to the crossbar interconnection strategy has been the hypercube architecture
shown in Figure 5 for three dimensions. In general consider the set of all points in a d-
dimensional space with each coordinate equal to' zero or one. These points may be con-
sidered as the comers of a d-dimensional cube. If each point is thought of as a
processor/memory pair and a communication link is established for every pair of processors
that differ in a single coordinate, the resulting architecture is called a hypercube. A number
 405

of architectures can be mapped to the hypercube in the sense that the hypercube may be
configured to simulate the behavior of the architecture. The hypercube architecture
possesses a number of useful properties and many symmetric multiprocessors implemented
in the late 1980s were based on the hypercube architecture.

Example: Intel Hypercube

Figure 5. Hypercube architecture in three dimensions.

The cost of the hypercube interconnection strategy scales as O(p log(p» where p is the
number of processors. Figure 6 illustrates the pipeline architecture which is widely used.

processors are stages of pipeline

Figure 6. Pipeline archi tecture performs like an assembly line.

In fact many microprocessors utilize pipelines to accelerate perfonnance. A pipeline with k

stages is designed by breaking up a task into k subtasks. Each stage perfonns one of the k
subtasks and passes the result to the next processor in the pipeline (like an automobile
assembly line). The task is complete when the result emerges from the last stage in the pipe-
line. The pipeline architecture is only effective if several identical tasks need to be pro-
cessed. If each subtask wkes unit time and there are a large number of tasks (say n) to be
processed, the speedup of a pipeline architecture with k stages is kn/Ck-l + n) or approxi-
mately k. In practice, pipeline architectures are limited by how fast data can be fed into
them. The feed rate is controlled by the memory subsystem design.
 406

2.4.4.2. Memory Subsystems

Memory subsystem perfonnance is usually enhanced by organlZmg memory modules to
function in parallel. Assuming that a memory module can supply data at a unit rate, k inter-
leaved memory modules can supply data at a rate of k units. In order for this to work, data
has to be organized in parallel in such a fashion that the k pieces of simultaneously available
data are useful. This places a burden on the algorithm designer or a compiler to situate data
so that it may be accessed concurrently.
Figure 7 illustrates a memory caching system which is necessary because processors
are typically much faster than the available commodity memory. Memory sufficiently fast to
keep up with processors is available but is very expensive. In order to counter this situation.
computer architects use a memory cache. A small amount of fast memory (Ml) is used at the
top of the cache to feed the processor. Whenever the processor asks for an address that is not
available in MI, the cheaper but slower memory M2 is searched. If M2 contains the
requested memory, an entire block of addresses is transferred from M2 to MI including the
requested address. By interleaving the memories comprising M2, the block transfer can
occur much more rapidly than the access rate of individual memory units. In a similar
fashion the cheapest but slowest memory M3 supports M2. Thus if most memory address
requests are supported by MI, the memory cache will appear to execute at nearly the speed
of MI but the overall cost of the memory system will be held down.

Quantity Speed

10 1000X

10Q 10X

1000Q 1X
Figure 7. Memory caching system.

If fl and f2 are the fraction of time Ml and M2, respectively, contain the requested address,
then the effective perfonnance speed of the memory cache will appear to be fllOOOX + fl
lOX + (l-fl - f2) X. Because sequential algorithms tend to access data in a unifonn fashion,
caching usually results in dramatic perfonnance ,improvements with little effort, however
caching considerations are often important to considering the placement of data in parallel
computing algorithms.
 407

3. Designing Parallel Algorithms

In order to achieve a high degree of speedup, data must be correctly positioned so that com-
putations may proceed without waiting for data transfers, i.e. data must be positioned "close"
to processing elements and data movement must occur concurrently with computations.
Sometim.es the design of a parallel algorithm necessarily entails solving a difficult combina-
torial optimization problem. To appreciate this aspect of parallel algorithm design. consider
how a simple Gauss-Seidel iteration scheme may be parallelized. A dependency graph
G=(V.A) is a useful and compact way to think about the communication requirements of an
iterative method. Each vertex in V represents an unknown and an arc (i.k) is present in A if
and only if function It depends on variable Xi' Finding a Gauss-Seidel update ordering that
maximizes speedup is equivalent to a graph coloring problem on the dependency graph [3].
This graph coloring problem is NP-complete which means finding an algorithm to optimally
solve a particular instance may entail great effort. The implications for parallel computing
are that designing good algorithm can necessarily involve significant work.
As an illustration of many of the aspects of parallel algorithm design. consider the
two-dimensional heat conduction equation on a rectangular region subject to boundary con-
ditions [3].

such that
T(XO,y)=!I(Y)

T(x loY) =f2(Y)

T(x.Yo)=h(x)

T(x.y 1) =!4(X)

Suppose the x and y axis are discretized into equally spaced intervals using points (i AxJ ~y),
where Ax is the size of the x-interval and ~y is the size of the y-interval and i=O •...• M and
j=O•...• N. Furthennore. denote the temperature at space point (i AxJ ~y) as TiJ . Finite differ-
ence approximation of the derivatives yields:
iJ 2 T T;-I.j-2T;J+Ti+I.j
iJx 2 = Al;2

iJ 2 T Ti.j - l - 2 T;J + Ti.j +l

iJy2 = ~y2

which may be combined:

Ti- 1J -2T;.j+Ti+I.j + Ti.j-I-2Ti.j+Ti.j+1 0
Al;2 ~y2

Thus TiJ may be computed from the four surrounding space points. The (symmetric) depen-
dency graph for the system of equations implied by this difference equation for M=N=4 is
shown in Figure 8. The circles represent the temperature at the space points and the edges
represent the symmetric dependency of adjacent space points for updated values in a Gauss-
 408

Seidel solution scheme. The dependency graph may be minimally colored with two colors
so that no two adjacent vertices possess the same color. White vertices may be updated
simultaneously in a Gauss-Seidel scheme. Likewise for black vertices. In an actual imple-
mentation of a Gauss-Seidel algorithm, a processor takes responsibility for updating an equal
number of white and black vertices in some localized region of space. Each processor
updates white vertices using the information currently stored by the processor for black ver-
tices, exchanges the new white vertex values with neighbors, and then computes new values
for black vertices. The process continues until appropriate accuracy is obtained.

(0,4) (4,4)

(0,0) (4,0)
Dependency Graph for the Two-Dimensional Coloring Scheme
Heat Conduction Equation (parallelization using
five processors)
Figure 8. Coloring scheme for finite difference grid.

Because of the localized communication, a number of parallel architectures would be

appropriate for the Gauss-Seidel algorithm implied by this dependency graph. A large
number of P.D.E.'s are amenable to Gauss-Seidel parallelization via coloring schemes
including the Vorticity Transport Equation, Poisson's Equation, Laminar How Heat-
Exchanger Equation, Telephone Equation, Wave· Equation, Biharmonic Equation, Vibrating
Beam Equation, and the Ion-Exchange Equation.
As a practical matter, some parallel computer vendors include optimized algorithms
for solving linear systems on their hardware. M:u-ket pressure should increase this trend.
Even so, for large scale problems, applications specialists are still forced to develop their
own tailored algorithms for solving linear systems. The need arises because the linear sys-
tem is very large, ill- conditioned, or the structure of the matrix lends itself to the develop-
ment of very fast special purpose algorithms. The large community of applied
 409

mathematicians and applications specialists in SIAM concern themselves with computational

research into quickly solving linear systems. A number of the leading linear algebra pack-
ages are being adapted to exploit parallelism [28], e.g. Linear Algebra Package (LAPACK).
In addition, a number of special purpose parallel computers already exist for solving linear
systems and more are under development [23]. From a process system point of view, the
ongoing research into parallel solution of linear systems is important since almost all process
system algorithms rely on the solUtion of linear systems.

3.1. Role of Special Purpose Algorithms

The only purpose of parallel computing is to improve algorithm performance. Tailoring
algorithms to particular problems also provides a compatible option for enhancing perfor-
mance. The form of this tailoring varies considerably, but generally speaking, exploiting
structure involves developing theoretical results and data structures specifically tuned to
exploiting problem characteristics. For example, very efficient solvers have been developed
for sparse systems of linear equations such as tridiagonal systems [13]. As a more detailed
measure of the value of special purpose algorithms, consider the assignment problem which
is a highly structured linear program.

minimize LLcijxij

.
LXij=l
i=lj=l

j=l •...• n

.
;=1

LXij=l i=l •...• n

j=!

X;j~ i,j=l, ...• n

The general purpose simplex method available in GAMS (BDMLP) requires 17.5 seconds of
CPU time when applied to an assignment problem of size n=40 (1600 variables) using a Sun
3/280 computer. On the other hand the specialized assignment problem algorithm of [2]
requires 0.08 seconds of CPU time to solve the same size problem. In parallel computing
terms, the special purpose algorithm achieves a speedup of 218.8 over the general purpose
approach. Furthermore, special purpose algorithms often possess easily exploitable parallel-
ism compared to general purpose approaches. Indeed the special purpose assignment prob-
lem algorithm of [2] was designed for parallel execution although it is among the best avail-
able assignment problem codes when executed in sequential mode. In principle, special pur-
pose algorithms may be developed for every problem. In practice, the expense of developing
special purpose algorithms is too great for most applications. This means the development
of methods for reducing the costs of special purpose algorithms, such as computer aided
software engineering techniques, is as important an area of research as parallel computing.
In fact, for enumerative algorithms such as branch and bound in mathematical programming
and alpha-beta search in Artificial Intelligence, exploiting parallel computing is clearly of
secondary importance when compared to problem structure exploitation since worst case
performance of these paradigms results in unreasonably long execution times on any foresee-
able computer architecture.
 410

3.2. Hierarchical ParalleUsm

In engineering computations, opportunities exist for exploiting parallelism at many different
levels of granularity. Indeed, the only way to circumvent Amdahl's law for most calcula-
tions is to exploit the parallelism available at these different levels. To illustrate the notion
of hierarchical parallelism, consider the solution of discrete optimization problems that arise
in a number of process systems applications. Exact solutions for the majority of these prob-
lems can only be obtained through an enumerative algorithm using some form of branch and
bound. Table 1 illustrates the levels of parallelism which can be exploited in a branch and
bound algorithm.

Table 1 - Levels of Parallelism for Branch and Bound

Mode of Task
parallelism granularity
competing very high
search trees
tree search high
algorithm moderate
components
function low
evaluations
machine very low
instructions

Parallelization at the three coarsest levels of granularity is the responsibility of the applica-
tions expert while the two finest granularity levels are usually the responsibility of computer
architects and compiler authors.
In the context of discrete optimization, the branch and bound tree (search tree)
represents a means of partitioning the set of feasible solutions. This partitioning is never
unique and quite often different partitioning strategies can lead to widely different search
times even for identical lower and upper bounding techniques although the best partitioning
strategy usually is not known in advance. The availability of multiple partitioning strategies
provides an excellent opportunity for exploiting parallelism. In particular, the availability of
k partitioning strategies implies that k different search trees can be created. A group of pro-
cessors can be used to explore each search tree (see below) and all search trees can be
exploited whenever one group proves optimality of a solution. As an example of this type of
parallelism letf(x) be the probability that a branch and bound algorithm requires time x and
g(x) be the probability that k competing trees require time x based on the same lower and
upper bounding technique. If the probability of finding a solution in a given time using a
 411

particular partitioning strategy is assumed to be independent with respect to other partition-

ing strategies then g(x) is related to [(x) (a reasonable assumption with a strong problem
relaxation).
g(x)=k [l-F(x)]k-l [(x)

where
x
F(x) = I[ (s) ds
o
For example, if
[(x)=_x_
12 e .Jx
which is a realistic probability distribution function for a branch and bound algorithm with
strong bounding techniques, then the expected speedup for searching k branch and bound
trees in parallel is given by

Ix[(x)dx
Speedup, St= ..:~--­
Ixg(x)dx
o

For example, S2=1.969 and S.=3.59 so that a parallelization strategy based on competing
search trees is quite effective. In practice, the speedups given by this analysis are conserva-
tive since competing search trees can share feasible solution information resulting in syner-
gism which can accelerate the search. Additional information on competing search trees
may be found in [30).
Each of the competing search trees may also be investigated in parallel. A great deal
of research has been done in this area. In particular, the papers by [20,31] summarize exist-
ing strategies for parallelizing the search of a single branch and bound tree. The work
1
reported in [22 outlines conditions under which a parallel search strategy results in super-
linear speedup , near linear speedup, and sublinear speedup. In intuitive terms, the oppor-
tunities for parallelism can be related to search tree structure. Superlinear speedup is possi-
ble when there is more than one vertex with a relaxation value equal to the optimal solution.
Near linear speedup is possible whenever the number of vertices divided by maximum tree
depth exceeds the number of processors involved in the search. Sublinear speedup results
whenever the number of vertices divided by maximum depth is less than the number of pro-
cessors.
Finer granularity methods are required to develop parallel algorithms for the com-
ponents of branch and bound such as the solution of problem relaxations, heuristics, and
branching rules. However, parallelization of the branch and bound components comple-
ments the concurrency that may be exploited at the competing search tree level and

t Speedup greater than the number of processors.

 412

individual search tree level. The results given in [2,30,31) suggest a speedup of at least 70
to be possible for combined parallelization at each of these levels ( a product of a factor of 1
to 12 at competing search tree level, factor of 10 to 50 at search tree level, and a factor of 7
to 14 at the component level) for an exact asymmetric traveling salesman problem algorithm
based on an assignment problem relaxation. The expected speedup from hierarchical paral-
lelization for this problem is dependent on the ultimate sizes of the search trees with the larg-
est search trees yielding the greatest speedup and virtually no speedup possible for very
small search trees.
As another example of a class of process systems algorithms that may exploit
hierarchical parallelism, consider decomposition methods for mathematical programming
problems, e.g. Bender's Decomposition [11] and the Outer Approximation Method [8]. Both
of these decomposition methods iterate through a sequence of master problems and heuristic
problems. The master problems invariably rely on enumerative methods, usually branch and
bound, so that the hierarchical parallelization described above is applicable in master prob-
lem solution. Furthermore, the sequence of iterations provides yet another layer of parallel-
ism to be exploited. Namely, each iteration of the decomposition method produces a set of
constraints to further strengthen the bounds produced from master problem solution. Groups
of processors could pursue different decomposition paths by, say, using different initial
guesses for complicating variables. Each of the decomposition paths would produce dif-
ferent cuts to strengthen the master problem relaxation and the processor groups could share
these cuts to create a pool from which they could all draw upon. Additional synergism
among processor groups is possible by sharing feasible solutions. Experimental work is
necessary to determine the degree to which parallel decomposition could succeed, however
experience suggests that the best decomposition algorithms require only a small number of
iterations. Thus parallelization of high quality decomposition methods would offer only a
small sp~dup. However, when this small speedup is amplified by parallelization at lower
algorithmic levels, the potential overall speedup promises to be quite high.

4. Distributed Computing
The last few years have seen a surge of interest in distributed computing systems. A distri-
buted computing system involves processors a large physical distance apart (say more than
10 meters) and interprocessor communication delays are unpredictable. A set of computers
on a corporate or academic site connected using Ethernet technology [39] provides an exam-
ple of a distributed computing system. Distributed computing systems offer the greatest per-
formance potential and offer the most general view of parallel computing systems.
Engineering computations do not tend to be homogeneous in time in the sense that
the workload alternates between disparate types of calculations (e.g. calculation of physical
properties, construction of a Hessian matrix, solution of a linear system of equations, etc.).
Computational experience over a large number of algorithms shows that different types of
calculations yield varying degrees of processor efficiency for anyone architecture. Thus no
one parallel computing architecture is ideally suited for sophisticated engineering calcula-
tions. However, networking technology is evolving to the point where it will soon be
 413

possible to decompose a complex calculation into components, distribute the components to

the appropriate architecture for calculations, and the reassemble the final result in an
appropriate location. Such distributed computing could become a dominant mode of paral-
lelism. This leads to the notion of network based parallelism which is defined as the applica-
tion of locally distributed and widely distributed computers in the cooperative solution of a
given problem over short time scales. Network based parallelism is currently under investi-
gation at a number of locations within the United States, e.g. [19,41] and preliminary pro-
gress suggests that the necessary communication hardware could become more widely avail-
able by the middle of this decade with routine application possible by early next century. Of
course existing computer networks can support parallelism, however it is severely limited in
terms of the amount of information that can be transmitted and the number of algorithms that
can simultaneously be supported. As mentioned above, bandwidth and latency are the prin-
ciple factors in the calculus of network based parallelism. The principle challenge is to con-
struct algorithms that mitigate latency, which cannot be circumvented due to the fundamen-
tal speed of light, and exploit bandwidth, which can be made arbitrarily large. The ideal net-
work algorithms communicate infrequently, although when they do, they may exchange
enormous quantities of information.

4.1. Controlling Network Communication

The key consideration in the design of network algorithms is how to control communication.
The primary choice lies with the degree of control which the algorithm designer wishes to
exercise over communication. At the lowest level, algorithms may be implemented using
low level paradigms such as TCPIIP socket streams, datagrams, and remote procedure calls
[39] which tend to be tedious but offer a high degree of control. At a higher level, packages
such as ISIS (Cornell University) and C-Linda (Scientific Computing Associates) are avail-
able which hide network communication details through abstract paradigms. The principle
drawback of these paradigms is that one must use a given communication model which may
only offer awkward support for an algorithm. Existing research promises to offer a number
of support mechanisms for network based parallelism. In particular research is proceeding
on developing virtual memory across machines in a network whereby a process on one
machine could effortlessly reserve and use memory on any of a number of machines. A
number of vendors and academic researchers are exploring automatic migration of computa-
tional processes through which a process spawned on one machine would quickly locate and
use unburdened machines [40]. Provided efficiency issues can be adequately addressed, such
paradigms offer great promise to application specialists for implementing network algo-
rithms. Application specific development environments are another area being explored to
aid in the implementation of network algorithms. For example, as discussed above, parallel
computing offers great promise in reducing the execution times of branch and bound compu-
tations [32]. However, the time and effort needed to parallelize algorithms exacerbates the
already arduous task of algorithm development. This has prevented the routine use of paral-
lel and distributed computers in solving combinatorial optimization problems. Given that all
branch and bound algorithms can utilize the same mode of parallelism, tools can be
specifically developed to reduce the burden associated with designing and implementing
 414

branch and bound algorithms in a distributed environment [18J.

As an example of network based parallelism using existing communication technol-
ogy, consider the computationally intense task of multiplying two fully dense matrices to
produce a third matrix, i.e. C=AB. This calculation can be parallelized in a number of ways
but, for simplicity, consider distributing matrix A and B to each of a number of machines on
a network. The task of computing C can then be partitioned among the machines so that
each machine computes a portion of C according to its relative capability (see Figure 9).

C=A *B A,B,C E Ff n=700

Computation Schematic:

Sparcstation 1

Sparcstation 1

Generate Sparcstation 1+ Collect

Matrix Submatrices
Sparcstation Sparcstation 1+
Sparcstation 1

Sparcstation 1

Figure 9. Distributed matrix multiplication scheme.

For square double precision matrices of size 700, the sequential computing time on a Sun
Microsystem Sparcstation 1+ is 751.2 seconds. Performance for the network based algo-
rithm depicted in Figure 9 is given in Table 2. In Table 2, the critical path is defined as that
string of computations which controlled the wall clock time of the algorithm. In particular.
the transmission of the 7.84 megabytes of matrix data (A and B) to each of the four Sparc Is
from the Sparc 1+, the portion of matrix multiplication done on the Sparc 1+, and the collec-
tion of the resulting matrix back on the Sparc 1+ controlled the wall clock execution time.
The overall speedup for the algorithm computed,as the parallel wall clock execution time
divided by the Sparcstation 1+ sequential execution time is 3.72 yielding an efficiency of
93% using four processors. This simple example points out a difficulty of using the usual
definitions of speedup and efficiency for network calculations.
 415

Table 2 - Distributed Matrix Computation Results

Operation time (sec)
Multiplication Time (Sparc 1) 186.7
Multiplication Time (Spare 1+) 151.0
Latency (along critical computation path) 0.50
Transmission Time (along crit. path) 14.46
Multiplication Time (along crit. path) 186.7
Wall Clock Execution Time 201.7

Namely, the Sparc 1+ is approximately 33% faster than a Sparc I while the normal definition
of speedup and efficiency assume all processors to be of equal capability. The conservative
approach is to perform speedup and efficiency calculations using the fastest processor to col-
lect sequential times.

5. Additional Reading
The last several years have seen a dramatic increase in the number of publications addressing
various parallel computing issues [5]. Th~ textbook [3] addresses a number of issues
relevant to the design of parallel algorithms. A number of references discuss the trends in
computer architecture [14,29,38] and implementation technology [6,24,37]. Molecular
simulation [9] and transport phenomena [16,35] offer a number of opportunities for applying
parallel computing. In addition to the areas discussed above, there has been considerable
research into the parallelization of continuous optimization methods [25]. A number of
excellent references exist for the parallelization of generic numerical methods [13,27,28]
and parallel programming languages [1,4].

References
1. Babb. R. G., Programming Parallel Processors, Addision·Wesley, 1988.
2. Balas, E., D. L. Miller, I. F. Pelrny, and P. Toth, "A Parallel Shortest Path Algorithm for the Assign-
ment Problem," Journal of the Associationfor Computing Machinery, vol. 38, pp. 985-1004,1991.
3. Bertsekas, D. P. and I. N. Tsitsildis, Parallel and Distributed Computation, Prentice Hall, Englewood
Cliffs, 1989.
4. Brawer, S.,lntroduction to Parallel Programming, Academic Press, 1989.
5. Corcoran, E., "Calculating Reality," Scientific American, vol. 264, no. I, pp. 100-109, 1991.
6. Corcoran, E., "Diminishing Dimensions," Scientific American, vol. 263, no. 5, pp. 122-131, 1989.
7. Defense, U. S. Department of, Critical Technologies Plan (Chapter 3), 1990.
8. Duran, M. A. and 1. E. Grossmann, "An Outer-Approximation Algorithm for a Class of Mixed-Integer
Nonlinear Programs," Mathematical Programming, vol. 36, pp. 307-339, 1986.
9. Fincham, D., "Parallel Computers and Molecular Simulation," Molecular Simulation, vol. 1, pp. 1-45,
1987.
10. Flynn, M. 1., "Very High-Speed Computing Systems," Proc. IEEE, vol. 54, pp. 1901-1909, 1966.
11. Geoffrion, A. M., "Generalized Benders Decomposition," Journal o/OptimiZalion Theory and Appli·
cations, vol. 10, no. 4, pp. 237-260, 1972.
12. Ghezzi, c., Fundamentals of Software Engineering, Prentice-Hall, Englewood Cliffs, NJ, 1991.
 416

13. Golub, G. H. and C. F. Van Loan, Matrix Computations (2nd edition), John Hopkins, Baltimore, 1989.
14. Hillis, W. D., The Connection Machine, MIT Press, 1985.
15. Hwang, K. and F. A. Briggs, Computer Architecture and Parallel Processing, McGraw-Hili, New
York,1984.
16. Jespersen, D. C. and C. Levit, "A Computational Fluid Dynamics Algorithm on a Massively Para!lel
Computer," International Journal of Supercomputing Applications, vol. 3, pp. 9-27,1989.
17. Kim, S. and S. J. Karrila, Microhydrodynamics: Principles and Selected Applications, Butterwont·
Heinemann, Boston, 1991.
18. Kudva, G. and J.F. Pekny, "DCABB: A Distributed Control Architecture for Branch and Bound!
Calculations", Computers and Chemical Engineering, vol. 19, pp. 847-865,1995.
19. Kung, H. T., E Cooper, and M. Levine, "Gigabit Nectar Testbed," Corporation for National Research
Initiatives Grant Proposal (funded), School of Computer Science, Carnegie Mellon University, 1990.
20. Lavallee, I. and C. Roucairol, "Parallel Branch and Bound Algorithms," MASI Research Repon,
EURO VII, Bologna, Italy, 1985.
21. "Leebaert, D., Technology 2001, MIT Press, 1991.
22. Li, G. and B. W. Wah, "Computational Efficiency of Parallel Approximate Branch-and-Bound Algo-
rithms," International Conference on Parallel Processing, pp. 473-480, 1984.
23. McCanny, 1. F., J. McWhirter, and E. E. Swartzlander, Systolic array processors: contributions by
speakers at the International Conference on Systolic Arrays, held at Killarney, Co. Kerry, Ireland,
1989, Prentice Hall, 1989.
24. Meindl, J. D., "Chips for Advanced Computing," Scientific American, vol. 257, no. 4, pp. 78-89,
1987.
25. Meyer, R. R. and S. A. Zenios, "Parallel Optimization on Novel Computer Architectures," Annals of
Operations Research, vol. 14, 1988.
26. Miller, D. L., "Parallel Methods in Combinatorial Optimization," in Invited Lecture, Purdue Univer-
sity, West Lafayette, lN, 1991.
27. Modi, J. J., Parallel Algorithms and Matrix Computation, Oxford University Press, 1988.
28. Ortega, J. M., Introduction to Parallel and Vector Solution ofLinear Systems, Plenum Press, 1988.
29. Patterson, D. A., Computer Architecture: A Quantitative Approach, Morgan Kaufman Publishers, San
Mateo, CA, 1990.
30. Pekny, J. F., "Exact Parallel Algorithms for Some Members of the Traveling Salesman Problem Fam-
ily," Ph. D. Dissertation, Carnegie Mellon University, Pittsburgh, PA 15213, 1989.
31. Pekny, J. F. and D. L. Miller, "A Parallel Branch and Bound Algorithm For Solving Large Asym-
metric Traveling Salesman Problems," MathemaIical Programming, vol. 55, pp. 17-33, 1992.
32. Pekny, J. F., D. L. Miller, and G. Kudva, "An Exact Algorithm for Resource Constrained Sequencing
With Application to Production Scheduling Under an Aggregate Deadline," Computers and Chemical
Engineering, vol. 17, pp. 671-682, 1993.
33. Pekny, J. F., D. L. Miller, and G. J. McRae, "An Exact Parallel Algorithm for Scheduling When Pro-
duction Costs Depend on Consecutive System States," Computers and Chemical Engineering, vol. 14,
pp. 1009-1023, 1990.
34. Pita, J., "Parallel Computing Methods in Chemical Engineering," in Invited Lecture, Purdue Univer-
sity, West Lafayette, lN, 1991.
35. Saati, A., S. Biringen, and C. Farhat, "Solving Navier-Stokes Equations on a Massively Parallel Pro-
cessor: Beyond the One Gigaflop Performance," International Journal of Supercomputing Applica-
tions, vol. 4, pp. 72-80, 1990.
36. Sierra, H. M., An Introduction to Direct Access Storage Devices, Academic Press, 1990.
37. Stix, G., "Second-Generation Silicon," Scientific American, vol. 264, no. I, pp. 110-111, 1991.
38. Stone, H. S., High-PerforTlUlnce Computer Architecture, Addison-Wesley, Menlo Park, 1987.
39. Tanenbaum, A. S., Computer Networks, Prentice Hall, Englewood Cliffs, 1981.
40. Tazelaar, J. M., "Desktop Supercomputing," BITE, vol. 15, no. 5, pp. 204-258,1990.
41. Tesler, L. G., "Networked Computing in the 19905," Scientific American, vol. 265, no. 3, pp. 86-93,
1991.
42. Weiser, M., "The Computer for the 21st Century," Scientific American, vol. 265, no. 3, pp. 94-104,
1991.
Optimization

Arthur W. Westerberg

Dept. of Chemical Engineering and the Engineering Design Research Center, Carnegie Mellon University,
Pittsburgh, PA 15213, USA

Abstract: This paper is a tutorial on optimization theory and methods for continuous variable
problems. Its main purpose is to provide geometric insights. We introduce necessary
conditions for testing the optimality of proposed solutions for unconstrained, equality
constrained and then inequality constrained problems. We draw useful connections between
constrained derivatives and Lagrange theory. The geometry of the dual is exploited to explain
its properties; we use the ideas to derive the dual for linear programming. We cover pattern
search and then show the key ideas behind generalized reduced gradient and sequential
quadratic programming methods. Using earlier insights, linear programming becomes a special
case of nonlinear programming; we readily explain the Simplex algorithm. The paper ends
with presentations on interior point methods and Benders' decomposition. For nonlinear
problems, the paper deals only with finding and testing oflocal solutions.

Keywords: optimization, constrained derivatives, Lagrange multipliers, Kuhn-Tucker

multipliers, generalized dual, pattern search, generalized reduced gradient, sequential quadratic
programming, linear programming, interior point algorithms, Benders' decomposition.

Introduction

Optimization should be viewed as a tool to aid in decision malcing. Its purpose is to aid in the
selection of the better values for the decisions which can be made by the person in solving a
problem. To formulate an optimization problem, one must resolve three issues. First, one
must have a representation of the artifact which can be used to detennine how the artifact
performs in response to the decisions one makes. This representation may be a mathematical
model or the artifact itself Second, one must have a way to evaluate the performance - an
objective function - which is used to compare alternative solutions. Third, one must have a
 418

method to search for the improvement. In this paper, we shall be concentrating on the third
issue, the methods one might use. The first two items are difficult ones, but discussing them at
length is outside the scope of this paper.
Example optimization problems are: (1) determine the optimal thickness of pipe insulation;
(2) find the best equipment sizes and operating schedules for the design of a new batch process
to make a given slate of products; (3) choose the best set of operating conditions for a set of
experiments to determine the constants in a kinetic model for a given reaction; (4) find the
amounts of a given set of ingredients one should use for making a carbon rod to be used as an
electrode in a arc welder.
For problem (1), one will usually write a mathematical model of how insulation of varying
thickness restricts the loss of heat from a pipe. Evaluation requires one develop a cost model
for the insulation (a capital cost in dollars) and the heat which is lost (an operating cost in
dollarslyr). Some method is required to permit these two costs to be compared such as a
present worth analysis. Finally, if the model is simple enough, the method one can use is to set
the derivative of the evaluation function to zero with respect to wall thickness to find candidate
points for the its optimal thickness. For problem (2), selecting a best operating schedule
involves discrete decisions which will generally require models that have integer variables.
Such problems will be discussed at length in the paper by Grossmann in this AS!.
It may not be possible to develop a mathematical model for problem (4) as we may not
know enough to characterize the performance of a rod versus the amounts of the various
ingredients used in its manufacture. Here, we may have to manufacture the rods and then
judge them by ranking the rods relative to each other, perhaps based partially or totally on
opinions. Pattern search methods have been devised to attack problems in this class; we shall
consider them briefly later.
For most of this paper, we shall assume a mathematical model is possible for the problem
to be solved. The model may be encoded in a subroutine and be known to us only implicitly,
or we may know the equations explicitly. A general form for such an optimization problem is:
minF =F(z)
S.t. h(z) = 0
g(z) ~ 0
where F represents a specified objective function that is to be minimized. Functions hand g
represent equality and inequality constraints which must be satisfied at the final problem
solution.
 419

Variables z are used to model such things as flows, mole fractions, physical properties,
temperatures and sizes. The objective function F is generally assumed to be a scalar function,
one which represents such things as cost, net present value, safety or flexibility. Sometimes
several objective functions are specified (e.g. minimize cost while maximizing reliability); these
are commonly combined into one-function, or else one is selected for the optimization while
the others are specified as constraints. Equations h(z)=O are typically algebraic equations,
linear or nonlinear, when modeling steady-state processes, or algebraic coupled with ordinary
and/or partial differential equations when optimizing time varying processes. Inequalities g(z)

~ 0 put limits on the values variables can take such as a minimum and maximum temperature
or they restrict one pressure to be greater than another.
One set of issues about practical optimization we shall be unable to address but which is
critical if one wishes to solve large problems are numerical analysis issues. We shall not
discuss what happens when a matrix is almost singular or how to factor a sparse matrix or how
to partition and precedence order a set of equations. Another topic we shall completely
exclude is the optimization of distributed systems. This topic is in Biegler et al. [4]. A good
text on this topic is by Bryson and Ho [6]. Finally we shall not consider so-called genetic
algorithms or those based on simulated annealing. These latter are best suited for solving
problems involving a very large number of discrete decisions, the topic of one of the papers by
Grossmann.
For further reading on optimization, readers are directed to the following books [17,31].

Packages

There are a number of packages available for optimization. Following is a list of some of them.
(1) Frameworks
GAMS. This framework is commercially available. It provides a uniform language
to access several different optimization packages, many of them listed below. It
will convert the model as expressed in "GAMS" into the form needed to run the
package chosen.
AMPL. This framework is by Fourier and co workers [14] at Northwestern
University. It is well suited for constructing complex models.
 420

ASCEND. This framework is our own. Featuring an object-oriented modeling

language, it too is well suited for constructing complex models.

(2) Algebraic optimization with equality and inequality constraints

SQP. A package by Biegler in our Chemical Engineering Department.

MINOS5.4. A package available from Stanford Research Institute (affiliated with

Stanford University). This package is the state of the art for mildly nonlinear
programming problems.

GRG. A package from Lasdon at the U. of Texas, Dept. of Management Science.

(3) Linear programming

MPSXfromIDM
SCICONIC from the company of that name.

MINOS5.4

Cplex. A package by R. Bixby at Rice University and Cplx, Inc.

Most current commercial codes for linear programming extend the Simplex algorithm, and
they can typically handle problems with up to 15,000 constraints.

Organization of Paper

Several sections of this paper are based on a chapter on optimization which this author
prepared with Biegler and Grossmann and which appears in Ullmann's Encyclopedia of
Industrial Chemistry [4]. In that work and here, we partition the presentation on optimization
into two parts. There is the theory needed to determine if a candidate point is an optimal one,
the theme ofthe next section of the paper. The second part covers various methods one might
use to find candidate points, the theme of the subsequent section. Our goal throughout this
paper is to provide physical insight.

In the next section, we state conditions for a point to be a local optimum for an
unconstrained problem, then for an equality constrained one, and finally for an inequality
constrained one. To obtain the conditions for equality constrained problems, we introduce
constrained derivatives as they directly relate the necessary conditions for the constrained
problem to those for the unconstrained one. We show a very nice way to compute them. We
next introduce Lagrange multipliers and the associated Lagrange function. Lagrange theory
 421

and constrained derivatives are elegantly related. This insight aids in explaining the methods
we shall describe in the following section to find candidate points.

Contrary to most presentations, we shall present linear programming as a special case of

nonlinear programming, making it possible to explain why the Simplex algorithm is as it is. An
exciting development in the solving of Linear Programming are the recent interior point
algorithms, which we shall discuss.
We end this paper by looking at Benders' decomposition as a method to solve problems
having special structure.

Conditions for Optimality

We start by stating both necessary and sufficient conditions for a point to be the minimum for
an unconstrained problem.

Local Minimum Point for Unconstrained Problems

Consider the following unconstrained optimization problem

Min (F(u) I uERr}

u
If F is continuous and has continuous first and second derivatives, it is necessary that F is
stationary with respect to all variations in the independent variables u at a point u which is
proposed as a minimum to F, i.e.,

of =0, i=I,2, .. r or VuF=O at u =li (I)

Ou;
These are only necessary conditions as point u may be a minimum, maximum or saddle point.

Sufficient conditions are that any local move away from the optimal point u gives rise to
an increase in the objective function. We expand F in a Taylor series locally around our
candidate point u up to second order terms

+ ".

= F(u) + VuFTI li (u-li) + t (u-li) T iuuFl li (u-li) + ...

 422

Ifu satisfies necessary conditions (1), the second term disappears in this last line. For this case
we see that sufficient conditions for the point to be a local minimum are that the matrix of

second partial derivatives V~ is positive definite. This matrix is synunetric so all of its
eigenvalues are real; to be positive definite, they must all be greater than zero.

Constrained Derivatives - Equality Constrained Problems

Consider minimizing our objective function F written in terms of n variables z and subject to m
equality constraints h(z)=O, i.e.,

Min (F(z) I h(z) = 0, zERn, h:Rn --+ Rm) (2)

z
We wish to test point z to see if it could be a minimum point. It is necessary that F is
stationary for all infinitesimal moves for z that satisfy the equality constraints. We discover the
appropriate necessary conditions, given this goal, by linearizing the m equality constraints
around z, getting

h(z + Az) = h(z) + VzhTI z 11

(3)

where Az = z - z.
We want to characterize all moves Az such that the linearized equality constraints remain at
zero. There are m constraints here so m of the variables are dependent leaving us with r=n-m
independent variables. Partition the variables Az into a set of m dependent variables I1x and
r=n-m independent variables l1u. Eqn (3), rearranged and then rewritten in terms of these
variables becomes

I1h=VxhTlz I1x+Vuh Tlz l1u=O

Solving for dependent variables I1x in terms of the independent variables l1u, we get

(4)

Note that, we must choose which variables are the dependent ones to assure that the Jacobian

matrix Vxh evaluated at our test point is nonsingular. This partitioning is only possible if the

rank: of the m by n matrix Vzh is of rank: m. Eqn (4) states that the changes in m dependent
variables x can be computed once we specify the changes for the r independent variables u.
Linearize the objective function F(z) in terms of the partitioned variables
 423

M=VXFTI z~ + VuFTI z~u

and substitute out variables ~ using eqn (4).

LW = {VxPT - VupT [VxhTJ-l VuhTfz ~u

r (5)
= (dE)T L\u = L {.dE} ~Uj
du Ah;O j;l dUj llh;Q

There is one term for each ~Ui in the row vector which is in the curly braces {}. These

terms are called constrained derivatives. They tell us how the objective function will change if
we change the independent variables Ui while changing the dependent variables Xi to keep the
constraints satisfied.
Necessary conditions for optimality are that these constrained derivatives are zero, i.e.,

'\.!lEd) = 0 , i=I,2, .. r
Uj llh;Q

An Easy Way to Compute Constrained Derivatives

Form the Jacobian matrix for the equality constraints h(z) augmented with the objective
function with respect to variables z at z.
VzhTI A &=0
z m rows
VzFTI A &=0
z I row
Note that, there are n variables z in these m+ I linearized equations. Perform a forward gauss

elimination on these equations including the last row (VzF ~ = 0 ) but do not pivot within that
row. One will select m pivots. Select them so the m by m pivoted portion of the matrix is
nonsingular. The pivoted variables are the dependent variables x for the problem, while the
unpivoted are the independent variables u. Fig. 1 shows the structure of the result. The
nonzero portion of the last row beneath the variables u contains exactly the numerical
evaluation for the constrained derivatives given in eqn (5). One can prove this statement by
carrying out the elimination symbolically and noting that this part of the matrix is algebraically
the constrained derivatives as noted.

Equality Constrained Problems - Lagrange Multipliers

Form a scalar function, which we shall term the Lagrange function, by adding each of the
 424

columns x columns u

rowsh

rowF

Fig. 1 Panitioning the Variables and Computing Constrained Derivatives

in a Single Step Using Gaussian Elimination

equality constraints multiplied by an arbitrary multiplier to the objective function.

m T
L{x,U,A.} = F(x, u) + L A.ihi(X,U) = F(x,u) + A. h(x,u)
i=l

At any point where the functions h(z) are zero, the Lagrange function equals the objective
function.
Next, write the stationarity conditions for L with respect to variables X, u and A..

VxLTlz = VxFTlz +A,TVh!lz =OT (6)

VuLTI z = VufTI z + A,TVh!1 z = OT (7)

VALTI z = hT{x,ll) =OT

Solve eqn (6) for the Lagrange multipliers

A.T = _VxFT[ Vh~]-1

(8)

and then eliminate these multipliers from eqn (7).

VulT = VuPT- VxFT[ vh~l VhJ = 0 (9)

We see by comparing eqn (9) to eqn (5) that Vu are equal to the constrained derivatives for
our problem, which, as before, should be zero at the solution to our problem. Also these
stationarity conditions very neatly provide us with the necessary conditions for optimality of an
equality constrained problem.
 425

Lagrange multipliers are often referred to as shadow prices, adjoint variables or dual
variables, depending on the context. Assume we are at an optimum point for the our problem.
Perturb the variables such that only constraint hi changes. We can write

6L = 6F + l..i6hi = 0
which is zero because, as just shown, the Lagrange function is at a stationary point at the
optimum. Solving for the change in-the objective function
6F=- ~Ah

The multiplier tells us how the optimal value of the objective function changes for this small
change in the value of a constraint while holding all the other constraints at zero. It is for this
reason they are often called shadow prices.

Equality and Inequality Constrained Problems - Kuhn-Tucker Multipliers

In the previous section, we considered only equality constraints. We now add inequality
constraints and examine how we might test a point to see if it is an optimum. Our problem is
Min (F(z) I h(z) = 0, g(z) ~ 0, ZE R", F:R" ~ R 1, h:R" ~ Rffi, g:R" ~ RP}
z
The Lagrange function here is similar to before.

L(z,I..,Jl) '" F(z) + I.. Th(z) + JlTg(z)

only here, we also add each of the inequality constraints gi(Z) multiplied by what we shall call a
Kuhn-Tucker multiplier, Ili. The necessary conditions for optimality, called the Karush-Kuhn-
Tucker conditions for inequality constrained optimization problems, are

VzL Iz Iz Iz
= VzF + Vzh I.. + V ~ Iz 11 = 0
VJL=h(z)=O
g(z) ~ 0
lligi(Z) = 0, i=1,2, ... p (10)
Ili ~ 0 , i=1,2, ... p
Conditions (10), called complementary slackness conditions, state that either the constraint
gi(Z)=O and/or its corresponding multiplier Ili is zero. If constraint gi(Z) is zero, it is behaving

like an equality constraint, and its multiplier Ili is exactly the same as a Lagrange multiplier for
an equality constraint. If the constraint is away from zero, it is not a part of the problem and
should not affect it. Setting its multiplier to zero removes it from the problem.
 426

As our goal is to minimize the objective function, releasing the constraint into the feasible
region must not decrease the objective function. Using the shadow price argument above, it is
evident that the multiplier must be nonnegative [24].

Constraint Qualifications

The necessary conditions above will not hold if, for example, two nonlinear inequality
constraints form a cusp as shown in Fig. 2, and the optimum is exactly at that point. The
optimum requires both constraints for its definition, even though they are both collinear at the
solution. The equation
VF + 1.11 Vg\ + 1l2Vgl = 0
which states that the gradient of the objective function can be written as a linear combination
of the gradients of the constraints can be untrue at this point, as Fig. 2 illustrates. One usually
states that the constraints have to be independent at the solution when stating the Karush-
Kuhn-Tucker conditions, a constraint qualification.

o~imum point
,

Vg
2

Fig.:Z Necessary Conditions Can Fail at a Cusp

Kuhn-Tucker Sufficiency Conditions

Sufficiency conditions to assure a Kuhn-Tucker point is a local minimum point require one to
prove that the objective function will increase for any feasible move away from such a point.
To carry out such a test one has to generate the matrix of second derivatives of the Lagrange
function with respect to all the variables z evaluated at z. The test is seldom done as it
requires too much work.

The Generalized Dual

Consider the following problem, which we shall call the primal problem.
 427

F" '" F(z") = Min (F(z) I h(z) = 0, ZES) (11)

z

Let us write the following "restricted" Lagrange function for this problem. We call it restricted

as it is defined only for the restricted set of values ofzES.

T
L(z, A) = {F(z) + A h(z) I ZE S}

Pick a point zES and plot its corresponding point F(z) versus h(z}. Repeat for all ZES,
getting the region R shown in Fig. 3. Region R is defined as

R = {(F(z),h(z) I for all ZE S }

Pass a hyperplane (line) through any point in R with a slope of -A, as illustrated. The intercept
where h(z) = 0 can be seen to be exactly equal to the Lagrange function for this problem.

F(z)

-------t--'~-------- h(z)
L(z,A) = F(z) - (-A) h(z) = F(z) + Ah(z)

Fig. 3 The Lagrange Function in the Space ofF(z) vs. h(z)

If we minimize the Lagrange function over all points in R for a fixed slope -A, we obtain
the minimum intercept possible for that hyperplane, which is illustrated in Fig. 4. Note that this
hyperplane supports Gust touches) the region R, with all ofR being on only one side of it. This
support function

D(A)=Min {L(Z,A) IZES}

z

is known as the generalized dual function for our original problem. The minimization is carried
out over z, and the global minimum must be found.
By examining the geometric interpretation for the dual, we immediately see one of its most
important properties. It is always below the region R on the intercept where h(z) = o. As such
it must be a lower bound for the optimal value for our objective junction for the primal
problem defined by eqn (11), namely
 428

F z)

F+, the minimum solution

to the primal problem

h(z)

DO.) = min {L(z, l) I z in S}

z
Fig. 4 A Geometrical Interpretation of The Generalized Dual Function, D(A.)

D(A) ~ F*
We can now vary the slope A to find the maximum value that this dual function can attain.
D * == D(A *) = Max {D(A) I AERm} ~ F * == F(z*) = Min {F(z) I h(z) = 0, ZE S)
A Z

If a hyperplane can support the region R at the point where h(z) = 0, then we note that this
maximum exactly equals the minimum value for our original objective at its optimal solution.
It may be that the region R cannot be supported at F"', in which case the optimum of the dual
is always less than the optimum for the primal, as illustrated in Fig. 5. Note there are two
support points (at least) ifthis is the case, neither of which satisfies the equality constraints for
the primal problem.
Further properties of dual: Examining its geometrical interpretation, we can note some
further important properties of the dual.
First, if the region R fails to cover any part of the axis where h(z) = 0, then the primal
problem can have no solution. It is infeasible. For this case, we can find a hyperplane to
support the region R that will intersect this vertical axis at any value desired. The maximum
value will be positive infinity. If the original problem is infeasible, the dual is feasible but
unbounded Conversely, ifR covers any part of the axis, the dual cannot be unbounded, so
the previous statement is really an if and only if statement.
Second, consider the case we show in Fig. 6 where the region R is unbounded below
but where it does not cover the negative vertical axis. We note there are multipliers(slopes),
as the hyperplane labeled 1 illustrates, that will lead from region R to an intersection with
vertical axis (h(z)=O) at negative infinity; i.e., for those values the dual function is unbounded
 429

F(z)

F*, the minimum solution

to the primal problem

h(z)
D*, the maximum value
of the dual function
Fig. 5 A Nonconvex Region R with No Support Hyperplane at the Solution to the Primal Problem

below. However, there are also multipliers where the support plane is finite as hyperplane 2
illustrates. Since the dual problem is to search over the hyperplane slopes to find a maximum
to the intercept with the vertical axis, we can eliminate looking over multiplier values where
the dual function is unbounded below. We shall define the dual function as being infeasible
for these values of the multipliers.

I
Finite dual function
RegionR
unbounded in
this direction

I
Fig. 6 Case Where Region R Is Unbounded In Negative F(Z) Direction. A Support Hyperplane
With Slope Parallel To I Yields A Dual Solution Of Negative Infinite
While One With Slope Of 2 Yields A Finite Dual Solution.
 430

Third, if the region R in Fig. 6 covers the entire negative vertical axis where h(z) is
zero, then the dual function is negative infinity for all values of the m~ltipliers. Continuing our
ideas just above, the dual is infeasible everywhere. Thus we have a symmetry: if the primal is
infeasible, the dual is unbounded. If the primal is unbounded, the dual is infeasible.
While it is not immediately obvious, the dual of the dual is closely related to the primal
problem. It corresponds to an optimization carried out over the convex hull of the region R.
It is left as an exercise for the reader to prove this statement.

Example: Let us find the dual for the following problem.

Min {cTu IAu ~ b, u ~ O}
u

We can introduce slack variables and rewrite this problem as follows

Min
II,S
{cTulb+s-Au=O, u, s~O}

The constrained Lagrange function can then be written as

T T T T
L(u, s, A) = {CTU+A (b+s-Au)lu,s~O}={(cLA A)u+A S+A b)lu,s~O}
from which we derive the dual function
D(A)=Min L(U,S,A)
II,S

The minimization operation can be carried out term by term, giving

Min (Cj - ~TA.) _I 0 ifCj-ATAj~O \ d Mi ~ _I 0 inj~O \

/\, e>j Ui -
\ -00 ifCj-ATAj<O I an n /\"S'
JJ
-
\ -00 inj<O I
where Ai is the i-th column of A. For the primal problem to have a bounded solution, we can
eliminate looking over multipliers where the dual is unbounded below (see Fig. 6 earlier).
ATA'5. c and A~ 0

The dual function becomes

D(A) = Min
II,S
L(U,S,A) =

= {O + 0 + ATb IAT A '5. c, A ~ O} = {bTAI AT A '5. C, A ~ O}

and the. dual optimization problem

(12)
We can show that R is convex by noting if the points (u(l),s(l» and (u(2),s(2» are both in
R then so is their convex combination a(u(I),s(I» + (1-a.)(u(2),s(2» and that it maps
precisely into the point
 431

{acTu(1) + (1-a)cTu(2), a(b + s(l) - Au(l)) + (1-a)(b + s(2) - Au(2»}

in R. If R is convex, then it has a support hyperplane everywhere. Thus the value of the
objective function for both the primal and the dual will be equal at the solution.

Strategies of Optimization

The theory just covered can tell us if a candidate point is, or more precisely, is not the optimum
point, but how do we find candidate point? The simplest strategy is to place a grid of points
throughout the feasible space, evaluating the objective function at every grid point. If the grid
is fine enough, then the point yielding the highest value for the objective function can be

selected as the optimum. 20 variables gridded over only 10 points would take place 1020
points in our grid, and, at one nanosecond per evaluation, it would take in excess of four
thousand years to carry out these evaluations.
Most strategies limit themselves to finding a local minimum point in the vicinity of the
starting point for the search. Such a strategy will find the global optimum only if the problem
has a single minimum point or a set of "connected" minimum points. A "convex" problem has
only a global optimum.

Pattern Search

Suppose the optimization problem is to find the right mix of a given set of ingredients and the
proper baking temperature and time to make the best cake possible. A panel of judges can be
formed to judge the cakes; assume they are only asked to rank order the cakes and that they
can do that task in a consistent manner. Our approach will be to bake several cakes and ask
the judges to rank order them. For this type of problem, pattern search methods can be used
to find the better conditions for manufacturing the product. We shall only describe the ideas
behind this approach. Details on implementing it can be found in Umeda and Ichikawa [35].
The complex method is one such pattern search method, see Fig. 7. First form a "complex"
of at least r+1 (r = 2 and we used 4 points in Fig. 7) different points at which to bake the cakes
by picking a range of suitable values for the r independent variables for the baking process.
Bake the cakes and then ask the judges to identify the worst cake.
For each independent variable, form the average value at which it was run in the complex.
Draw a line from the coordinates of the worst cake through the average point - called the
 432

centroid - and continue on that line a distance that is twice that between these two points.
This point will be the next test point. First decide if it is feasible. If so bake the cake and
discover if it leads to a cake that is better than the worst cake from the last set of cakes. If it is
not feasible or it is not better, then return half the distance toward the average values from the
last test and try again. If it is better, toss out the worst point of the last test and replace it with
this new one. Again, ask the judges to find the worst cake. Continue as above until the cakes
are all the same quality in the most recent test. It might pay to restart at this point, stopping
finally if the restart leads to no improvement. The method takes large steps if the steps are
being successful in improving the recipe. It collapses onto a set of point quite close to each
other otherwise. It works reasonably well, but it requires one to bake lots of cakes.

1
• 2 worst

•
3

•
4

Fig. 7 Complex Method, a Pattern Search Optimization Method

Generalized Reduced Gradient (GRG) Method

We shall develop next a method called the generalized reduced gradient (GRG) method for
optimization. We start by developing a numerical approach to optimize an unconstrained
problem.

Optimization of Unconstrained Objective: Assume we have an objective function F which

is a function of independent variables Uj, i = Lr. Assume we can have a computer program
which, when supplied with values for the independent variables, can feed us back both F and its
derivatives with respect to each Uj. Assume that F is we)) approximated as an as yet unknown
quadratic function in u.
 433

F ~ a+ bTU +luTQu
2
where a is a scalar, b a vector and Q an TXT symmetric positive definite matrix. The gradient of
our approximate function is

VuF=b+Qu

which, when we set it to zero, allows to find an estimate for it minimum

u=_Q-1b (13)

We do not know Q and b at the start so we can proceed as follows. b contains r unknown
coefficients and Q another r(r+l)I2. To estimate b and Q, we can run our computer code
repeatedly, getting r equations each time - namely

(VuFXl) =b + Qu(1)
(VuFX2) = b + Qu(2)
(14)
(VuFXt) = b + Qu(t)

As soon as we have written as many independent equations from these computer runs as there
are unknown coefficients, we can solve these linear equations for band Q. A proper choice of
the points u(i) will guarantee getting independent equations to solve here.
Given b and Q, eqn (13) provides us with a new estimate for u as a candidate minimum
point. We run the subroutine again to obtain the gradient ofF at this point. If the gradient is
essentially zero, we can stop; we have a point which satisfies the necessary conditions for
optimality. If not, we write equations in the form of (14) for this new point, add them to the
set while removing the oldest set of equations. We solve these equations for band Q and
continue until we are at a minimum point. Ifremoval of the oldest equations from the set (14)
leads to a singular set of equations, then different equations have to be selected for removal.
We can keep all the older equations, with the new ones added to the top of the list. Pivoting
can be done by proceeding down the list until a nonsingular set of equations is found. We use
the older equations only if they have to be. Also, since only one set of equations is being
replaced, clever methods are available to find the solution to the equations with much less
work than is required to solve the set of equations the first time [10,34].

Quadratic Fit for the Equality Constrained Case: We wish to solve a problem of the form
of eqn (2). We proceed as follows. For each iteration k:
1. Enter with values provided for variables u(k).
 434

2. Given values for u(k), solve equations h(x,u) = 0 for x(k). These will be m equations in m
unknowns. If the equations are nonlinear, solving can be done using a variant of the
Newton-Raphson method.
3. Use eqns (8) to solve for the Lagrange multipliers, A.(k). If we used the Newton-Raphson
method (or any or several variants to it) to solve the equations, we will already have

generated the Jacobian matrix VJh Iz(k) and its L\U factors so solving eqns (8) requires
very little effort.

4. Substitute A(k) into equation (7), which in general will not be zero. The gradient Vu (k)

computed will be the constrained derivatives ofF with respect to the independent variables
u(k).
5. Return.
We enter with given values for the independent variables u and exit with the (constrained)
derivatives of our objective function with respect to them. We have just described the routine
we indicated was needed for the unconstrained problem above where we use a succession of
quadratic fits to move toward the optimal point for an unconstrained problem. Apply that
method.
This approach is a form of the generalized reduced gradient (GRG) approach to
optimizing, one of the better ways to carry out optimization numerically.

Inequality Constrained Problems: To solve inequality constrained problems, we have to

develop a strategy that can decide which of the inequality constraints should be treated as
equalities. Once we have decided, then a GRG type of approach can be used to solve the
resulting equality constrained problem. Solving can be split into two phases: phase 1 where
the goal is to find a point that is feasible with respect to the inequality constraints and phase 2
where one seeks the optimum while maintaining feasibility. Phase 1 is often accomplished by
ignoring the objective function and using instead

P {if(z) if gi(Z) > D)}

F=L
i=1 0 otherwise

until all the inequality constraints are satisfied.

Once satisfied, we then proceed as follows. At each point check which of the inequality
constraints are active, i.e., exactly equal to zero. These can be placed into the active set and
 435

treated as equalities. The remaining can be put aside to be used only for testing. A step can
then be proposed using the GRG algorithm. If it does not cause one to violate any of the
inactive inequality constraints, the step is taken. Otherwise one can add the closest inactive
inequality constraint to the active set. Finding the closet inactive equality will almost certainly
require a line search in the direction proposed by the GRG algorithm.
When one comes to a stationary point, one has to test the active inequality constraints at
that point to see if they should remain active. This test is done by examining the sign (they
should be nonnegative if they are to remain active) of their respective Kuhn-Tucker multipliers.
If any should be released, it has to be done carefully as the release of a constraint changes the
multipliers for all the constraints. One can find oneself cycling through the testing to decide
whether to release the constraints. A correct approach is to add slack variables s to the
problem to convert the inequality constraints to equalities and then require the slack variables

to remain positive. The multipliers associated with the inequalities s ~ 0 all behave
independently, and their sign tells one directly to keep or release the constraints. In other
words, simultaneously release all the slack variables which have multipliers strictly less than
zero. If released, the slack variables must be treated as a part of the set of independent
variables until one is well away from the associated constraints for this approach to work.

Successive Quadratic Programming (SQP)

The above approach to finding the optimum is called a feasible path method as it attempts at all
times to remain feasible with respect to the equality and inequality constraints as it moves to
the optimum. A quite different method exists called the Successive Quadratic Programming
(SQP) method which only requires one be feasible at the final solution. Tests which compare
the GRG and SQP methods generally favor the SQP method so it has the reputation of being
one of the best methods known for nonlinear optimization for the type of problems we are
considering in this paper.
Assume we can guess which of the inequality constraints will be active at the final solution.
The necessary conditions for optimality are

V z L(z,~,)..) = VF + VgNl + Vh).. = 0

gA(Z) = 0
h(z) = 0
 436

Then, one can apply Newton's method to the necessary conditions for optimality, which are
a set of simultaneous (non)linear equations. The Newton equations one would write are

VzzUz(i), u(i), A(i») V!tA(z(i») Vh(z(i») Az(i) 1 [ VzUz(i), Il~i), A(i») j

[ V!tA(z(i»)T o o 1[AJ.L(i) =- gNz(I))
Vh(z(i»)T o o AA(i) h(z(i»)

A sufficient condition for a unique Newton direction is that the matrix of constraint
derivatives is of full rank (linear independence of constraints) and the Hessian matrix of the
Lagrange function (V zzL(z,Il,A.) projected into the space of the linearized constraints is
positive definite. The linearized system actually represents the solution of the following
quadratic programming problem:

Min VF(z(i»)TAz + 12 Alv zzL(z(i), Il(i), A.(i») A

!!.z

subject to
!tA(z(i») + V!tA(z(i»)TAz = 0
h(z(i») + Vh(z(i»TAz = 0

Reformulating the necessary conditions as a linear quadratic program has an interesting

side effect. We can simply add linearizations of the inactive inequalities to the problem and let
the active set be selected by the algorithm used to solve the linear quadratic program.
Problems with calculating second derivatives as well as maintaining positive definiteness of
the Hessian matrix can be avoided by approximating this matrix by B(i) using a quasi-Newton
formula such as BFGS [5,9,11,12,13,19,34]. One maintains positive definiteness by skipping
the update if it causes the matrix to lose this property. Here gradients of the Lagrange function
are used to calculate the update formula [22,30]. The resulting quadratic program, which
generates the search direction at each iteration i becomes:

subject to
g(z(i») + Vg(~i»)TAz S; 0
h(z(i») + Vh(z(i»)TAz = 0

This linear quadratic program will have a unique solution if B(i) is kept positive definite.
Efficient solution methods exist for solving it [16,20,25,38].
 437

Finally, to ensure convergence of this algorithm from poor starting points, a step size ex is
chosen along the search direction so that the point at the next iteration (zi+l=z4a.d) is closer
to the solution of the NLP [7,22,32].
These problems get very large as the Lagrange function involves all the variables in the
problem. If one has a problem with 5000 variables z and the problem has only 10 degrees of
freedom (i.e., the partitioning will select 4990 variables x and only 10 variables u), one is still
faced with maintaining a matrixB which is 5000x5000. Bema and Westerberg [3] proposed a
method that kept the quasi-Newton updates for B rather than keeping B itself They were able
to reduce the computational and space requirements significantly, exactly reproducing the steps
taken by the original algorithm. Later Locke et al [26] proposed a method which permitted B
to be approximated only in the space of the degrees of freedom, very significantly reducing the
space and computational requirements. More recently a "range and null space" decomposition
approach [29,36,37] has been proposed solving the problem. This decomposition loses
considerably on sparsity but is numerically more reliable. Lucia and Kumar [27] proposed and
tested explicitly creating the second derivative information for the Hessian and then exploited
its sparsity. In an attempt to keep the sparsity of the Locke et al approach and improve its
numerical reliability, Schmid and Biegler [33] have recently proposed methods to estimate the
terms which are missing in the Locke et al algorithm. Finally Schmid and Biegler are
developing a linear quadratic programming algorithm based on ideas in Goldfarb and Idnani
[20] which is much faster than available library codes.

Linear Programming

If the objective function and all equality and inequality constraints are linear, then a very
efficient means is available to solve our optimization problem. Considering only inequalities,
we can write

11in {cTu I Au ~ b, u ~ O} (primal LP)

We label this problem a prima/linear program as we shall later examine a corresponding dual
linear programming when we look at an example problem. Fig. 8 illustrates the appearance of
a small linear program.
To solve we change all inequalities into equalities by introducing slack variables

Min {cTu Ib + S - Au = 0, u, s ~ O}
u,s (15)
 438

and then observe, as Dantzig [8] did, that our solution can reside at a comer point for the
feasible region, such as points a, b, c and d in Fig. 8. If the objective exactly parallels one of
the boundaries, then the whole boundary - including its corner points - are solutions. If the
objective is everywhere equal, then all points are solutions, including again any of the corner
points. It is for this reason that we stated the solution can always reside at a corner point.

;"
~ ;"
~:ing;" ;"
objective
;"

;"

feasible region

Fig. 8 A Small Linear Program

How might we find these corner points? They must occur at intersection points where r of
the constraints are exactly zero. Intersection points are points a through d again plus points
like e which are outside the feasible region. Point a corresponds to Ul and U2 being
simultaneously zero. Point b corresponds to Ul and SI being zero, while point c corresponds to
Sl and S2 being zero. Ifwe examine the degrees of freedom for our problem, we see there are r

variables u, p variables sand p equality constraints. Thus, there are r degrees of freedom, the
number of variables u that exist in the problem. If we set r variables from the set u and s to
zero, solving the equations will provide us with an intersection point. If the remaining
variables are nonnegative, then the intersection point is feasible. Otherwise, it is outside the
feasible region.

The Simplex Algorithm: Dantzig [8] developed the remarkably effective Simplex algorithm
that allows one to move from one feasible intersection point to another, always in a downhill
 439

direction. Such a set of moves ultimately leads to the lowest corner point for the feasible
region which is then the solution to our problem. Each step in the Simplex algorithm is
equivalent to an elimination step in a Gauss elimination for solving linear equations and an
examination step of the result to discover which adjacent intersection points are feasible and
downhill.

Let us mentally apply the Simplex algorithm to the example in Fig. 8. Suppose we are
currently at point d. We examine how the objective function changes if we move along either
of the constraints which are active at d. To see what we really are doing, let each of the
variables which are zero at the corner point be our independent variables for the problem at
this point in time; here we identifY variables U2 and S2 as our independent variables. Increase
one while holding the remaining one(s) at zero. We find we move exactly along an edge
whose constraint(s) correspond to the variable(s) being held at zero. In particular release U2

and move along g2 = O. Release S2 and move along U2 = O.

The constrained derivatives for the degrees of freedom for the problem defined at the
current corner point tell us precisely how the objective function will change if we increase one
of the independent variables while holding the rest at zero, precisely what we are doing. So we
will need constrained derivatives, which we shall see are readily available if we do things
correctly.

Next, we need to know how far we can go. As we proceed, we generally encounter other
constraints. Suppose we have selected U2 to increase, moving along constraint g2. We will

encounter the constraint where S1 becomes zero. In effect, we trade U2 = 0 for S1 = 0 to arrive
at the adjacent point. The first variable to become zero as we increase U2 tells us where to stop.

We are now at point c. We start again with SI and S2 being our independent variables. We
must compute the constrained derivatives for them, select one, and move again. This time we
move to point b. At point b, the constrained derivatives are all positive indicating there is no
downhill direction to move. We have located the optimum point.

Problems occur when the region is unbounded if the objective function decreases in that
direction. The examination to discover which constraint is closest in the direction selected
finds there is no constraint in the way. The algorithm simply stops, reports an unbounded
solution exists and indicates the direction in which it occurs.
 440

Also, there can be degeneracy which occurs when more than r variables are zero at a
comer point. In Fig. 8, degeneracy would be equivalent to three constraints intersecting at one
point. Obviously only two are"needed to define the point. This redundancy can cause the
Simplex algorithm to cycle if care is not taken. If one encounters degeneracy, the normal
solution is to perturb enough of the constraints to have no more than r intersecting at the point.
One then solves the perturbed problem to move from the point, if movement is required,
removing the perturbations once away from the point.

Example: We shall carry out the solution for the following very small linear program.

Min F=2uI +3U2-U3

subject to
g,: u, +U2;S; 10
g2: u3;S; 4
hI: 2U2-SU3=6
u" U2, U3 ~ 0
We shall first put this problem into the form indicated by eqn (IS) to identify A, b and c
properly. The inequality constraints have to be written in the form Au = b + s.
-UI-U2 =-10+s,
- U3 = -4 + S2
We shall write the equality constraint with a special slack variable called an artificial
variable, using the same sign convention as for the inequality constraints.
-2 ul + SU3 = -6 + al
The artificial variable al must be zero at the final solution which we can accomplish by giving it
a very large cost as follows.

F(a) = cTu + (big number) a, = [2,3,-1] [~~] + 1000 al

The solution to the optimization problem will then make it zero, its least value. If it cannot be
removed from the problem in this manner, the original problem is not feasible. The constraints
can be put into matrix form, getting

Au-s=b => [-~ -~ _~][~~] -[:~]=[ -_~]

o -2 S U3 a, -6
uI, u2, u3, sl, s2, al ~ 0
We choose to transform each of the constraints so each of their RHS terms is positive by
multiplying each by -1 to form the following Simplex tableau.
 441

~
RHS
~
basic U2 U3 51 52
gl: 51 1 1 10
g2: 52 I I 4
hi: al I 2 -5 1 I 6
F(a):
F: I
2
2
3
3
-I
-I
1000 I 0
0
We have included six variables for this problem and three constraints. The first constraint,
for example, can be read directly from the tableau; namely, UI+U2+S1 = 10. We have set up
two objective function rows. The former has a large cost for the artificial variable al while the
latter has its cost set to zero. We shall use the row called F(a) as the objective function until
all of the artificial variables are removed (i.e., get set to zero) from the problem. Then we shall
switch the objective function row to F for the remainder of the problem. Once we have
switched, we will no longer allow an artificial variable to be reintroduced back into the
problem with nonzero value.
Each variable under the column labeled "basic" is the one which is being solved for using
the equation for which it is listed. They are chosen initially to be the slack and artificial
variables. All remaining variables are called "nonbasic" variables; they are chosen initially to be
all the problem variables (here Ul, U2 and U3) and will be treated as the current set of
independent variables. As we noted above, we set all independent (nonbasic) variables to zero.
The dependent (basic) variables of SI, S2 and al have a current value equal to their
corresponding RHS value, namely, 10, 6 and 4, respectively. Note that the identity matrix
appears under the columns for the basic variables. We put a zero into the RHS position for
both of the objective function rows F(a) and F.
If we reduce the row F(a) to all zeros below the dependent (basic) variables, then, as we
discussed earlier in our section on constrained derivatives and in Fig. 1, the entries below the
independent variables will be the constrained derivatives for the independent variables. To
place a zero here requires that we multiply row hI by 1000 and subtract it from row F(a),
getting the following tableau.

~ ~
basic U2 U3 51 52 RHS
gl: 51 1 1 10
~: 52 I 1 I 4
hi: al I 2 -5 I 6

I
F(a): 2
I
-1997 -I 0 -6000
F: 2 3 -1 0 0
 442

The value appearing in the RHS column for F( a) and F are the negative of the respective
objective function values for the current solution - namely for SI = 10, S2 = 4, al = 6, and Ul =
U2 = U3 = O. The constrained derivatives for Ul, U2 and U3 are 2, -1997 and -1 respectively.

We see that increasing U2 by one unit will decrease the objective function by 1997. U2 has the

most negative constrained derivative so we choose to "introduce" it into the "basis" - i.e., to
make it a dependent (basic) variable. We now need to select which variable to remove from
the basis - i.e., return to zero. We examine each row in tum.
Row gt: We intend to increase U2 while making the current basis variable, S1, go from 10 to O.

U2 will increase to 10/1 = 10. The "1" used here is the coefficient under the column U2

in row gl.

Row g2: A zero in this row under U2 tells us that U2 does not appear in this equation. It is,
therefore, impossible to reduce the basis variable for that row to zero and have U2
increase to compensate.
Row ht: Here, making the basis variable al go to zero will cause U2 to increase to 6/2 = 3.

If any of the rows had made the trade by requiring U2 to take a negative value, we skip this

row. It is saying U2 can be introduced to an infinite positive amount without causing the

constraint it represents to be violated.

We can introduce U2 at most to the value of3, the lesser of the two numbers 10 and 3. If
we go past 3, al will go past zero and become negative in the trade. We introduce U2 into the
basis and remove al. To put our tableau into standard form, we want the column under
variable U2 to have a one in row hI and zeros in all the other rows. We accomplish this by
performing an elimination step corresponding to a Gaussian elimination. We first rescale row
hI so a 1 appears in it under U2 by dividing it by 2 throughout. We subtract (1) times this row

from row gl to put a zero in that row, subtract (-1997) times this row from row F(a) and

finally (3) times this row from row F, getting the following tableau.

basic UI U2 U3 SI S2 al I RHS

gl: SI 0 -2.5 ~.5 I 7

g2: S2 I 4
hI: U2 -2.5 0.5 I 3
F(a): 2 0 6.5 998.51 -9
F: 2 0 6.5 -1.5 I -9
 443

Note that, we have indicated that U2 is now the basic variable for row hi. All the artificial

variables are now removed from the problem. We switch our attention from objective function
row F(a) to row F from this point on in the algorithm. Constrained derivatives appear under
the columns for the independent variables UI, U3 and al. We ignore the constrained derivative

for the artificial variable al as it cannot be reintroduced into the problem. It must have a zero

value at the final solution Constrained derivatives for uland U2 in row F are positive

indicating that introducing any of these will increase the objective function. We are at a
minimum point. The solution is read straight from the tableau: Sl = 7, S2 = 4, U2 = 3, Ul = U3
= al = O. The objective function is the negative of the RHS for row F, i.e., F = 9. Since the
artificial variable a 1 is zero, the equality constraint is satisfied, and we are really at the solution.

In the example given to illustrate the generalized dual, we in fact developed the dual to a
linear program (as may have been evident to the reader at the time). Eqn. (12) gives the dual
formulation for our problem, namely

Max {bTAJ ATA S; c, 1.<': O}

A.
which for our example problem is

Max -10 Al - 4 1.2 - 61.3

subject to
-AI S; 2
-AI - 2 1.3 S; 3
-1.2 + 51.3 s;-1
1.1,1.2<':0
The third Lagrange multiplier, 1.3, does not have to be positive as it corresponds to the equality

constraint. To convert this problem to a linear program where all variables must be positive,
we split 1.3 into two parts, a positive part and a negative part. We can also add in slack

variables al to a3 at the same time and write the following equivalent optimization problem.

Max -10 Al - 4 1.2 + 6 (A; - 1.3)

subject to
-AI + al = 2
-AI +2 (A; - 1.3) + a2 = 3
-1.2 - 5 (A; - 1.3) + a3 = -1
AI, A2, A;, A3, aI, a2, a3<': 0

One can show that the values for the constrained derivatives in the final primal tableau

provide us with the solution to the corresponding dual problem, namely: al = 2, a2 = 0, a3 =

 444

6.5, Al = 0, A2 = 0 and A3 = -1.5 (i.e., A; = 0, A3 = 1.5). The numbers are in the objective
function row F in the order given here. The reader should verify that this is the solution.
The correspondence is seen as follows. The first column of the original tableau gives us
the first equation in the dual; its slack 0"1 corresponds to this column. A similar observation

holds for columns 2 and 3. The multipliers Al to A3 are the Lagrange multipliers for the
equations in the primal problem. Their values are the constrained derivatives for the slack
variables for the constraints for primal problem. So, for example, Al appears under the column

for Sl.

Interior point algorithms for Linear Programming Problems

There has been considerable excitement in the popular press about so-called interior point
algorithms [23] for solving extremely large linear programming problems. Computational
demands for these algorithms grow less rapidly than for the Simplex algorithm, with a break-
even point being a few thousand constraints. We shall base the presentation in this section on a
recent article by MacDonald and Hrymak [28] which itself is based on ideas in references
[1,21].
A key idea for an interior method is that one heads across the feasible region to locate the
solution rather than around its edges as one does for the Simplex algorithm. This move is
found by computing the direction of steepest descent for the objective function with respect to
changing the slack variables. Variables u are computed in terms of the slack variables by using
the inequality constraints. The direction of steepest descent is a function of the scaling of the
variables used for the problem.
A really clever and second key idea for interior point algorithms is the way the problem is
scaled. At the start of each iteration one rescales the space one searches so that current point
is a unit distance from the constraint boundaries. In general one must rescale all the variables.
The algorithm has to guarantee that it will never quite reach the constraint boundaries at the
end of an iteration so that one can then rescale the variables at the start of the next to be one
unit from the boundary. A zero distance cannot be rescaled to be a unit distance.
One terminates when changes in the unscaled variables become negligible from one
iteration to the next. The distances to the constraints which finally define the optimal solution
have all been magnified significantly. One will really be close to them.
 445

Consider the following linear program.

Min
u.s
{F = cTu Ib + S - Au = 0, u, s ~ O}

where slack variables have been introduced into the problem to convert inequalities into
equalities. Assume matrix A has p rows and r columns and that there are more constraints than
there are variables u, i.e., p ~ r.
Assume we are at some point in the interior of the feasible region, (u(k), s(k)). Let Ds and
Du be diagonal matrices whose i-th diagonal elements are sj(k) and uj(k) respectively. Define
rescaled variable vectors as
s' = nsl s and u' = nul u
and replace the variables in the constraints by these rescaled variables, getting
nslb + s' -nslAD,.u' =0
We wish to solve these equations for u' in tenns of s'; however, the coefficient matrix DsADu is
of dimension mxr. We must first make it square which we do by premultiplying by its
transpose (rxp x pxr --> rxr where p ~ r has been assumed above) to get
-DuATnsl(Dslb + s') + (DuATns2ADu)u' = 0
These equations can now be solved, giving
u' = (DuATns2ADurl DuATns1(DSlb + s')
We can now determine the gradient of the objective function, F = cTu, with respect to the
rescaled slack variables s', getting

Vs·F = (Vs·u')Tc = Dsl ADu(DuATns2ADJ- l c

The direction of steepest descent for changing the rescaled slack variables is the negative of
this gradient direction. We let the step in s' be in this direction. The corresponding steps for u'
and the unsealed variables s and u follow directly.
!ls' = _Dsl ADu(DuATns2ADurl c
!lu' =- (DuATDs2ADJ-l c
ns
&= -ADu (DuAT 2ADJ- 1 c
!lu =- Du(DuATns2ADJ- l C
The above defines the direction to move. We want to move close to but not exactly onto
the edge of the feasible region. We encounter the edge when one of the variables in s or u
becomes zero while the rest stay positive while taking our step. The variable with the most
 446

negative value for sj/,1sj or uj/,1uj, as appropriate, is the one that will hit the edge first. One

typically takes a step that goes more than 99% but not 100% of the way toward the edge.
The last issue to settle is how to get an initial point which is strictly inside the feasible
region. By introducing slack variables and artificial variables as we did above for our linear
programming example, we can pick a point which is feasible but on the constraint boundary.
Pick such a point but then make all the variables which are normally set to zero just slightly
positive.
All the work in this algorithm is in factoring the matrix (DuATns2ADu) which is far less
sparse that the original coefficient matrix A. Because the algorithm is useful only when the
problems get really large, one must use superb numerical methods to do these computations.
MacDonald and Hrymak show that the Karmarker algorithm (closely related to the above
algorithm) is a special case of the Newton Barrier Method [18]. The constraints stating the
nonnegativity of the problem variables u and the slack variables s are replaced by adding a term
to the objective function that grows to infinity as anyone of these variables decreases to zero:
r p
F = cTu - 11 (~ In Uj + ~ In Sj )
i=1 j=1

where 11 is a positive fixed scalar. One needs a feasible starting point strictly interior to the
feasible region.
MacDonal~ and Hrymak also discuss a method that attacks the problem from both the
primal and dual formulations simultaneously. Both forms are written with slack variables
where the nonnegativity for all the variables is maintained by erecting barrier functions. The
necessary conditions for optimality are stated separately for both the primal and dual with their
respective barrier functions in place. The necessary conditions are then combined. Solving
them computes a direction in which to move.

Benders' Decomposition

One way to solve the following problem

Min {F(u,y) Ig(u,y):5: 0, h(u,y) = O}

u,y

IS to use an approach called projection. Projection breaks the problem into an outer
optimization which contains an inner one, such as the following for the above problem
 447

Min { Min {F(u,y) 'g(u,y):;; 0, h(u,y) = O}}

y u

There is no advantage to using projection unless the inner problem possesses a special
structure which is very easy to solve. For example the inner problem might be a linear
program while the overall problem is a nonlinear problem. Benders [2] presented a method for
precisely this case which Geoffiion subsequently generalized [15]. We shall look at the
problem Benders presented here, using ideas presented earlier in this paper. Benders' problem
was the following

Min {cTu + F(y)

u,y
,Au +f(y) ~ b, u~ 0, YES}
which, when projected, yields the following two-level optimization problem

Min {F(y)+ Min {cTu IAu ~ b - f(y), u ~ 0 }IYES}

y u

With y fixed the inner problem is a linear program. We can replace the inner problem with
its equivalent dual representation

Min {F(y) + Max {(b - f(yWA,' ATA,:;; c, A, ~ 0 }, YES}

y A

Solving either form gives us a solution for the original problem; however, there are some
advantages to this latter form. The solution to the dual problem occurs at one of the comer

points for the feasible region. The feasible region is defined by the constraints ATA, :;; c, A, ~ 0,
and these constraints do not contain the outer problem variables y so the comer points are not
a function ofy.

If we are given a value for y and if we had a list of all the comer points A,(l ),A,(2) ... , we
could solve the inner problem by picking that comer point having the maximum value for its

corresponding objective function (b - f(y»TA,(k). Ifwe know only some of the comer points,
we can search over just those for the best one, hoping the optimal solution for the given value
of y will reside there. Not having all the points, the value found will be less than or at most
equal to the maximum; it will provide an lower bound on the maximum. We can subsequently
solve the inner problem given y. If we obtain the same comer point, then the inner problem
was exactly solved using only the known subset of the corner points. If we do not, then we
can add the new corner point to the list of those found so far.

The feasible region may be unbounded for the dual problem which occurs when the primal
inner problem is infeasible for the chosen value of y. The directions in which the inner dual
 448

problem are unbounded are not a function of y, as we established above. To understand the
geometry of an unbounded problem, consider Fig. 9.

(b)
Fig. 9 Unbounded feasible region for the dual. (a) shows the constraints for the original dual problem.
(b) shows the constraints shifted to pass through zero.

We shift all the constraints to pass through the origin. Then any convex combination of the
constraints bounding the feasible region in (b) represents a direction in which the problem in
(a) is unbounded. Any v which satisfies the following set of constraints is an unbounded
direction:
{ATV SO, V ~ 0, L vi= 1 }
We can find it by reposing the dual problem as follows: (1) zero the coefficients of the
objective function, (2) zero the right-hand-side of the inequality constraints and (3) add in the
constraint that the multipliers add to unity.
Such an unbounded direction imposes a constraint on y, and these constraints are called
cutting planes. To preclude an infinite solution for the dual - and therefore an infeasible
solution for the primal inner problem and thus the problem as a whole, we can state that the
inner dual objective function must not increase in this unbounded direction, namely

(b - f(yW v(k) S 0

We add this constraint to our problem statement. It is a constraint on the value of y, the
variables for the outer optimization problem. The inner problem has fed back a constraint to
the outer problem on the values of y it should consider when optimizing over them.
The algorithm is as follows.
1. Set the iteration counter k=O. Define empty sets C (for corner points) and U (for
unbounded directions).
 449

2. Solve the following optimization problem for y (if set U is empty, set 0 equal to zero)

Min {F(y) +0 I 0 ~ [b-f(y)]TA,(i) Ifor all A,(i)eC}, yeS, [b-f(yWv(j)~ ofor all vG)eU}
y
This problem tells us to find the value of y that minimizes the sum of F(y) and the
maximum of the inner problem objective over all the comer points found so far, subject
to y being in the set S and satisfying all the cutting plane constraints found so far. The
sets C and U are initially empty so the first time through, one initially finds a value of y
that minimizes F(y) subject to y being in the set S. Exit if there is no value of y
satisfying all the constraints, indicating that the original problem is infeasible.
3. For this value ofy, solve the inner problem

Max {(b - f(Y)lA,1 ATA, ~ c, A, ~ O}

A.
which will give rise to a new comer point A,(k) or, ifthe problem is unbounded, a
direction v(k). Place whichever is found in the set C or U as appropriate.
4. Ifthe solution in step 3 is bounded and has the same value for the inner objective as
found in step 2, exit with the solution to the original problem. Else increment the
iteration counter and return to step 2.
The problem in step 2 grows with each iteration. Thus, these problems can get large.
Geoffiion [15] has generalized Benders' method. The generalized method is used to solve
mixed integer programming problems, for example, where the inner problem variables are
required to take on values of zero and one only. Such use is indicated in the papers by
Grossmann at this Advanced Study Institute.

References

1. Alder, II, N. Kannarker, M. Resende and G. Veigo, "Implementation ofKannarkar's Algorithm for Linear
Programming," Mathematical Programming, 44, 297-335 (1989)
2. Benders, J.F., "Partitioning Procedures for Solving Mixed-variables Progranuning Problems," Numerische
Mathematik, 4, 238 (1962)
3. Berna, T.J., Locke, M.H., and A.W. Westerberg, "A New Approach to Optimization of Chemical
Processes," AlChE J, ~ 37-43 (1980)
4. Biegler, L.T., I.E. Grossmann, and A.W. Westerberg, "Optimization," in Ullmann's Encyclopedia of
Industrial Chemistry, Bl, Mathematics in Chemical Engineering, Chapt. 10, 1-106 to 1-128,
Weinheim:VCH Verlagsgesellschaft 1990
5. Broyden, C.G., "The Convergence of a Class of Double Rank Minimization Algorithms," 1. lnst. Math.
Applic., 6, 76 (1970)
6. Bryson, A.E. and Y-C Ho, Applied Optimal Control, Washington D. C.:Hemisphere Publishing 1975
7. Chamberlain, R.M., C. Lemarechal, H.C. Pedersen, and MJ.D. Powell, "The Watchdog Technique for
Forcing Convergence in Algorithms for Constrained Optimization," Math. Prog. Study 16,
Amsterdam:North Holland 1982
 450

8. Dantzig, G., "Linear Programming and Extensions", Princeton:Princeton University Press 1963
9. Davidon, W.C., "Variable Metric Methods for Minimization," AEC R&D Report ANL-5990 (rev.) (1959)
10. Dennis, J.E. and J.J. More,"Quasi-Newton Methods, Motivation and Theory," SIAM Review, 21, 443
(1977)
II. Fletcher, R and Powell, MJ.D., "A Rapidly Converging Descent Method for Minimization," Computer 1.,
6, 163 (1963)
12. Fletcher, R, "A New Approach to Variable Metric Algorithms," Computer 1., 13, 317 (1970)
13. Fletcher, R, Practical Methods o/Optimization, New York:WiJey 1987
14 Fourer, R, D.M. Gay and B.W. Kerninghan, "A Modeling Lanugage for Mathematical Programming,"
Management Science, 36(5), 519-554 (1990).
15. Geoffrion, AM., "Generalized Benders Decomposition", JOTA, 10,237 (1972)
16. Gill, P. and W. Murray, "Numerically Stable Methods for Quadratic Programming," Math. Prog., 14, 349
(1978)
17. Gill, P.E., W. Murray and M. Wright, Practical Optimization, New York:Academic Press 1981
18. Gill, P.E., W. Murray, M. Saunder, 1. Tomlin and M. Wright, "On Projective Newton Barrier Methods for
Linear Programming and as Equivalence to Caretakers Projective Method," Math. Prog., 36, 183, (1984)
19. Goldfarb, D., "A Family of Variable Metric Methods Derived by Variational Means," Math. Comp., 24, 23
(1970)
20. Goldfarb, D. and A. Idnani, "A Numerically Stable Dual Method for Solving Strictly Convex Quadratic
Programs," Math. Prog., 27, I (1983)
21. Goldfarb, D. and MJ. Todd, "Linear Programming", Chapter II in Optimization (eds. G.L. Nemhauser,
AH.G. Rinnoy Kan and MJ. Todd), Arnsterdam:North Holland 1989
22. Han, S-P, "A Globally Convergent Method for Nonlinear Programming," 1. Opt. Theo. Applics., 22, 297
(1977)
23. Karmarker, N., "A New Polynomial-Time Algorithm for Linear Programming," Combinatorica, 4, 373-
395 (1984)
24. Kuhn, H.W., and Tucker, AW., "Nonlinear Programming," in Neyman, 1. (ed), Proc. Second Berkeley
Symp. Mathematical Statistics and Probability, 402-411, Berkeley, CA:Univ. California Press 1951
25. Lemke, C. E., "A Method of Solution for Quadratic Programs," Management Science, 8, 442 (1962)
26. Locke, M.H., A.W. Westerberg and RH. Edahl, "An Improved Successive Quadratic Programming
Optimization Algorithm for Engineering Design Problems," AIChE J, Vol. 29, pp 871-874 (1983)
27. Lucia, A and A. Kumar, "Distillation Optimization," Compo Chern. Engr., 12, 12, 1263 (1988)
28. MacDonald, W.E., AN. Hrymak and S. Treiber, "Interior Point Algorithms for Refinery Scheduling
Problems," in Proc. 4th Annual Symp. Process Systems Engineering, Montibello, Quebec, Canada, pp
III. 13. 1-16, Aug. 5-9, 1991
29. Nocedal, Jorge and Michael L. Overton, "Projected Hessian Updating Algorithms for Nonlinearly
Constrained Optimization", SIAM J. Numer. Anal., 22, 5 (1985)
30. Powell, MJ.D., "A Fast Algorithm for Nonlinearly Constrained Optimization Calculations," Lecture Notes
in Mathematics 630, Berlin:Springer Verlag 1977
31. Reklaitis, G.V., A Ravindran, and K.M. Ragsdell, Engineering Optimization Methods and Applications,
New York: Wiley 1983
32. Schittkowski, K., "The Nonlinear Programming Algorithm of Wilson, Han and Powell with an Augmented
Lagrangian Type Line Search Function," Num. Math., 38, 83 (1982)
33. Schmid, C. and L.T. Biegler, "Acceleration of Reduced Hessian Methods for Large-scale Nonlinear
Programming," paper presented at AIChE Annual Meeting, Los Angeles, CA, Nov. 1991
34. Shanno, D.F., "Conditioning of Quasi-Newton methods for function minimization," Math. Comp., Co 24,
647 (1970)
35. Umeda, T. and A Ichikawa, "I&EC Proc. Design Develop., ill 229 (1971)
36. Vasantharajan, S. and L.T. Biegler, "Large-Scale Decomposition Strategies for Successive Quadratic
Programming," Camputers and Chemical Engineering, 12, 11, 1087 (1988)
37. Vasantharajan, S., 1. Viswanathan and L.T. Biegler, "Large Scale Development of Reduced Successive
Quadratic Programming," presented at CORSrrIMS/ORSA Meeting, Vancouver, BC, May, 1989
38. Wolfe, P., "The Simplex Method for Quadratic Programming," Econometrica, 27, 3, 382 (1959)
Mixed-Integer Optimization Techniques for the
Design and Scheduling of Batch Processes

Ignacio E. Grossmann, Ignacio Quesada, Ramesh Raman and Vasilios T. Voudouris

Department of Chemical Engineering and Engineering Design Research Center, Carnegie Mellon
University, Pittsburgh, PA 15213, U.S.A.

Abstract: This paper provides a general overview of mixed-integer optimization

techniques that are relevant for the design and scheduling of batch processes. A brief
review of the recent application of these techniques in batch processing is fIrst presented.
The paper then concentrates on general purpose methods for mixed-integer linear (MILP)
and mixed-integer nonlinear programming (MINLP) problems. Basic solution methods as
well as recent developments are presented. A discussion on modelling and reformulation is
also given to highlight the importance of this aspect in mixed-integer programming.
Finally, several examples are presented in various areas of application to illustrate the
performance of various methods.

Keywords: mathematical programming, mixed-integer linear programming, mixed-

integer nonlinear programming, branch and bound, nonconvex optimization, reformulation
techniques, batch design and scheduling

Introduction

The design, planning and scheduling of batch processes is a very fertile area for the
application of mixed-integer programming techniques. The reason for this is that most of
the mathematical optimization models that arise in these problems involve both discrete and
continuous variables that must satisfy a set of equality and inequality constraints, and that
must be chosen so as to optimize a given objective function. While there has been the
recognition that many batch processing problems can be posed as mixed-integer
optimization problems, the more extensive application of these techniques has only taken
place in the recent past.
 452

It is the purpose of this paper to provide an overview of mixed-integer optimization

techniques. We will fIrst present a brief review of the application of these techniques in
batch processing. We then provide a brief introduction to mixed-integer programming in
order to determine a general classifIcation of major problem types. Next we concentrate in
both mixed-integer linear (MILP) and mixed-integer nonlinear programming (MINLP)
techniques, introducing fIrst the basic methods and then the recent developments that have
taken place. We then present a discussion on modelling and refonnulation, and fInally,
some numerical examples and results in various areas of application.

Review of applications

In this section we will present a brief overview of the application of mixed-integer

programming in batch processing. More extensive reviews can be found in [55] and
[66,67].

Mixed-integer nonlinear programming techniques have been applied mostly to

design problems. Based on the problem considered by Sparrow et al [81], Grossmann and
Sargent [28] were the fIrst to formally model the design of multiproduct batch plants with
parallel units and with single product campaigns as an MINLP problem. These authors
showed that if one relaxes the numbers of parallel units to be continuous, the associated
NLP corresponds to a geometric program that has a unique solution. Rather than solving
the problem directly as an MINLP, the authors proposed a heuristic rounding scheme for
the number of parallel units using nonlinear constraints based on the solution of the relaxed
NLP. Since this problem provides a valid lower bound to the cost, optimality was
established within the deviation of the rounded solution. This MINLP model was
subsequently extended by Knopf et al [36] in order to handle semi-continuous units. A
further extension, was the MINLP model for a special type of multipurpose plants by
Suharni and Mah [83] in which simultaneous production was only allowed if products did
not require the same processing stages. This model was subsequently modified by
Vaselenak et al [90] and by Faqir and Karimi [19] to embed the selection of production
campaigns. However, all these works did not rigorously solve the MINLP, but they relied
on the rounding scheme by Grossmann and Sargent [28] for obtaining an integer number
of parallel units.

The first design application in which an MINLP model was rigorously solved was
[90] who considered the retrofit design of multiproduct batch plants. These authors applied
 453

the outer-approximation method by Duran and Grossmann [18] with a modification to

handle separable non convex terms in the objective. Recently, [20] removed the
assumptions of equal volume for units operating out of phase by Vaselenak et al [90], and
formulated a new MINLP model that again was solved by the outer-approximation method.
Also, [38] formulated the MINLP model by Grossmann and Sargent [28] in terms of 0-1
variables for the parallel units and solved it rigorously with the outer-approximation method
as implemented in DICOPT. Subsequently, [96] applied this computer code to an MINLP
model for multiproduct plants under uncertainty with staged expansions.

An important limitation in all the above applications was that convexity of the
relaxed MINLP problem was a major requirement. Also, it became apparent that the
solution of larger design problems could become expensive. The first difficulty was
partially circumvented with the augmented penalty version of the outer-approximation
algorithm proposed by Viswanathan and Grossmann [92] and which was implemented in
the computer code DICOPT++. This code was applied by Birewar and Grossmann [9] for
the simultaneous synthesis, sizing and scheduling of multiproduct batch plants which gives
rise to a nonconvex MINLP model.

Papageorgaki and Reklaitis [53,54] developed a comprehensive MINLP model for

the design of multipurpose batch plants which involved nonconvex terms. They found that
the code DICOPT would get trapped into suboptimal solutions and that the computation
time was high. For this reason they proposed a special decomposition method in which the
subprqblems are NLP's with fixed 0-1 variables and campaign lengths and the master
problem corresponds to a simplified MILP. Faqir and Karimi [19] also modelled a special
class of multipurpose batch plants with multiple production routes and discrete sizes as an
MINLP problem that involves nonconvexities in the form of bilinear constraints. These
authors proposed valid underestimators for these constraints and reduced the design
problem to a sequence of MILP problems. Recently, [93] have shown that several batch
design problems, convex and nonconvex, can in fact be reformulated as MILP problems
when they involve discrete sizes. Examples include the design of multiproduct batch plants
with single product campaigns and the design of multipurpose batch plants with multiple
routes. These authors have also developed a comprehensive MILP synthesis model for
multiproduct plants in which the cost of inventories is accounted for [95]. Finally, [65]
have reported computational experience in solving a variety of batch design problems as
MINLP problems using the computer code DICOPT++, while [82] have applied it in the
optimization of flexibility of mUltiproduct batch plants.
 454

As for scheduling and planning, there have been a large number of MILP models
reported in the Operations Research literature. However, in chemical engineering the first
major MILP model for batch scheduling was proposed by [68] for the case of multipurpose
batch plants in which the products were preassigned to processing units. They used the
computer code LINDO [74] to solve this problem, and later extended it to handle the
production of a product over several predefined sets of units [69]. Ku and Karimi [42]
developed an Mll..P model for selecting production sequences that minimize the inakespan
in multiproduct batch plants with one unit per stage. Their model, which can accomodate a
variety of storage policies, was also solved with the computer code LINDO.

A very general approach to the scheduling of batch operations was proposed by

Kondili et al [40] in which they developed a state-task network representation to model
batch operations with complex process network structures. By discretizing the time
domain they posed their problem as a multiperiod MILP model that has the flexibility of
accomodating variable batch sizes, splitting and mixing of batches, finite, unlimited or no
storage, various transfer policies and resource constraints. Furthermore, the model has the
flexibility of assigning equipment to different tasks. Recently, [77] have been able to
considerably tighten the LP relaxation for this problem and develop a special purpose
branch and bound method with which these authors have been able to solve problems with
more than one thousand 0-1 variables. These authors have also extended their MILP model
to some design and planning problems [76].

For the case of the no-wait flowshop scheduling problem [49] (see also [57])
formulated the problem as an asymmetric traveling salesman problem (see [30]). For this
model they developed a parallel .branch and bound method that was coupled with a
matching algorithm for detecting Hamiltonian cycles. The specialized implementation of
their algorithm has allowed them to solve problems to optimality with more than 10,000
batches, which effectively translates to problems with more than 20,000 constraints and
100,000,000 0-1 variables.

Finally, MINLP models for scheduling of mUltipurpose batch plants have been
formulated by Wellons and Reklaitis [97] to handle flexible allocation of equipment and
campaign formations. Due to the large size of these problems, these authors developed a
special decomposition strategy for their solution. Sahinidis and Grossmann [70]
considered the cyclic scheduling of continuous multiproduct plants with parallel lines and
formulated the problem as a large-scale MINLP problem. They developed a solution
 455

method based on Generalized Benders decomposition for which they were able to solve
problems with up to 800 0-1 variables, 23,000 continuous variables and 3000 constraints.

In summary, what this brief review shows is that both MILP and MINLP
techniques are playing an increasingly important role in the modelling and solution of batch
processing problems. This review also shows the importance of exploiting the structure of
these problems for developing reasonably efficient solution methods. It should also be
mentioned that while there might be the temptation to resort to simpler optimization
approaches such as simulated annealing, mixed integer programming provides a rigorous
and deterministic framework, although it is not always the easiest one to apply. On the
other hand, many mixed-integer problems that were regarded as unsolvable 10 years ago
are currently being solved to optimality with reasonable computing requirements due to
advances in algorithms and increased computer power.

Mixed-integer Programming

In its most general form a mixed-integer program corresponds to the optimization problem,

min Z = f(x,y) (MIP)

S.t. h(x,y)=0
g(x,y)::;; 0
xe RD ye N+m

in which x is a vector of continuous variables and y is a vector of integer variables. The

above problem (MIP) specializes to the two following cases:

I. Mixed-integer linear programming (MILP). The objective function f, and the constraints
h and g are linear in x and y in this case. Furthermore, most of the applications of interest
are restricted to the case when the integer variables y are binary, i.e. ye {0,1 }m. A
number of important classes of problems include the pure integer linear programming
problem (only integer variables) and a large number of specialized combinatorial
optimization problems that include for instance the assignment, knapsack, matching,
covering, facility location, networks with fixed charges and traveling salesman problems
(see [51]).
 456

II. Mixed integer nonlinear programming (MINLP). The objective function and/or
constraints are nonlinear in this case. The most common form is linear in the integer
variables and nonlinear in the continuous variables [27]. More specialized forms include
polynomial 0-1 programs and 0-1 multilinear programs which can be transformed into
MlLP problems (eg see [6]). The difficulty that arises in the solution of MlLP and MINLP
problems is that due to the combinatorial nature of these problems, there are no optimality
conditions like in the continuous case that can be directly exploited for developing efficient
solution methods. In this paper we will concentrate on the modelling and solution of
unstructured MlLP problems, and MlNLP problems that are linear in the 0-1 variables.
Both types of problems correspond to the more general type of mixed-integer optimization
problems that arise in batch processing. It is very important however, to recognize that if
the model has a more specialized structure, general purpose techniques will be inefficient
for solving large scale version of these problems, and specialized combinational
optimization algorithms should be used in this case.

Mixed-integer Linear Programming (MILP)

We will assume the more common case in which the subset of the integer variables y are
restricted to take only 0 or 1 values. This then gives rise to the MlLP problem:

min Z = cTx + bTy (MlLP)

s.t. Ax + By ~ d
x ~ 0, YE {O,1}m

In attempting to develop a solution method to solve problem (MlLP), the first

obvious alternative would be to solve for every combination of 0-1 variables the
corresponding LP problem in terms of the variables x, and then pick as the solution the 0-1
combination with lowest objective function. The major drawback with such an approach is
that the number of 0-1 combinations is exponential. For example, an MILP problem with
10 0-1 variables would require the solution of 2 10 = 1024 LPs, while a problem with 50 0-
1 variables would require the solution at 250 =1.13xl015 LPs! Thus, this approach is, in
general, computationally infeasible.

A second alternative is to relax the 0-1 requirements and treat the variables y as
continuous with bounds, 0 ~ y ~ 1. The problem with such an approach, however, is that
 457

except for few special cases (e.g. assignment problem), there is no guarantee that the
variables y will take integer values at the relaxed LP solution. As an example, consider the
pure integer programming problem,

min Z = -1.2Yl - Y2
s.t. 1.2Yl + 0.5Y2::;; I (1)
Yl +Y2::;; 1
Yl , Y2 = 0,1

By relaxing Yl and Y2 to be continuous the solution yields the noninteger point

Y1=0.715, Y2=0.285, Z= -1.143. Assume we simply round the variables to the nearest
integer value, namely Yl = 1, Y2 =O. This, however, is an infeasible solution as it violates
the first constraint. In fact, the optimal solution is Yl =0, Y2 = 1, Z = -1. Thus, solving
the MlLP problem by relaxation of the Y variables and rounding them to the nearest integer,
will in general not lead to the correct solution. Note, however, that the relaxed LP has the
property that its optimal objective value provides a lower bound to the integer solution.

In order to obtain a rigorous solution to the problem (MlLP) the most common
approach is the branch and bound method which originally was proposed by Land and
Doig [44] and later formalized by Dakin [16]. In the branch and bound technique the
objective is to perform an enumeration without having to examine all the 0-1 combinations.
The basic idea is first to represent all the 0-1 combinations through a binary tree such as the
example shown in Fig. 1. Here at each node of the tree the solution of the linear program
subject to integer constraints for the subset of the Y variables that are fixed in previous
branches is considered. For example, in node A the root of the tree involves the solution of
the relaxed LP, while node B involves the solution of the LP with fixed Yl = 0, Y2 = 1 and
with 0 ::;; Y3 ::;; 1.

In order to avoid the enumeration of all the nodes in the binary tree, we can exploit
the following basic properties. Let k denote a descendent node of node 1 in the tree (e.g.
k=B, 1=A) and let (pk) and (pI) denote the corresponding LP subproblems. Then the
following properties can be easily established:
 458

1. If(pl) is infeasible then (pk) is also infeasible.

2. If (pk) is feasible then (pI) is also feasible, and (Zl )*:;; (Zk)*. That is, the optimal
objective of subproblem (pI) corresponds to a lower bound of the optimal objective at
subproblem (pk).

3. If the optimal solution of subproblem (pk) is such thal y = 0 or 1, then (Zk)* :::: Z*.
That is, the optimal objective of subproblem (pk) corresponds to an upper bound of
Z*, the optimal Mn..P solution.

The above properties can be used to fathom nodes in the tree within an enumeration
procedure. The question of how to actually enumerate the tree involves the use of
branching rules. Firstly, one does not necessarily have to follow the order of the index of
the variables y for branching as might be implied in Fig. 1. A simple alternative is to
branch instead on the 0-1 variable that is closest to 0.5. Alternatively, one can specify a
priority for the 0-1 variables, or else use a more sophisticated scheme that is based on the
use of penalties [17, 86]. Secondly, one has to decide as to what node should be examined
next having solved the LP at a given node in the tree. Here the two major alternatives are to
use a depth-first (last in-first out) or a breadth-first (best second rule) enumeration. In the
former case one of the branches of the most recent node is expanded first; if all of them
have been examined we backtrack to another node. In the

Fig. 1 Binary tree representation for three 0-1 variables

 459

latter case the two branches of the node with the lowest bound are expanded successively;
in this case no backtracking is required. While the depth-first enumeration requires less
storage, the breadth-first enumeration requires in general an examination of fewer nodes.
In practice the most common scheme is to use depth first, but by branching on both the 0
and 1 values of a binary variable at each node.

In summary, the branch and bound method consists in first solving the relaxed LP
problem. If y takes integer values we stop. Otherwise we proceed to enumerate the nodes
in the tree according to some specified branching rules. At each node the corresponding LP
subproblem is solved, typically by updating the dual LP problem of the previous node
which requires few pivot operations. By then making use of the properties cited before,
we either fathom the node (if infeasible or if lower bound ~ upper bound) or keep it open
for further examination. Clearly the computational efficiency is largely dependent on the
quality of the lower bounds of the LP subproblems.

As an example, consider the following MILP problem involving one continuous

variable and three 0-1 variables:

min Z = x + YI + 3Y2 + 2Y3

S.t. -x + 3YI + 2Y2 + Y3 S; 0 (2)
- 5Yl - 8Y2 - 3Y3 S; - 9
x ~ 0 , Yl, n, Y3 = 0,1

The branch and bound tree using a breadth-first enumeration is shown in Fig. 2.
The number in the circles represents the order in which 9 nodes out of the 15 nodes in the
tree are examined to find the optimum. Note that the relaxed solution (node 1) has a lower
bound of Z = 5.8, and that the optimum is found in node 9 where Z = 8, YI=O, n=Y3=1,
andx=3.
 460

Z=9
[1.1.0)

Infeas.

Fig. 2 Branch and bound tree for example problem (2)

The branch and bound method is currently the most common method used for
MILP in both academic and commercial computer software (eg. LINDO, ZOOM,
SCICONIC, OSL, CPLEX, XA). Some of these codes feature a number of special
features that can help to reduce the enumeration in the tree search. Perhaps one of the most
noteworthy are the generalized upper bound constraints [7] which are integer constraints of
the form,

(3)

In this case instead of performing branching on individual variables, the branching

is performed by partitioning the variables into two subsets (commonly of equal size). As a
simple example consider the problem:

min. Z = Yl + 2Y2 + 3Y3 + 4Y4

S.t. Yl + Y2 - Y3 -Y4 ~ 0 (4)
Yl + Y2 + Y3 +Y4 = I
Yi =0,1 i = 1,4
 461

The relaxed LP solution of this problem is Z = 2, Yl = Y3 = 0.5, Y2 = Y4 = O. If a

standard branch and bound search is performed, 4 nodes are required for the enumeration
as shown in Fig. 3a. However, if we instead treat the last constraint as a generalized upper
bound constraint, only two nodes are enumerated as shown in Fig. 3b.

Z=2.0

Z=4
Y4 =1

®
7 =2.5 1

\ =40

4 Z = Infeas.
Z=3

(a) Branching on individual variables

Z=3

(b) Branching with generalized upper bounds

Fig. 3 Standard branching rule and generalized upper bounds

Closely related to the generalized upper bound constraints, are the special ordered
sets (see [7, 87]). The most common are the SOSI constraints that have the form,

x= Lam (5)
ieI

in which the second constraint is denoted as a reference row where x is a variable and ai
are constants with increasing value. In this case the partitioning of the 0-1 variables at each
node is performed according to the placement of the value of the continuous variable x
 462

relative to the points ai. SOS2 constraints are those in which exactly two adjacent 0-1
variables must be nonzero, and they are commonly used to model piecewise linear concave
functions. Again, considerable reductions in the enumeration can be achieved with these
types of constraints.

Another important capability in branch and bound codes are preprocessing

techniques that have the effect of fixing variables, eliminating redundant constraints, adding
logical inequalities, tightening variable bounds and/or performing coefficient reduction (see
[12, 15, 47]). A simple example of coefficient reduction is for instance converting the
inequality 2Yl + Y2 ~ 1 into Yl + Y2 ~ 1 which yields a tighter representation in the 0-1
polytope. An example of a logical constraint are minimum cover constraints. For instance,
given the constraint 3Yl + 2Y2 + 4Y3 :::; 6, Yl + Y3:::; 1 is a minimum cover since it
eliminates the simultaneous selection of Yl = Y3 = 1 which violates this constraint.
Preprocessing techniques can often reduce the integrality gap of an MILP but their
application is not always guaranteed to reduce the computation time.

Although the LP based branch and bound method is the dominant method for MILP
optimization, there are other solution approaches which often complement this method.
These can be broadly classified into three major types: cutting plane methods,
decomposition methods and logic based methods. Only a very brief overview will be given
for these methods.

The basic idea of the original cutting plane methods was to solve a sequence of
successively tighter linear programming problems. These are obtained by generating
additional inequalities that cut-off the fractional integer solution. Gomory [26] developed a
method for generating these cutting planes, but the computational performance tends to be
poor due to slow convergence and the large increase in size of the LP SUbproblems. An
alternative approach is to generate strong cutting planes that correspond to facets, or faces
of the integer or mixed-integer convex hull. Strong cutting planes are obtained. by
considering a separation problem to determine the strongest valid inequality that cuts off the
fractional solution. This, however, is computationally a difficult problem (it is NP-hard)
and for this reason unless one can obtain theoretically these cuts for problems with special
structure, only approximate cutting planes are generated. Also, strong cutting planes are
generated from the LP relaxation and during the branch and bound search to tighten the LP.
Crowder et al [15] developed strong cutting planes for pure integer programming problems
by considering each constraint individually and treating each of them as knapsack
 463

problems. Van Roy and Wolsey [89] considered special network structures for MILP
problems to generate strong cutting planes. In both cases, substantial improvements were
obtained in a number of test problems.

A more recent approach for cutting plane methods has been based on the important
theoretical result that it is possible to transform an unstructured MILP problem into an
equivalent LP problem that corresponds to the convex hull of the MILP. This involves
converting the MILP into a nonlinear polynomial mixed integer problem which is
subsequently linearized through variable transformations (45, 80]. Unfortunately, the
transformation to the LP with the convex hull is exponential. However, these
transformations can be used as a basis for generating cutting planes within a "branch and
cut" enumeration, and this is for instance an approach that is being explored by Balas et al
[5].

As for decomposition methods for MILP, the most common method is Benders
decomposition [8]. This method is based on the idea of partitioning the variables into
complicating variables (commonly integer variables in the MILP) and noncomplicating
variables (continuous variables in MILP). The idea is to solve a sequence of LP
subproblems for fIxed complicating variables yk,

Zk = min cTx + bTyk (LPB)

s.t. Ax::;; d - Byk
x~O

and master problems that correspond to projections in the space of the binary variables and
that are based on dual representations of the continuous space. The form of the master
problem given K feasible and M infeasible solution points for the subproblems is given by:

zf.= min a
a ~O"k) T(d - Byk) k = 1, .. K (MB)
(J.Lm) T(d - Byk)::;; ° m = 1, .. M
aeR', ye (O,1)m

Since the master problem provides valid lower bounds and the LP subproblems
upper bounds, the sequence of problems is solved until equality of the bounds is achieved.
 464

Benders decomposition has been successfully applied in some problems (eg. see [24], but
it can also have very slow convergence if the LP relaxation is not tight (see [46]).
Nevertheless, this method is in principle attractive in large multiperiod MlLP problems.
Finally, another type of decomposition techniques are Lagrangean relaxation methods
which are applied when complicating constraints destroy the special structure of a problem.

Logic based methods were developed by taking advantage of the analogy between
binary and boolean variables. Balas [3] developed Disjunctive Prograrruning as an alternate
form of representation of mixed-integer programming problems. MlLP problems are
formulated as linear programs with disjunctions (sets of constraints of which at least one
must be true). Balas [4] characterized the family of all valid cutting planes for a disjunctive
program. Using these cuts, disjunctions were re-expressed in terms of binary variables
and the resulting mixed-integer problem is solved.

Another class of logic based methods are based on using symbolic inference
techniques for the solution of pure integer programming problems. Hooker [32]
demonstrated the analogy between unit resolution and first order cutting planes. leroslow
and Wang [35] solved the satisfIability problem using a numerical branch and bound based
scheme but solving the nodal problems using unit resolution. An alternate symbolic based
branching rule was also proposed by these authors. Motivated by the above ideas, [62]
considered the incorporation of logic in general mixed-integer programming problems in
the form of redundant constraints to express with logic propositions the relations among
units in superstructures. Here one approach is to convert the logic constraints into
inequalities and add them to the MILP. Although this has the effect of reducing the
integrality gap, the size of the problem is often greatly increased [63]. Therefore, these
authors considered an alternate scheme in which symbolic inference techniques were used
on a the set of logical constraints which are expressed in either the disjunctive or
conjunctive normal form representations. The idea is to perform symbolic inference at each
node during the branch and bound procedure in order to perform branching on the variables
so as to fix additional binary variables. Orders of magnitude reductions have been reported
by these authors using this approach [64].

Finally, it should be noted that the more recent computer codes for MlLP, such as
OSL (IBM, 1992) and MINTO [73] have an "open software architecture" that gives the
user considerably more flexibility to control the branch and bound search. For instance,
 465

these codes allow the addition of cutting planes and modification of branching rules
according to procedures supplied by the user.

Mixed-integer Nonlinear Programming (MINLP)

Although the problem (MIP) given earlier in the paper corresponds to an MINLP
problem, for most applications the problem is linear in the 0-1 variables and nonlinear in
the continuous variables x; that is.

min Z = f(x) + cTy

S.t. h(x) = 0 (MINLP)
g(x) + By s; 0
xeRD, ye{O,1)m

This mixed-integer nonlinear program can in principle also be solved with the
branch and bound method presented in the previous section ([31, 50, 11 D. The major
difference here is that the examination of each node requires the solution of a nonlinear
program rather than the solution of an LP. Provided the solution of each NLP subproblem
is unique, similar properties as in the case of the MILP would hold with which the rigorous
global solution of the MINLP can be guaranteed.

An important drawback of the branch and bound method for MINLP is that the
solution of the NLP subproblems can be expensive since they cannot be readily updated as
in the case of the MILP. Therefore, in order to reduce the computational expense involved
in solving many NLP subproblems, ~e can resort to two other methods: Generalized
Benders decomposition [23] and Outer-Approximation [18]. The basic idea in both
methods is to solve an alternating sequence of NLP subproblems and MILP master
problems. The NLP subproblems are solved by optimizing the continuous variables x for
a given fixed value of y, and their solution yields an upper bound to the optimal solution
of (MINLP). The MILP master problems consist of linear approximations that are
accumulated as iterations proceed, and they have the objective of predicting new values of
the binary variables y as well as a lower bound on the optimal solution. The alternate
sequence of NLP subproblems and MILP master problems is continued up to the point
where the predicted lower bound of the MILP master is greater or equal than the best upper
bound obtained from the NLP subproblems.
 466

The MILP master problem in Generalized Benders decomposition (assuming

feasible NLP subproblems) is given at any iteration K by:

7IS .
'"'UB =mm a (MGB)
s.t. a ~ f(xk) + cTy + (~k)T [g(xk) + By] k=1,2 ... K
aeRl , ye {O,l}ID

where a is the largest Lagrangian approximation obtained from the solution of the K NLP
subproblems; xk and ~k correspond to the optimal solution and multiplier of the kth NLP
subproblem; z& corresponds to the predicted lower bound at iteration K.

In the case of the Outer-Approximation method the MILP master problem is given
by:

~A=mina (MOA)
s.t. a ~ f(xk) + Vf(xk)T(x-xk) + cTY)
°
°
TkVh(xk) T (x-xk) ::;; k=I,2, ... K
g(xk) + Vg(xk)T(x-xk) + By::;;
ae R I , xe R n , ye {O, 1 }m

where a is the largest linear approximation of the objective subject to linear approximations
of the feasible region obtained from the solution of the K NLP subproblems. Tk is a
diagonal matrix whose entries tfi =sign (Ah where A~ is the Lagrange multiplier of
equation hi at iteration k, and is used to relax the equations in the form of inequalities [37].
This method has been implemented in the computer code DICOPT [38].

Note that in both master problems the predicted lower bounds, ~B' and ~A'
increase monotonically as iterations K proceed since the linear approximations are refined
by accumulating the lagrangian (in MGB) or linearizations (in MOA) of previous iterations.
It should be noted also that in both cases rigorous lower bounds, and therefore
convergence to the global optimum, can only be ensured when certain convexity conditions
hold (see [23, 18]).

In comparing the two methods, it should be noted that the lower bounds predicted
by the outer approximation method are always greater than or equal to the lower bounds
predicted by Generalized Benders decomposition. This follows from the fact that the
Lagrangian cut in GBD represents a surrogate constraint from the linearization in the OA
 467

algorithm [60]. Hence, the Outer-Approximation method will require the solution of fewer
NLP subproblems and Mn..P master problems (see example 4 later in the paper). On the
other hand, the Mn..P master in Outer-Approximation is more expensive to solve so that
Generalized Benders may require less time if the NLP subproblems are inexpensive to
solve. As discussed in [72], fast convergence with GBD can only be achieved if the NLP
relaxation is tight.

As a simple example of an MINLP consider the problem:

min Z = Yl + 1.5Y2 + 0.5Y3 + Xl 2 + X2 2

s.t. (XI-2)Lx2~0
Xl - 2Y1 ~O
Xl - x2 - 4(1-Y2) ~ 0
xI-(1-Yl)~O
x2 - Y2 ~ 0 (6)
Xl + x2~ 3Y3
YI + Y2 + Y3 ~ I
o~ Xl ~ 4, 0 ~ x2 ~ 4
YI,Y2,Y3=0,1

Objective function

10~~._._ Upper
bound
Upper bound
.. ?BD
OA _ •···· ..··::0
.............
Lower ••••
bound 0 ....... - ..... ·.. ··.0··.. Lower bound
-5 OA... GBD
.,.•.
-10 • 0 0'

,1/
-IS

-20
...........
O·
~~--------~--------~--------~--lrernriom

Fig. 4 Progress of iterations of OA and GBD for MINLP in (6)

Note that the nonlinearities involved in problem (6) are convex. Fig. 4 shows the
convergence of the OA and the GBD methods to the optimal solution using as a starting
 468

point Yl = Y2 = Y3 = 1. The optimal solution is Z =3.5, with Yl = 0, Y2 = 1, Y3 = 0, xl =

1, x2 = 1. Note that the OA algorithm requires 3 major iterations, while GBD requires 4,
and that the lower bounds of OA are much stronger.

Other related methods for MINLP include the extension of the OA algorithm by
Yuan et at [99] who considered nonlinear convex terms for the 0-1 variables, and the
feasibility technique by Mawekwang and Murtagh [48] in which a feasible MINLP solution
is obtained from the relaxed NLP problem. The latter method has been recently extended
by Sugden [84].

In the application ofGeneraIized Benders decomposition and Outer-Approximation,

two major difficulties that can arise are the computational expense involved in the master
problem if the number of 0-1 variables is large, and non-convergence to the global
optimum due to the nonconvexities involved in the nonlinear functions.

To circumvent the first problem, Quesada and Grossmann [59] have proposed for
the convex case an LPINLP based branch and bound method in which the basic idea is to
integrate the solution of the MILP master problem and the NLP subproblems which are
assumed to be inexpensive to solve. This is accomplished by a tree enumeration in which
an NLP is ftrst solved to construct an initial linear approximation to the problem. The LP
based branch and bound search is then applied; however when an integer solution is found
a new NLP subproblem is solved from which new linear approximations are derived which
are then used to update the open nodes in the tree. In this way the cold start for a new
branch and bound tree for the MILP master problem is avoided. It should be noted that this
computational scheme can be applied to Generalized Benders and Outer-Approximation.
As mentioned before the latter will yield stronger lower bounds. However, in this
integrated branch and bound method the size of the LP subproblems can potentially become
large. To handle this problem Quesada and Grossmann [59] proposed the use of partial
surrogates that exploit the linear substructures present in an MINLP problem.

In particular, consider that the MINLP has the following structure,

Z =min cTy + aTw + r(v)

st Cy+Dw+t(v)~O (MINLP')
Ey+Fw+Gv ~b
y E Y, w E W, VE V
 469

in which the equality constraints are relaxed to inequalities according to the matrix Tk and
included in the inequality set. Here the continuous variables x have been partitioned in two
subsets w and v such that the constraints are divided into linear and nonlinear constraints.
and the continuous variables into linear and nonlinear variables. In this representation
[Dw + t(v)]
f(x)=aT w + rev). BT = [ClEF. g(x)= Fw + Gv • and X= WxV. By constructing a
partial surrogate constraint involving the linearization of the nonlinear terms in the objective
and nonlinear constraints. the modified master problem has the form:

ZLK = min ~ (MMOA)

st cTy+aTw+b-n=O
~ <'! r(vk) + O,k )T [ Cy + Dw + t(vk )] - (Jlk)T G ( v - vk) k=l .... K
Ey +Fw +Gv :;;b
y e Y. w e W. ve V. n e RI. ~ e R'

where Ak and Jlk are the optimal multipliers of the kth NLP subproblem. It can be seen that
as opposed to the Benders cuts. the linearizations are defined in the full space of the
variables. requiring only the addition of one new constraint for the nonlinear terms. It can
be shown that the lower bound ZLK predicted by the above master problem is weaker than
the one of OAt but stronger than the one by GBD. Computational experience has shown
that the predicted lower bounds are in fact not much weaker than the ones by the OA
algorithm.

As for the question of nonconvexities. one approach is to modify the definition of

the MILP master problem so as to avoid cutting off feasible mixed-integer solutions.
Viswanathan and Grossmann [92] proposed an augmented-penalty version of the MILP
master problem for outer-approximation. which has the following form:

K
z& = min n + L (pk)T(pk + qk)
k=' (MOAP)
S.t. n <! f(xk) + Vf(xk)T(x_Xk) + CTY)
TkVh(xk)T(x_xk):;; pk k=1.2 .... K
g(xk) + Vg(xk)T(x_xk) + By:;; qk
neR'. xeRD.ye{O.l}m
 470

in which the slacks pk, qk, have been added to the function linearizations, and in the
objective function with weights pk that are sufficiently large but fmite". Since in this case
one cannot guarantee a rigorous lower bound, the search is tenninated when there is no
further improvement in the solution of the NLP subproblem. This method has been
implemented in the computer code DICOPT++ which has shown to be successful in a
number of applications. It should also be noted that if the original MINLP is convex the
above master problem reduces to the original OA algorithm since the slacks will take a
value of zero.

An important limitation with the above approach is that it does not address the
question whether the NLP subproblems may contain multiple local solutions. Recently
there has an important effort to address the global optimization of nonconvex nonlinear
programming problems. The current methods are either stochastic or detenninistic in
nature. In the fonner, generally no assumption about the mathematical structure of the
problem is made. Simulated annealing js an example of a method that belongs to this
category which in fact has been applied to batch process scheduling[43. 56]. This method
however has the disadvantages that no strict guarantee can be given about global optimality
and that its computational expense can be high. Deterministic methods require the problem
to have some particular mathematical structure that can be exploited to ensure global
optimality.

Floudas and Visweswaran [21] developed a global optimization algorithm for the
solution of bilinear programming problems. Valid lower and upper bounds on the global
optimal solution are obtained through the solution of primal and relaxed dual problems.
The primal problem arises by fixing a subset of complicating variables which reduces the
bilinear NLP into an LP subproblem." The relaxed dual problems arise from the master
problem of GBD but in which the Lagrangian function is linearized and partitioned into
subregions to guarantee valid lower bounds. An implicit partition of the feasible space is
conducted to reduce the gap between the lower and upper bounds. A potential limitation of
this method is that the number of relaxed dual problems to be solved at each iteration can
grow exponentially with the number of variables involved in the nonconvex terms.

Another approach for solving nonconvex NLP problems in which the objective
function involves bilinear terms is the one presented by AI-Khayyal and Falk [1]. These
authors make use of the convex envelopes for the individual bilinear terms to generate a
 471

valid lower bound on the global solution. An LP underestimator problem is imbedded in a

spatial branch and bound algorithm to find the global optimum. Sherali and Alameddine
[78] presented a reformulation-linearization technique which generates tight LP
underestimator problems that dominate the ones of AI-Khayyal and Falk. A similar branch
and bound search is conducted to fmd the global solution. Although this method requires
the enumeration of few nodes in the branch and bound tree. it has the main disadvantage
that the size of the LP underestimator problems grows exponentially with the number of
constraints.

Swaney [85] has addressed the problem in which the objective function and
constraints are given by bilinear terms and separable concave functions. A comprehensive
LP underestimator problem provides valid lower bounds that are used within a branch a
bound enumeration scheme in which the partitions do not increase exponentially with the
number of variables.

Quesada and Grossmann [60] have considered the global optimization of

nonconvex NLP problems in which the feasible region is convex and the objective involves
rational and/or bilinear terms in addition to convex functions. The basic idea is based on
deriving an NLP underestimator problem that involves both linear and nonlinear estimator
functions that provide an exact approximation of the boundary of the feasible region. The
linear understimators are similar to the ones by AI-Khayyal and Falk [1]. but these are
strengthened in the NLP by nonlinear convex underestimators. The NLP underestimator
proble!ll. which allows the generation of tight lower bounds of the original problem. is
coupled with a spatial branch an bound search procedure for fmding the global optimum
solution.

Modelling and reformulation

One of the difficulties involved in the application of mixed-integer programming

techniques is that problem formulation is not always trivial. and that the way one
formulates a problem can have a very large impact in the computational efficiency for the
solution. In fact, it is not uncommon that for a given problem one formulation may be
essentially unsolvable. while another formulation may make the problem much easier to
solve. Thus. model formulation is a crucial step in the application of mixed-integer
programming techniques.
 472

While model fonnulation still remains largely an art, a number of guiding principles
are starting to emerge that are based on a better understanding of polyhedral theory in
integer programming (see [51]). In this section we will present an overview of modelling
techniques and illustrate them with example problems. These techniques can be broadly
classified into logic based methods, semi-heuristic guidelines, refonnulation techniques and
linearization techniques.

It is often the case in mixed-integer programming ·that it is not obvious how to

fonnulate a constraint in the first place, let alone fonnulate the best fonn of that constraint.
Here the use of propositional logic and its systematic transfonnation to inequalities with 0-1
variables can be of great help (e.g. see [98, 14,62]). In particular, when logic expressions
are converted into the conjunctive nonnal fonn. each clause has the fonn,

PI v P2 v ... v Pm (7)

where Pi is a proposition and v is the logical opemtor OR. The above clause can be readily
tmnsfonned into an inequality by relating a binary variable Yi to the truth value of each
proposition Pi (or 1 - Yi for its negation). The fonn of the inequality for the above clause
is,

YI + Y2 + ... + Ym ~ 1 (8)

As an example consider the logical condition, PI v P2 ~ P3 which when

converted into conjuctive normal fonn yields, (-, PI v P3) 1\ (..., P2 v P3). Each of the
two clauses can then be tmns1ated into the inequalities,

1-YI+Y3~·1 Y3 ~Yl (9)

1 - Y2 + Y3 ~ 1 or Y3 ~ Y2

Similar procedures can be applied when deriving mixed-integer constraints.

Once constraints have been fonnulated for a given problem, the question that arises
is whether alternative fonnulations might be better suited for the computation, and here the
first techniques to be considered are semi-heuristic guidelines. These are rules of thumb on
how to fonnulate "good" models. A simple example are variable upper bound constraints
for problems with fixed charges,
 473

(10)

Here it is well known that although for representation purposes the upper bound U can be
large, one should try to select the smallest valid bound in order to avoid a poor LP
relaxation. Another well known example is the constraint that often arises in multiperiod
MILP problems (selecting unit.z implies possible operation Yi in period i, i=I,2, ...n),
n
LYi -nzSO (11)
i=!

Here the disaggregation into the constraints,

Yi - z SO i=I,2, .. n (12)

will produce a much tighter representation (in fact the convex hull of the 0-1 polytope).
These are incidentally the constraints that would be obtained if the logical conditions for
this constraint are expressed in conjuctive normal form.

The main problem with the disaggregated form of constraints is the potentially large
number of them when compared with the aggregated form. Therefore, one has to balance
the trade-offs between model size and tighter relaxations. For example [93] report a model
in which the disaggregated variable upper bound constraints were used in the form

Wijsn S Uijsn Yjsn V i,j, s, n (13)

An equivalent aggregated form of the above constraints is

Lw ijsn S L U ijsn Yjsn v j, s, n (14)

When the first set of constraints is used the model involved 708 constraints and required
233 CPUsec using SCICONIC on a VAX 6320. When the second set of constraints is
used, the model required only 220 constraints, but the time was increased to 649 sec
because of the looser relaxation of (14).

While the above modelling schemes are somewhat obvious, there are some which
are not. A very good example arises in the MILP scheduling model of [40]. In this model,
the following constaints apply:
 474

(a) At any time t, an idle item of equipment j can only start at most one task
i E Ij-

(b) If the item j does start perfonning a given task i E ~, then it cannot start any
other task until the current one is finished after Pi time units.

Kondili et al [46] formulated the two above constraints as:

I, Wijt ~ 1 'v' j, t
ielj (15)
t+prI
I, I, Wi'jt' - 1 s M(1 - Wijt) 'v' i, j E Kio t
t'=t i'elj

where M is a suitably large number. Note that the second constraint has the effect of
imposing condition (b) if Wijt = 1. As one might expect the second constraint yields a poor
relaxation due to the effect of the "big M". Interestingly, [77] found an equivalent
representation for the above two constraints, which is not only much tighter, but requires
much fewer constraints! These are given by,

t-pi+I
I, ~ Wijt'S 1 'v' j, t (16)
ielj t=t

Thus, this example clearly shows that formulation of a "proper" model is not always trivial or
even well understood. Nevertheless, an analysis of the problem based on polyhedral theory
can help one understand the reason for the effectiveness of the constraint. A detailed proof
for constraint (16) is given in the Appendix.

However, not everything in MILP modelling is an art. A more rational approach that
has been emerging is the idea of reformulation techniques that are based on variable
disaggregation (e.g. see [61, 33, 34,]), and which have the effect of tightening the LP
relaxation. The example par excellence is the lot sizing problem that in its "naive" form is
given by the MILP (see [51]):
 475

NT
min I. (PtXt + ht St + CtYt>
t=1
S.t. St-I + Xt = dt + St t = I, NT (17)
Xt:S; xUYt t = I, NT
So = 0
St,Xt~ 0, YtE (O,I} t= I, NT

where Xt is the amount to be produced in period 1, Yt is the associated 0-1 variable, and St is
the the inventory for period t; Ct.Pt.ht. are the set-up, production and storage costs for time
period t, t = I, NT.

As has been shown by Krarup and Bilde [41] the above MILP can be reformulated by
dis aggregating the production variables Xt into the variables qtt, to represent the amount
produced in period t to satisfy the demand in period 't ~ t; that is,

(18)

The MILP is then reformulated as,

NT NT NT
min I. I. (Pt + ht + ht+1 + ...+ ht-I) qn + I. CtYt (19)
t=1 t=t t=1

t = I, NT
t=1
qt't :S; dtYt t = I,NT, 't = t,NT
qt't ~ 0, Yt = (O,l)

As it turns out this reformulation yields the absolute tightest LP relaxation since it yields 0-1
values for the Y variables; thus this problem can be solved as an LP and there is no need to
apply a branch and bound search as there is in the original MILP (17). It should be noted that
although this example is quite impressive, it is not totally surprising from a theoretical
viewpoint. The lot sizing problem is solvable in polynomial time and therefore one would
expect that it should be possible to formulate this problem as an LP that is polynomially sized
in the number of variables and constraints. It should be noted that the lot sizing problem is
often embedded in MILP planning and scheduling problems with which one can reformulate
these problems to tighten the LP relaxation as discussed in [71].
 476

Finally, another case that often arises in the modelling of mixed-integer problems are
nonlinearities such as bilinear products of 0-1 variables or products of 0-1 with continuous
variables. Nonlinearities in binary variables usually involve the transformation of a nonlinear
function into a polynomial function of 0-1 variables and then transforming the polynomial
function into a linear function of 0-1 variables [79]. For cross products between binary and
continuous variables [58] proposed a linearization method which was later extended by
Glover [25]. The main idea behind these linearization schemes was the introduction of a new
continuous variable to represent the cross product. The equiValence between the bilinear term
and the new variable was enforced by introducing a set of equivalence constraints. For the
specific case in which the model has a multiple choice structure, an efficient linearization
scheme was proposed by Grossmann et al [29]. This scheme compared to the one proposed
by Glover gives tighter LP relaxations with fewer number of constraints. The multiple choice
structure usually arises in discrete design models, in which the design variables instead of
being continuous, take values from a finite set. Batch process design problems often involve
discrete sizes and as such the latter linearization scheme is well suited. As an example,
consider the the bilinear constraints:

N(i)
aij L disYisvj - ~ijWj ~ 0 je JCi), i=l, .. n (20)
5=1

in which Yis is a 0-1 variable and Vj is continuous, and where the following constraint holds:

N(i)
L Yis=! (21)
s=!

In order to remove the bilinear terms YisVj in (20), define the continuous variables Vijs such
that

N(i)

Vj = L Vijs je J(i) , i=l, .. n (22)

s=1
VjL Yis ~ Vijs ~ vp Yis je J(i), s=l, N(i), i=l..n (23)

where VjL, Vju are valid lower and upper bounds. Using the equations in (21) to (23), the
constraints in (20) can be replaced by the linear inequalities

N(i)

aij L disvijs - ~ijWj ~ 0 je J(i) , i=I, .. n (24)

s=1
 477

The bilinear constraints in (20) can also be linearized by considering in addition to the
inequalities in (24), the following constraints proposed by Glover [25]:

VjL Yis ::; Vijs ::; Vp Yis

Vijs ~ Vj - vp(1- Yis) jeJ(i), s=l, N(i), i=l..n (25)
Vijs ::; Vj - VjL(1- Yis)

This linearization, however, requires almost twice as many constraints as (22), (23) and (24)
Furthermore, while a point (Vijs, Vj, Yis) satisfying (22) and·(23) satisfies the inequalities in
(25), the converse may not be true. For instance, assume a non-integer point Yis such that
Vijs = vPYis. Using (21) it follows from (25) that

(26)

while (22) yields Vj = VjU. Thus, the inequalities in (25) may produce a weaker LP
relaxation.

For the case when the bilinear constraints in (20) are only inequalities, Torres [88]
has shown that it is sufficient to consider the following constraints from (25):
VjL Yis ::; Vijs
Vijs ~ Vj - VjU(1- Yis) jeJ(i), s=l, N(i), i=l..n (27)

which requires fewer constraints than the proposed linearization in (22) and (23).
However, the above inequalities also can produce a weaker LP relaxation. For instance,
setting Vijs = VjLyis for a non-integer point Yis yields,

Vj ::; vp - (vp - vr) Yis (28)

while (22) yields Vj = vF

While the modelling techniques described in this section have been mostly aimed at
MILP problems, they are of course also applicable to MINLP problems. One aspect
however, that is particular to MINLP problems are the modelling of nonlinearities of the
continuous variables. In such a case it is important to determine whether the nonlinear
constraints are convex or not. If they are not, the first attempt should be to try to convexify
the problem. The most common approach is to apply exponential transformations of the
form x = exp(u), where x are the original continuous variables and u the transformed
 478

variables; if the original nonlinearities correspond to posynomials these tranformations will

lead to convex constraints. A good example is the optimal design of multiproduct plants with
single product campaigns [28], with which [39] were able to rigorously solve the MINLP
problem to global optimality with the outer-approximation algorithm. When no
transformations can be found to convexify a problem, this does not necessarily mean that the
relaxed NLP has multiple local optima. However, non convexities in the form of bilinear
products and rational terms are warning signals that should not be ignored. In this case the
application of a global optimization method such as the ones· described previously will be the
only way to rigorously guarantee the global optimum.

Finally, it should be noted that another aspect in modelling is the computer software
that is available for formulating and solving mixed-integer optimization problems. Modelling
systems such as GAMS [13] and AMPL [22] have emerged as major tools in which
problems can be specified in algebraic form and automatically interfaced with codes for
mixed-integer linear and nonlinear optimization (e.g. MINOS, ZOOM, SCICONIC, OSL,
CPLEX, DICOPT++). Modelling tools such as these can greatly reduce the time that is
required to test and prototype mixed-integer optimization models.

Examples

In this section we will present several examples to illustrate a number of points and the
application of techniques for mixed-integer programming in batch processing.
Example 1: The MILP for the State Task Network model for the scheduling of
batch operations by Kondili et al. [40] has been used to compare the performance of three
MILP codes: ZOOM an academic code, and OSL and SCICONIC which are commercial
codes. This example which has 5 tasks, 4 units, 10 time periods and 9 states (see Fig. 5),
also demonstrates the effect of modelling schemes on the solution efficiency. The
objective is to maximize the production of the two final products. The resulting MILP
model, which incorporates the constraints in (16), involves 251 variables (80 binary) and
309 constraints (see [77]). The results of the benchmark comparison between the three
codes for this problem are shown in Table I. The problems were solved to optimality (0%
gap) by using GAMS as an interface. As can be seen in Table I the performance of the
codes is quite different. SCICONIC had the lowest computing requirements: about less
than a tenth of the requirements of ZOOM.
 479

01
Feed A
Heating ~o--
Hot A L-..,-_--'

Feed B

0--

Fig.5 State·task network for example problem

It should be noted that [77] solved this problem with their own branch and bound
method in two forms. In the first the MILP was identical as the one solved in this paper.
In this case 1085 nodes and 437 secs on a SUNI- Sparcstation were required to solve the
problem within 1% of optimality. In the second form, the authors applied a solution
strategy for reducing the size of the relaxed LP and for reducing degeneracies. In this case
only 29 nodes and 7 secs were required to solve the problem, which represents a
performance comparable to the one by SCICONIC.

To illustrate the effect that alternate formulations may have, two cases were
considered and the results are shown in Table 2a. Firstly, realizing thatthe objective
function does not contain any binary variable, the second column involves the addition of a
penalty to the objective function in which all the binary variables are multiplied by a very
small number so as not to affect the optimum solution. The idea is simply to drive the 0-1
variables to zero to reduce the effect of degeneracies. In the third column, 18 logic cuts in
the form of inequalities have been added to the MILP model to reduce the relaxation gap
[63]. These logic cuts represent connectivity of units in the state task network. For
example, in the problem of Fig. 5, since no storage of impure C is allowed, the separation
step has to be immediately performed after reaction 3. As can be seen from Table 2a, both
modelling schemes lead to a substantial improvement in the solution efficiency with OSL,
while with SCICONIC only the addition of logic cuts improves the solution efficiency.
Furthermore, the effect of adding these logic cuts in this problem have been studied for the
 480

case of 10, 20, 40 and 50 time periods. The results, shown in Table 2b, demonstrate an
increase in the effectiveness of the logic cuts in improving the efficiency of the branch and
bound procedure. The reduction in the number of nodes required in the branch and bound
search due to the logic cuts increases from a factor of 3 for the 10 period case to a factor of
more than 6 for the 40 period case. The 50 time period problem, with 1251 variables (400
binary) and 1509 constraints could not be solved by OSL within 100,000 iterations and 1
hour of CPU time on the mM POWER 530. With the addition of the logic cuts, the
problem is solved in 158.84 sec requiring only 698 nodes and 5017 iterations.

Table 1. Comparison with several MILP codes

nodes iterations CPU time *

ZOOM 410 7866 39.44
OSL 350 918 14.85
SCICONIC 61 318 3.63

* IBM POWER 530

Table 2a. Computational results with modified formulations (OSL 1 SCICONIC)

Original Model with altered Model with

Model Objective Function Loeic Cuts

number of nodes 350/61 40161 108/33

number of iterations 918/318 336/318 6201233
CPU time * 14.85/3.63 2.5113.65 5.98/2.13

Relaxed Optimum 257.2 257.2 257.2

Integer Optimum 241 241 241

* sec IBM POWER 530

 481

Table 2b. Example 1: Effect of logic cuts for different time periods

Original Model
Model with Lo~ic Cuts
10 Time Periods
251 variables
80 binarv

Constraints 309 327

Number of nodes 350 108
Number of Iterations 918 620
CPU Time • 14.85 5.98

20 Time Periods
501 variables
160 binarv

Constraints 609 643

Number of nodes 123 67
Number of Iterations 755 658
CPU Time' 10.22 7.81

40 Time Periods
1001 variables
320 binarv

Constraints 1209 1279

Number of nodes 2098 315
Number of Iterations 25964 3423
CPU Time • 424.68 67.17

SO Time Periods
1251 variables
400 binarv

Constraints 1509 1597

Number of nodes >20.000 698
Number of Iterations >100.000 5017
CPU Time' >3.600 158.84

• sec IBM POWER 530

Example 2. In order to illustrate the effect of preprocessing and the use of SOSI
constraints, consider the design of multiproduct batch plants with one unit per stage,
operating with single product campaigns, and where the equipment is available in discrete
sizes [93]. The MILP model is as follows:
 482

(RPl)

i=l, .. ,N , j=l, .. ,M

j=I, .. ,M
j=I, .. ,M , s=I, .. ,nsj

The example considered involves a plant with 6 stages and 5 products. To illustrate the
effect that the number of discrete sizes has in the size of model (RPl) as well as in the
computational performance, three problems one with 8, one with 15 and another with 29
discrete sizes were considered.The MILP problems were solved using SCICONIC 2.11
(SCICONIC, 1991) through GAMS 2.25 in a Vax-6420.

Table 3. Computational results for example 2

# Dis. Sizes I Constraints I Variables I 0-1 Vars's I Cpu-time I Iterations I Nodes

WITHOUT SOSI, DOMAIN REDUCTION AND CUTTOFF

8 38 54 48 2.93 181 89

IS 38 96 90 25.09 985 731

29 38 180 174 44.94 1203 979

WITH SOSI,DOMAINREDUCTION AND CUTTOFF

8 38 46 40 1.93 57 53

IS 38 82 76 2.85 91 64

29 38 154 148 6.64 182 150

• In VAX- 6420 SECONDS

As seen from Table 3, the number of discrete sizes has a significant effect in the
number of 0-1 variables, and hence in the number of iterations and the CPU time. One can,
however, reduce significantly the computational requirements by performing a domain
reduction of the 0-1 variables through the use of bounds to fix a subset of them to zero,
treating the multiple choice constraints as SOSI constraints and applying and objective
function cutoff as described in [93]. As seen in Table 3, reductions of up to one order of
magnitude are achieved.
 483

Example 3. This example will illustrate how strong cutting planes may
significantly improve the computational performance of MILP problems with poor
continuous relaxations. A good example are jobshop scheduling problems. Consider the
case in which one has to schedule a total of 8 batches, 2 for each one of 4 products A, B,
C, D so as to minimize the makespan in a plant consisting of 5 stages. The processing
times for each product are given in Fig. 6, where it can be seen that not all products require
all the stages, and that they all require zero-wait transfer policy.

8 7
Sig 1

SIg2 ~ ~
SIg 3 ~
3
~
9
r--L,
w., u..
Stg4
4 4
SlgS .......... ~

PrdA 0 PrdB D PrdC. PrdD ~

TIme

Fig. 6 Processing times of products in various stages

As noted in [94] the makespan minimization problem described above can be formulated as
an Mn..P problem of the form:

min Ms (PI)

M
s. t Ms ;::Si + L t ik V'i
k=l
k k -' l
Sj - Si + W(1-Yijk);:: (L tik - L tj0 V' (i ,j, k') E C
k=l k=l
k k -' l
Si - Sj + WYij k;:: (L tj k - L ti 0 V' (i , j, k' )E C
k=l k=l
YijkE {O,l} V'i,j,k Si;::O V'i

In the above formulation the potential clashes at every stage are resolved with a pair of
disjunctive constraints that involve a large bound W. The difficulty with these constraints
is that they are trivially satisfied when the corresponding binary variables are relaxed,
which in tum yields a poor LP relaxation. For this example, the LP relaxation had an
objective value of 18 compared to the optimal integer solution of 41 , which corresponds to
 484

a relaxation gap of 56%. The MILP was solved with SCICONIC 2.11 on a Vax-6420
requiring 55 CPUsecs, and the solution is shown in Fig. 7. In order to improve the LP
relaxation basic-cut inequalities have been recently proposed by Applegate and Cook [2],
and they have the form,

L tikSik~Brktik+ L tiktjk 'v'k

ie T Ii eT.jeT. kj}
L tik(Ms-Si0~FTktik+ L ~k + L
ieT ieT Ii eT.jeT. i<.il
M
where Si k = the starting time of job i on machine k (Si k =Si + t i0 L
T = a subset of the set of jobs k= 1

Ejk =the earliest possible starting time of jon k (which is just the sum of
j's processing times on the machines before k)
En =the minimum of Ejk over all JET
Fjk =the minimum completion time of j after it is processed on k (which is
just the sum of j's processing times on the remaining machines)
FTk =the minimum of Ejk over all JET

The impact of these constraints in the MILP was very significant in this example. The LP
relaxation increased to 38 which corresponds to a gap of only 7%. In this case the optimal
solution was obtained in only 8 CPUsecs.

.-
Stg 1 "

Slg 2 ===== ~

SIg3 ~ I ~ b3
SIg4
I h
SIg5 r----I I I I I I
I
PrdA D PrdB [] PedC. PedD ~
41 Time

Fig. 7 Optimal schedule for example 3

Example 4. The optimal design of multiproduct batch plants with parallel units
operating out of phase (see Fig. 8) will be used to illustrate the computational performance
of the different MINLP algorithms.
 485

•• •
Fig. 8 Multiproduct batch plant with parallel units

Two different cases are considered. One consists of 5 products in a plant with 6
processing stages process with a maximum of 4 parallel units per stage (batch5). The
second one has 6 products in a plant with 10 stages and a maximum of 4 parallel units per
stage (batch6). The MINLP formulation and the data for these examples are reported in
[38] and the size of the problems is given in Table 4. The model is a geometric
programming problem that can be transformed into a convex MINLP through exponential
transformations.
Table 4. Data for Example 4.

problem Binary Continuous Constraints

variables variables

batch5 24 22 73
batch6 40 32 141

The GBD and OA algorithms were used for the solution of both examples and the
computational results are given in Table 5. The GBD algorithm was implemented within
GAMS, while the version of the OA algorithm used was the one implemented in
DICOPT++ with the augmented penalty. MINOS 5.2 was used for the NLP subproblems
and SCICONIC 2.11 for the MILP master problems. Note that in both cases the OA
algorithm required much fewer iterations than GBD which predicted very weak bounds and
a large number of infeasible NLP subproblems during the iterations. For problem batch5
both algorithms found the global optimal solution. For batch6, both algorithms also found
the same solution which however is suboptimal since the correct optimum is $398,580. In
the case GBD, the algorithm did not converge as it had a large gap between the lower and
upper bounds after 66 iterations. In the case of the OA algorithm as implemented in
DICOPT ++ the optimal solution was suboptimal due to the termination criterion used in
this implementation.
 486

Table 5. Computational results for Example 4.

GBD abwrithm OA algorithm
Iproblem Solution iterations CPU time* Solution iterntions CPU time*

batch5 $285.506 67 766.88 $285.506 3 26.94

batch6 $402.496 66+ 2527.2 $402.496 4 108.58

sec Vax 6' ~:lU + Lonverg ence ot bounds was not achleved

In both the above examples the solution of the MIi..P master problem in the OA
algorithm was of the order of 80%. A rigorous implementation of the OA algorithm for the
convex case [18] and the LPINLP based branch and bound algorithm by Quesada and
Grossmann [59] were also applied to compare their computational performance with
respect to the number of nodes that are required for the MILP master problem. The results
are given in Table 6. As can be seen. both algorithms required the solution of 4 and 10
NLP subproblems. respectively, and they both obtained the same optimal solution.
However. the LPINLP based branch and bound required a substantially smaller number of
nodes (36% and 16% of the number of nodes required by OA).

Table 6. Results on MILP solution step for problems in Example 4.

optimal Outer Approximation LPINLP branch and bound

Iproblem solution nodes NLP nodes NLP

batch 5 $285. 506 90 4 32 4

batch 6 $398. 580 523 10 84 10

Example 5. In order to illustrate the effect of nonconvexities. consider the design

and production planning of a multiproduct batch plant with one unit per stage. The
objective is to maximize the profit given by the income from the sales of the products minus
the investment cost. Lower bounds are specified for the demands of the products and the
investment cost is assumed to be given by a linear cost function. Since the sizes of the
vessels are assumed to be continuous. this gives rise to the following NLP model:

max P= ~Pini Bi - ~!lj Vj

1 J
s.t. Vj ~ Sij Bj i=l. N, j=I.M (NLPP)
Inj Ij ~ H
j

QL _Bj ~ 0 i=l, N
nj
Vj.Bj,nj ~ 0
 487

where ni and Bi is the number and size of the batches for product i, and Vj is the size of the
equipment at stage j. The fIrst inequality is the capacity constraint in terms of the size
factors Sij, the second is the horizon constraint in terms of the cycle times for each product
Ti and the total time H, and the last inequality is the specification on lower bounds for the
demands QiL . Note that the objective function is nonconvex as it involves bilinear terms,
while the constraints are convex. The data for this example are given in Table 7. A
maximum size of SOOO L was specifled for the units in each stage.

Table 7. Data for Example 7

Tj Pi QL Sij(Ukg)
Product (hrs) ($/Kg) (Kg) 1 2 3
A 16 15 80000 2 3 4
B 12 13 SOOOO 4 6 3
C 13.6 14 SOOOO 3 2 S
D 18.4 17 2S000 4 3 4
(Xl = SO, (X2 = 80, (X3 = 60 ($/L); H = 8,000 hrs

When a standard local search algorithm (MINOS S.2) is used for solving this NLP
problem using as a starting point nA=DB=nc=60 and no=300 the predicted optimum profIt
is $8,043,800/yr and the corresponding batch sizes and their number are shown in Table 8.

Table 8. Suboptimal solution for Example 8

A B C D
B 125u 833.33 1000 1250
n 79.15 60 50 289.868

Since the formulation in (NLPP) is nonconvex there is no guarantee that this

solution is the global optimum. This problem can be reformulated by replacing the
nonconvex terms by underestimator functions to generate a valid NLP underestimator
problem as discussed in [60]. The underestimator functions require the solution of LP
subproblems to obtain tight bounds on the variables, and yield a convex NLP problem with
8 additional constraints.

The optimal profit predicted by the nonlinear underestimator problem is

$8,128,IOO/yr with the variables given in Table 9. When the objective function of the
original problem (NLPP) is evaluated for this feasible point the same value of the objective
function is obtained proving that it corresponds to the global optimal solution. This
problem was solved on a mMIR6000-S30 with MINOS 5.2, and 1.6 secs were required to
solve the LP bounding problems and 0.26 secs to solve the NLP underestimator problem.
 488

It is interesting to note that both the local and global solutions had the maximum equipment
sizes. The only difference was in the number of batches produced for products A and D.

Table 9. Global optimwn solution for Example 4.

A B C D
B 1250 833.33 1000 1250
n 389.5 60 50 20

Concluding Remarks

This paper has given a general overview of mixed-integer optimization techniques for
the optimization of batch processing systems. As was shown with the review of previous
work. the application of these techniques has increased substantially over the last few years.
Also. as was discussed in the review of mixed-integer optimization techniques. a number of
new methods are emerging that have the potential of increasing the size and scope of the
problems to be solved. While in the case of MILP branch and bound methods continue to
playa dominant role. the use of strong cutting planes. reformulation techniques and the
integration of symbolic logic hold great promise for reducing the computational expense for
solving large scale problems. Also. it will be interesting to see in the future what impact
interior point methods will have on MILP optimization (see for instance [10]. for preliminary
experience). While the solution time of large LP problems can be greatly reduced. the
subproblems in branch and bound cannot be readily updated as is the case with simplex
based methods. As was also shown with the results. different computer codes for MILP can
show very large differences in performance despite the fact that they all rely on similar ideas.
This clearly points to the importance of issues such as software and hardware
implementation. preprocessing. numerical stability and branching rules. However. care must
also be exercised in comparisons because any given method or implementation for MILP may
exhibit wide variations in performance by changes in data for the same model.

In the case of MINLP. the application of this type of models is becoming more
widespread with the Outer-Approximation and Generalized Benders Decomposition methods.
The former has proved to be generally more efficient. although the latter is better suited for
exploiting the structure of problems (e.g. see [70]). Aside from the issue of problem size in
MINLP optimization. nonconvexities remain a major source of difficulties. However.
significant progress is being made in the global optimization of non convex NLP problems.
and this will surely have a positive effect on MINLP optimization in the future. Finally. as
has been emphasized in this paper. problem formulation for MILP and MINLP problems has
 489

often a very large impact in the efficiency of the computations. and in many ways still
remains an art for the application of these techniques. However. a better understanding of
polyhedral theory and establishing firmer links with symbolic logic may have a substantial
effect on how to systematically formulate improved models for mixed-integer problems.

Acknowledgment

The authors gratefully acknowledge financial support from the National Science Foundation
under Grant CBT-8908735 and CTS-9209671. and from the Engineering Design Research
Center at Carnegie Mellon.

Appendix: On the reduced set of inequalities of [77J

In order to prove the equivalence of constraint (15) and the one in (16) by Shah et al [771.
one must first state the following lemma

Lemma : The integer constraint.

(I)

is equivalent to and sharper than the set of integer constraints

Yl+Y2+···+YK $; (AO)
YK+l + Yl $; 1 (AI)
YK+I + Y2 S I (A2)

YK+l + YK $; I (AK)

Proof:

First we note that (1) can easily be seen to be equivalent to the constraints (AD) to (AK)
since in these at most one variable Ycan take an integer value of I. Multiplying constraint
(AD) by (K-I) and adding it with the constraints in (Al)-(AK) yields.

K(Yl+Y2+ ..+YK+l) $; K-l + K

Yl+Y2+ .. +YK+l $; 1+ (K-l)/K

Since Yi E {D.I). the right hand side can be rounded below to obtain the inequality

Yl+Y2+ .. +YK+l $; 1

Proof of constraint

We know. from (15)

 490

I, Wi,j,t S; 1 V' j, t Wijt E {O,I} V'i,j,t

iEij

ie. Wil,j,t+Wi2,j,t+,,+Winj,j,t S; for allj,t

If Pil > 1, since no unit can process two tasks simultaneously,

Wil,j,t-I+Wil,j,t S; 1
Wil,j,t-I+Wi2j,t S; 1

Wil,j,t-I+Winj,j,t S; 1

From the lemma, we get

Wil,j,t-I+Wil,j,t+Wi2,j,t+,,+Winj,j,t S;

Ifpi2> I

Wi2,j,t-l+Wil,j,t S; 1
Wi2,j,t-I+Wi2,j,t S; 1
..
Wi2,j,t-I+Winj,j,t S;

Also, W i2,j,t-l+ Wil,j,t-IS;

This leads to Wi2,j,t-I+Wil,j,t-l+Wil,j,t+Wi2,j,t+,,+Winj,j,t S;

Repeat for all Wi',j,t-I where Pi' > 1 to get

W il,j,t-l+ W i2,j,t-I+ .. +Winj,j,t-l+ W il,j,t+ Wi2,j,t+ .. +W inj,j,t S;

Now, if Pil > 2

W il,j,t-2+ Wil,j,t-l S; 1
W il,j,t-2+ W i2,j,t-l S; 1

W il,j,t-2+ Winj,j,t-l S;. 1

Wil,j,t-2+Wil,j,t S; 1
Wil,j,t-2+Wi2,j,t S; 1

W il,j,t-2+ W inj,j,t S; 1

From the lemma, we get

Wil,j,t-2+W il,j,t-I+Wi2,j,t-l+,,+W inj,j,t-I+Wil,j,t+W i2,j,t+,,+Winj,j,t S;

Repeat for all Wi'j,t-2 for Pi' > 2

Repeat for all Wi'j,t-3 for Pi' > 3

Repeat for all Wi',j,t-pi+ I for Pi' > Pi-l

Finally, we get
 491

W i1.l.~:pil+1 +.+ WiI,j,l-1 + W i1,j,l+ W i2,j.l-pi2+1 + W i2,j,l-1 +.+ W i2,j,l+ W inj,j.l-pnj+l+ W inj.j,l-
1+,+Winj,j,1 ~ 1

Grouping tenus in the above inequality yields

I I I

L Wil,j,I'+ L W i2,j,t' + ... + L Winj,j,l'

t'=I-pl+1 1'=I-p2+1 1'=I-pnj+1

Further summing over all i, we get the constraint by Shah et al. (1991)
I

I. I. W i.j,I' ~
ie Ij I'=I-pi+ I

References

1. AI-Khayyal, F.A. and Fallc, J.E. (1983) Jointly constrained biconvex programming, Mathematics of
Operations Research 8, 273-286.
2. Applegate D. and Cook W. (1991). A Computational Study of the Job-Shop Scheduling Problem,
ORSA Journal on Computing, 3, No.2, pp 149-156.
3. Balas, E (1974). "Disjunctive Programming: Properties of the Convex Hull of Feasible Points."
MSRR #348, Camegie Mellon University.
4. Balas, E. (1975). "Disjunctive Programming: Cutting Planes from Logical Conditions".
Nonlinear Programming 2, O. L. Mangasarian et al., eds., Academic Press, 279-312.
5. Balas, E., Ceria, S. and Comuejols, G. (1993). A Lift-and-Project Cutting Plane Algorithm for
Mixed 0-1 Programs. Mathematical Programming, 58 (3), 295-324
6. Balas, E., and Mazzola, 1.8. (1984). Nonlinear 0-1 Programming: Linearization Techniques.
Mathematical Programming, 30, 1-21.
7. Beale, E. M. L, and Tomlin, J. A. (1970). Special Facilities in a Mathematical programming
System for Nonconvex problems using Ordered Set of Variables, in Proceedings of the Fifth
International Conference on Operational Research, J. Lawrence, ed., Tavistock Publications, pp 447-
454.
8. Benders,1. F. (1962). Partitioning Procedures for Solving Mixed Integer Variables Programming
Problems, Numerische Mathematik, 4, 238-252.
9. Birewar D.B and Grossmann I.E (1990). Simultaneous Synthesis, Sizing and Scheduling of
Multiproduct Batch Plants,Ind. Eng. Chem. Res., Vol 29, Noll, pp 2242-2251
10. Borchers, 8. and Mitchell, J.E. (1991). Using an Interior Point Method in a Branch and Bound
Method for Integer Programming, R.PJ. Math. Report No. 195.
II. Borchers, B. and Mitchell, J.E. (1991). An Improved Branch and Bound Algorithm for Mixed-Integer
Nonlinear Programs, R.P.1. Math. Report No. 200.
12. Brearly, A.L., Mitra, G. and Williams, H.P. (1975). An Analysis of Mathematical Programming
Problents Prior to Applying the Simplex Method, Mathematical Programming, 8,54-83.
13. Brooke, A., Kendrick, D. and Meeraus, A. (1988). GAMS: A User's Guide. Scientific Press, Palo
Alto.
14. Cavalier, T. M. and Soyster, A. L. (1987). Logical Deduction via Linear Programming. IMSE
Working Paper 87-147, Dept. of Industrial and Management Systems Engineering, Pennsyvaoia
State University.
15. Crowder. H. P.• Johnson, E. L.. and Padberg. M. W. (1983). Solving Large-Scale Zero-One Linear
Programming Problems. Operations Research. 31.803-834.
16. Dakin. R. 1. (1965). A Tree search Algorithm for Mixed Integer Programming Problems, Computer
Journal, 8,250-255.
17. Driebeek, N., J. (1966). An Algorithm for the solution of Mixed Integer Programming Problents,
Management Science, 12, 576-587.
18. Duran. M.A. and Grossmann, I.E. (1986). An Outer-Approximation Algorithm for a Class of
Mixed-Integer Nonlinear Programs. Mathematical Programming 36,307-339.
 492

19. Faqir N.M and Karimi LA (1990). Design of Multipurpose Batch Plants with Multiple Production
Routes, Proceedings FOCAPD'89, Snowmass Village CO, pp 451-468
20. Fletcher R, Hall I.A. and lohns W.R. (1991). Flexible Retrofit Design of Multiproduct Batch
Plants, Comp & Chem. Eng. 15, 843-852
21. Floudas, C.A. and Visweswaran, V. (1990). A global optimization algorithm (GOP) for certain
classes of nonconvex NLPs-I Theory, Computers chem. Engng. 14, 1397-1417
22. Fourer, R., Gay, D.M. and Kernighan, B.W. (1990). A Modeling Language for Mathematical
Programming, Management Science, 36, 519-554.
23. Geoffrion, A.M. (1972). Generalized Benders Decomposition. Journal o/Optimization Theory and
Applications, 10(4), 237-260.
24. Geoffrion,A.M. and Graves, G. (1974). Multicommodity Distribution System Design by Benders
Decomposition, Management Science, 20, 822-844.
25. Glover F.(l975). Improved Linear Integer Programming Formulations of Nonlinear Integer
Problems, Management Science, Vol. 22, No.4, pp 455-460
26. Gomory, R. E. (1960). An Algorithm for the Mixed Integer Problem, RM-2597, The Rand
Corporation ..
27. Grossmann, LE. (1990). Mixed-Integer Nonlinear Programming Techniques for the Synthesis of
Engineering Systems, Research in Eng. Design, I, 205-228.
28. Grossmann LE ·and Sargent R.W.H, (1979) Optimum Design of Multipurpose Chemical Plants ,
Ind.Eng.Chem.Proc.Des.Dev. , Vol 18, No.2, pp 343-348
29. Grossmann LE, Voudouris V.T., Ghattas 0.(1992). Mixed-Integer Linear Programming
Reformulation for Some Nonlinear Discrete Design Optimization Problems, Recent Advances in
Global Optimization (eels. Floudas, C.A .. and Pardalos, P.M.) ,pp.478-512, Princeton University
Press
30. Gupta, J.N.D. (1976). Optimal Flowshop Schedules with no Intermediate Storage Space. Naval Res.
Logis. Q. 23, 235-243.
31. Gupta, O.K. and Ravindran, V. (1985). Branch and Bound Experiments in Convex Nonlinear Integer
Programming. Management Science, 31(12), 1533-1546.
32. Hooker, J. N. (1988). Resolution vs Cutting Plane solution of Inference Problems: some
computational experience. Operations Research Letters, 7,1(1988).
33. Jeroslow, R. G. and Lowe, J. K. (1984). Modelling with Integer Variables. Mathematical
Programming Study,22, 167-184.
34. Jeroslow, R. G. and Lowe, J. K. (1985). Experimental results on the New Techniques for Integer
Programming Formulations, Journal of the Operational Research Society, 36(5), 393-403.
35. Jeroslow, R. E. and Wang, J. (1990). Solving propositional satisfiability problems, Annals 0/
Mathematics and AI,I, 167-187.
36. Knopf F.C, Okos M.R, and Reklaitis G.V. (1982). Optimal Design of BatchlSemicontinuous
Processes,lnd.Eng.Chem.Proc.Des.Dev. , Vol 21, No. I, pp 79-86
37. Kocis, G.R. and Grossmann, LE. (1987). Relaxation Strategy for the Structural Optimization of
Process Flowsheets.lndustrial and Engineering Chemistry Research, 26(9),1869-1880.
38. Kocis, G.R. and Grossmann, I.E. (1989). Computational Experience with DICOPT Solving
MINLP Problems in Process Synthesis Engineering. Computers and Chem. Eng. 13, 307-315.
39. Kocis G.R., Grossmann LE. (1988) Global Optimization of Nonconvex MINLP Problems in
Process Synthesis,lnd.Engng.Chem.Res. 27, 1407-1421.
40. Kondili E, Pantelides C.C and Sargent R.W.H. (1993). A General Algorithm for Short-term
Scheduling of Batch Operations. I. MILP Formulation. Computers and Chem. Eng .. , 17, 211-228.
41. Krarup, 1. and BiIde, O. (1977). Plant Location, Set Covering and Economic Lot Size: An O(mn)
Algorithm for Structured Problems in L. Collatz et al. (eds), Optimierung bei graphentheoretischen
und ganzzahligen Problemen, Int. Series of Numerical Mathematics, 36, 155-180, Birkhauser Verilig,
Basel.
42. Ku, H. and Karimi, I. (1988) Scheduling in Serial Multiproduct Batch Processes with Finite
Intermediate Storage: A Mixed Integer Linear Program Formulation, Ind. Eng. Chem. Res. 27,
1840-1848.
43. Ku, H. and Karimi, I. (1991) An evaluation of simulated annealing for batch process scheduling,lnd.
Eng. Chem. Res. 30, 163-169.
44. Land, A. H., and Doig, A. 0.(1960). An Automatic method for solving Discrete Programming
Problems, Econometrica, 28, 497-520.
45. Lovacz, L: and Schrijver, A. (1989). Cones of Matrices and Set Functions and 0-1 Optimization,
Report BS-R8925, Centrum voor Wiskunde en Informatica.
46. Magnanti, T. L. and Wong, R. T. (1981). Acclerated Benders Decomposition: Algorithm
Enhancement and Model Selection Criteria, Operations Research, 29, 464-484.
 493

47. Martin, R.K. and Schrage, L. (1985). Subset Coefficient Reduction Cuts for 0-1 Mixed-Integer
Programming, Operations Research, 33,505-526.
48. Mawekwang, H. and Murtagh, B.A. (1986). Solving Nonlinear Integer Programs with Large Scale
Optimization Software. Annals oj Operations Research,S, 427-437.
49. Miller D.L and Pekny J.F. (1991). Exact solution oflarge asymmetric traveling salesman problems,
Science, 251, pp 754-761.
50. Nabar, S.V. and Schrage (1990). Modeling and Solving Nonlinear Integer Programming Problems.
Paper No. 22a, Annual AIChE Meeting, Chicago, IL.
51. Nembauser, G. L and Wolsey, L (1988). Integer and Cominatorial Optimization. Wiley, New York.
52. OSL Release 2 (1991) Guide and Reference, IBM, Kingston, NY.
53. Papageorgaki S. and Reklaitis GN (1990) Optimal Design of Multipurpose Batch plants-I.
Problem Formulation ,lnd.Eng.Chem.Res.,VoI29, No. 10, pp 2054-2062
54. Papageorgaki S. and Reldaitis G.V (1990). Optimal Design of Multipurpose Batch plants-2. A
Decomposition Solution Strategy, Ind.Eng.Chem.Res.,Vol 29, No. 10, pp 2062-2073
55. Papageorgald S. and Reklaitis G.V. (1990). Mixed Integer Programming Approaches to Batch
Chemical Process Design and Scheduling, ORSAffIMS Meeting, Philadelphia.
56.· Patel A.N., Mah R.S.H. and Karimi lA. (1991). Preliminary design of multiproduct noncontinuous
plants using simulted annealing, Comp &: Chem Eng. 15,451-470 .
57. Pekny J.F and Miller D.L. (1991). Exact solution of the No-Wait F1owshop Scheduling Problem
with a comparison to heuristic methods, Comp &: Chem. Eng., Vol 15, No II, pp741-748.
58. Petersen C.C.(1991). A Note on Transforming the Product of Variables to Linear Form in Linear
Programs, Working Paper, Purdue University.
59. Quesada I. and Grossmann I.E. (1992). An LPINLP based Branch anc Bound Algorithm for Convex
MINLP Problems. Compo & Chem Eng., 16, 937-947.
60. Quesada I. and Grossmann I.E. (1992). Global Optimization Algorithm for Fractional and Bilinear
Progams. Submitted for publication.
61. Rardin, R. L. and Choe, U.(1979). Tighter Relaxations of Fixed Charge Network Flow Problems,
Georgia Institute of Technology, Industrial and Systems Engineering Report Series, #J-79-18,
Atlanta.
62. Raman, R. and Grossmann,l. E. (1991). Relation between MILP modelling and Logical Inference
for Process Synthesis, Computers and Chemical Engineering, 15(2),73-84.
63. Raman, R. and Grossmann,I.E. (1992). Integration of Logic and Heuristic Knowledge in MINLP
Optimization for Process Synthesis, Computers and Chemical Engineering, 16(3), 155-171.
64. Raman, R. and Grossmann, I.E. (1993). Symbolic Integration of Logic in Mixed-Integer
Programming Techniques for Process Synthesis, to appear in Computers and Chemical Engineering.
65. Ravenmark D. and Rippin D.W.T. (1991). Structure and equipment for Multiproduct Batch
Production, Paper No.133a, Presented in AIChE annulal meeting, Los Angeles, CA
66. Reklaitis G.V (1990) Progress and Issues in Cumputer-Aided Balch Process Design, FOCAl'/)
Proceedings, Elsevier, NY, PI> 241-275
67. Reklaitis G.V. (1991). "Perspectives on Scheduling and Planning of Process Operations",
Proceedings Fourth Inl.Symp. on Proc. Systems Eng., Montebello, Quebec, Canada.
68. Rich S.H and Prokopakis GJ. (1986). Scheduling and Sequencing of Batch Operations in a
Multipurpose Plant, Ind. Eng. Chem. Res, Vol. 25, No.4, pp 979-988
69. Rich S.H and Prokopakis GJ. (1987). Multiple Routings and Reaction Paths in Project Scheduling,
Ind.Eng.Chem.Res, Vol. 26, No.9, pp 1940-1943
70. Sahinidis, N.V. and Grossmann, I.E. (\991). MINLP Model for Cyclic Multiproduct Scheduling on
Continuous Parallel Lines, Computers and Chem. Eng., IS, 85-103.
71. Sahinidis, N.V. and Grossmann, I.E. (1991). Reformulation of Multiperiod MILP Models for
Planning and Scheduling of Chemical Processes, Computers and Chem. Eng., IS, 255-272.
72. Sahinidis, N.V. and Grossmann, I.E. (1991). Convergence Properties of Generalized Benders
Decomposition, Computers and Chem. Eng., IS, 481-491.
73. Savelsbergh, M.W.P., Sigismandi, G.C. and Nemhauser, G.L. (1991) Functional Description of
MINTO, a Mixed INTeger Optimizer, Georgia Tech., Atlanta.
74. Schrage, L. (\986). Linear,lnteger and Quadratic Programming with LINDO, Scientific Press, Palo
Alto.
75. SCICONICNM 2.11 (\991). Users Guide", Scicon Ltd, U.K.
76. Shah N. and Pantelides C.C. , (1991). Optimal Long-Term Campaign Planning and Design of Batch
Operations, Ind. Eng. Chem. Res., Vol 30, No. 10, pp 2308-2321
77. Shah N., Pantelides C.C. and Sargent, R.W.H. (1993). A Geneml Algorithm for Short-term
Scheduling of IIlItch Operalions. II. Computational Issues. COIII/lllter.1 all/I Cllelll. £11/1 •. , 17, 229-
244.
 494

78. Sherali, H.D. and Alameddine, A. (1992) A new reformulation-linearization technique for bilinear
programming problems, Journal of Global Optimization, 2,379-410.
79. Sherali H. and Adams W.(1988) A hierarchy of relaxations between the continuous and convex hull
representations for zero-one programming problems, Technical Report, Virginia Polytechnic
Institute.t
80. Sherali, H. D. and Adams, W. P. (1989). Hierarchy of relaxations and convex hull characterizations
for mixed integer 0-1 programming problems. Technical Report, Virginia Polytechnic Institute.
8!' Sparrow R.E, Forder GJ, Rippin D.W.T (1975) The Choice of Equipment Sizes for Multiproduct
Batch Plant. Heuristic vs. Branch and Bound, Ind. Eng. Chem.Proc.Des.Dev. , Vol 14, No.3,
ppI97-203
82. Straub, D.A. and I.E. Grossmann (1992). Evaluation and Optimization of Stochastic Flexibility in
Multiproduct Batch Plants, Comp.Chem.Eng., 16, 69-87.
83. Suharni I. and Mab R.S.H. (1982) Optimal Design of MultipurPose Batch Plants, Ind. Eng. Chem.
Proc. Des. Dev., Vol 21, No. I. pp 94-100
84. Sugden, SJ. (1992). A Class of Direct Search Methods for Nonlinear Integer Programming. Ph.D.
thesis. Bond University, Queensland, Australia.
85. Swaney, R.E. (1990). Global solution of algebraic nonlinear programs. Paper No.22f, AIChE
Meeting, Chicago, IL
86. Tomlin. 1. A. (1971). An Improved Branch aod Bound method for Integer Programming, Operations
Research, 19, 1070-1075.
87. Tomlin, J. A. (1988). Special Ordered Sets and an Application to Gas Supply Operations Planning.
Mathematical Programming, 42,69-84.
88. Torres, F. E. (1991). Linearization of Mixed-Integer Products. Mathematical Programming,
49,427-428.
89. Van Roy, T. J., and Wolsey, L. A. (1987). Solving Mixed-Integer Programming Problems Using
Automatic Reformulation, Operations Research, 35, pp.45-57.
90. Vaselenak J.A ,Grossmann I.E. and Westerberg A.W. (1987). An Embedding Formulation for the
Optimal Scheduling and Design of Multipurpose Batch Plants, Ind.Eng.Chem.Res,26, Nol, pp139-
148
9!. Vaselenak J.A • Grossmann I.E. and Westerberg A.W (1987) Optimal Retrofit Design of
multipurpose Batch Plants, Ind.Eng.Chem.Res, 26, No.4, pp718-726
92. Viswanathan, J. and Grossmann. I.E. (1990). A Combined Penalty Function and Outer-
Approximation Method for MINLP Optimization. Computers and Chem. Eng. 14(7),769-782.
93. Voudouris V.T and Grossmann I.E. (1992). Mixed Integer Linear Programming Reformulations for
Batch Process Design with Discrete Equipment Sizes, Ind.Eng.Chem.Res., 31, pp.1314-1326.
94. Voudouris V.T and Grossmann I.E. (1992). MILP Model for the Scheduling and Design of
Multipurpose Batch Plants. In preparation.
95. Voudouris, V.T. and Grosmann, I.E. (1993). Optimal Synthesis of Multiproduct Batch Plants with
Cyclic Scheduling and Inventory Considerations. To appear in Ind.Eng.Chem.Res.
96. Wellons H.S and Reklaitis G.V. (1989). The Design of Multiproduct Batch Plants under
Uncertainty with Staged Expansion, Com. & Chem. Eng., 13, No1l2, pp115-126
97. Wellons M.C and Rekiaitis,G.V. (1991). Scheduling of Multipurpose Batch Chemical Plants. !.
Multiple Product Campaign Formation and Production Planning, Ind.Eng.Chem.Res, 30, No.4,
pp688-705
98. Williams, P. (1988). Model Building in Mathematical Programming. Wiley, Chichester.
99. Yuan, X., Piboleau, S., and Domenech, S. (1989). Une Methode d'Optimisation Non Linaire en
Variables Mixtes pour La Conception de Procedes. RAIRO Recherche Operationnele
Recent Developments in the Evaluation and
Optimization of Flexible Chemical Processes

Ignacio E. Grossmann and David A. Straub

Deparunent ofChemicaJ Engineering. Carnegie Mellon University. Pittsburgh. PA 15213. USA

Abstract: The evaluation and optimization of flexible chemical processes remains one of the
most challenging problems in Process Systems Engineering. In this paper an overview of
recent methods for quantifying the propelty of flexibility in chemical plants will be presented.
As will be shown, these methods are gradually evolving from deterministic worst-case
measures for feasible operation to stochastic measures that account for the distribution
functions of the uncertain parameters. Another trend is the simultaneous handling of discrete
and continuous uncertainties with the aim of developing measures for flexibility and reliability
that can be integrated within a common framework. It will be then shown how some of these
measures can be incorporated in the optimization of chemical processes. In particular, the
problem of optimization of flexibility for multiproduct batch plants will be discussed.

Keywords: flexibility, design under uncertainty, worst-case analysis, statistical design.

1. Introduction
The problem of accounting for uncertainty at the design stage is clearly a problem of great
practical significance due to the vari'ations that are commonly experienced in plant operation
(e.g. changes in demands, fluctuations of feed compositions and equipment failure).
Furthermore, at the design stage one must rely on values of technical parameters which are
unlikely to be realized once a design is actually implemented (e.g. transfer coefficients and
efficiencies). Finally, models that are used to predict the performance of a plant at the design
stage may not even match the correct behavior trends of the process. In view of all these
uncertainties, the common practice is to overdesign processes and/or perform ad-hoc case
studies to try to verify the t1exibility or robustness of a design. The pitfalls of such
approaches, however, are well known and therefore have motivated the study and development
of systematic techniques over the last 20 years ([4], [5]).

It is the purpose of this paper to provide an overview of recent techniques that have been
developed for evaluating and optimizing t1exibility in the face of uncertainties of continuous
parameters and discrete states. This paper is in fact an updated version of a recent paper
 496

presented by the authors at the COPE Meeting in Barcelona ([7]). In this paper we will
emphasize work that has been developed by our group at Carnegie Mellon. This paper will be
organized as follows. The problem statements for the evaluation and optimization problems
will be given first for deterministic and stochastic approaches. An overview will then be
presented for different formulations and solution methods for the evaluation problems,
followed by similar items for the optimization design problems. As will be shown, the reason
for the recent trend towards the stochastic approach is that it offers a more general framework,
especially for integrating continuous and discrete uncertainties which often arise in the design
of batch processes. At the same time, however, the stochastic approach also involves a
number of major challenges that sti1l need to be overcome, especially for the optimization
problems. A specific application to multiproduct batch plants will be presented to illustrate
how the problem structure can be exploited in specific instances to simplify the optimization.

2. Problem Statements
It will be assumed that the model of a process is described by equations and inequalities of
the form:

h(d,z,x,9)=0
g(d,z,x,9):S;0 (1)
d=Dy

where the variables are defined as follows:

d - L vector of design variables that defines the structure and equipment sizes of a process
z - nz vector of control variables that can be adjusted during plant operation
x - nx vector of state variables that describe the behavior of a process
9 - np vector of continuous uncertain parameters
y - L vector of boolean variables that describes the unavailability (0) or availability (1) of
the con-esponding design variables d
D - diagonal matrix whose elements in the diagonal correspond to the design variables d

For convenience in the presentation it will be assumed that the state variables x in (1) are
eliminated from the equations h(d,z,x,9)=O; the model then reduces to

f(d,z,9):S;O
d=Dy (2)
The evaluation problems that can then be considered for a fixed design D are as follows:
 497

A) Deterministic Problems Let y be tixed. and S be described by a nominal value SN,

expected deviations in the positive and negative directions LlS+, LlS-, and a set of
inequalities r(S) ~ 0 to represent correlations of the parameters S:

a) Problem AI: Determine if the design d = Dy, is feasible for every point S in
T={ SISN -Lle- :'> S :'>SN+M+. r(S):'>O}

b) Problem A2: Determine the maximum deviation 0 that design d = Dy can tolerate such
that every point S in T(o)={SISN -Me-:'> S :'>SN+MS+, r(S):'>O} is feasible.

Problem (AI) corresponds to the feasibility problem discussed in Halemane and

Grossmann [8], while problem (A2) corresponds to the flexibility index problem discussed in
Swaney and Grossmann [25).

B) Stochastic Problems Let S be described by a joint probability distribution function j(S):

a) Problem B I: If y is fixed, determine the probability of feasible operation.

b) Problem B2: If the discrete probability PI for the availability of each piece of
equipment t is given. detelmine the expected probability of feasible operation.

Problem (B I) corresponds to evaluating the stochastic flexibility discussed In

Pistikopoulos and Mazzuchi [17). while problem (B2) corresponds to evaluating the expected
stochasticj1exibifity discussed in Straub and Grossmann [22).

As for the design optimization problems they will involve the selection of the matrix D so
as to minimize cost and either a) satisfy the feasibility test (A 1), or b) maximize the flexibility
measure as given by (A2). (B I) or (B2), where the latter problem gives rise to a multiobjective
optimization problem.

3. Evaluation for Deterministic Case

3.1 Formulations
In order to address problem (Al) for determining the feasibility of a fixed design d,
consider first a fixed value of the continuous parameter e. The feasibility for fixed d at the
given e value is then given by the following optimization problem [8) :

'l' (d, S) = min

u.z u (3)
S.t. fj (d, z, S):,> U jE J
 498

where \jI ~ 0 indicates feasibility and \jI > 0 infeasibility. Note that the objective in problem (3)
is to find a point z* such that the maximum potential constraint violation is minimized.
In terms of the function \jI(d.a). feasibility for every point ae T. can be established by the
formulation [8J:

X (d) = max \jI(d.

BeT
e) (4)
where Xed) ~ 0 indicates feasibility of the design d for every 'point in the parameter set T. and
Xed) > 0 indicates infeasibility. Note that the max operator in (4) determines the point a* for
which the largest potential constraint violation can occur.

As for the flexibility index problem (A2). the fOlmulation is given by (see [25]):
F = max 0
s.t. max \jI (d.
geT(O)
0 e)::; (5)

02:0

where the objective is to inscribe the largest parameter set T(o*) in the feasible region projected
in a-space. An alternative formulation to problem (A2) is.

(6)

where
o·(~ = max 0
O.z
s.t. fj(d.z.e)::;o jeJ (7)

e = eN + 08
02:0

and T = (81- l'1e-::; 8::; l'1e+)

The objective in (6) is to find the maximum displacement which is possible along the
displacement 9 from the nominal value aN. Note that in both (5) and (6) the solution of the
critical point a* lies at the boundary of the feasible region projected in a-space.

3.2 Methods
Assume that no constraints r(e)::; 0 are present for correlating parameter variations. Then
the simplest methods for solving problems (AI) and (A2) are vertex enumeration schemes
which rely on the assumption that the critical points a* lie at the vertexes of the sets T and
T(S*). Such an assumption is only valid provided certain convexity conditions hold (see [25]).
 499

Let V = {k} correspond to the set of vertices in T={ele N -~e- ~ e ~eN-~e+}. Then,
problem (4) can be reformulated as

X(d) = max uk
keY (8)

where
uk = min
u.z
u
(9)
s.t. ~ (d, z, ek) :0; u jE J

That is, the problem reduces to solving the 2"p optimization problems in (9).

Likewise, problem (6) can be reformulated as

F = min I)k (10)
keY'

where
I)k= max I)
Ii.z
s.t. fj (d, z, e):o; 0 jE J (11)
e = eN + ~ek
I)~O

and ~ek is the displacement vector to ve11ex k. This problem again reduces to solving 2"p
optimization problems in (11).

The problem of avoiding the exhaustive enumeration of all vertices, which increase
exponentially with the number of parameters, has been addressed by Swaney and Grossmann
[26] and Kabatek and Swaney [10] using implicit enumeration techniques. The latter authors
has been able to solve problems with up .to 20 parameters with such an approach.

An alternative method that does not rely on the assumption that critical points correspond
to vertices, is the active set strategy of Grossmann and Floudas [3]. This method relies on the
fact that the feasible region projected into the space of d and e,

R(d,e)={ 8 I 'I'(d,8):O;O} (12)

(see Figure 1) can be expressed in terms of active sets of constraints fj(d,z,8)=0, jE J~

k=l,oo.NAS.
 500

Figure 1 Constraints in the space of d and 8

These active sets are obtained from all subsets of non-zero multipliers that satisfy the Kuhn-
Tucker conditions of problem (3)

L Ajk=l. L A~~;=O (13)

jEJ~ jEJ~

Pistikopoulos and Grossmann [15] have proposed a systematic enumeration procedure to

identify the NAS active sets of constraints. provided that the corresponding submatrices in (13)
are of full rank.
The projected parameter feasible region in (12) can then be expressed as
R(d.8)={8 l'I'k(d.8)::;O. k=l... .• NAS} (14)

where
'l'k(d. 8) =min u
S.t. fj (d. z. 8) = u 1
je ~
(15)

The above active set strategy by Grossmann and Floudas [3] does not require. however, the
a-priori identification of constraints 'l'k. This is accomplished by reformulating problem (4)
with the Kuhn-Tucker conditions of (3) embedded in it. and expressed in terms of 0-1
variables Wj for modelling the complementarity conditions.

For the case of problem (A 1). this leads to the mixed-integer optimization problem
 501

X(d) = max u
S.t. sj+fjd,z,e)=u jeJ
L Aj= 1
jeJ

~
~
A/azfj =0
jeJ

Aj - wr5 1 \
jeJ
sr U (1 - Wj) $; 0 f
L Wj$; nz+ 1
jeJ (16)
eN _ ~e' $; e $; eN + ~e+
r(e) $; 0
wj=O,I; Aj;Sj~O jeJ

where U is a valid upper bound for the violation of constraints. For the case of problem (A2),
the calculation of the flexibility index can be formulated as the mixed-integer optimization
problem (17). In both cases, constraints fj that are linear in z and e give rise to MILP
problems which can be solved with standard branch and bound methods. For nonlinear
constraints models (16) and (17) give rise to MINLP problems which can be solved with
Generalized Benders Decomposition [2] or with any of the variants of the outer-approximation
method (e.g. [28]). Also, for the case when nz + 1 constraints are assumed to be active and the
constraints are monotone in z, Grossmann and Floudas [3] decompose the MINLP into a
sequence of NLP optimization problems, each corresponding to an active set which is
identified a-priori from the stationary conditions of the Lagrangian.

F=min 0
S.t. sj+fj(d,z,e)=o jeJ
L Aj = I
jeJ

~ A" afj =0
~
jeJ
Jaz
Aj - Wj $; 1 \
jeJ
sr U (1 - Wj) $; 0 I
LWj$;nz+l
jeJ (17)
eN _ Me-$; e $; eN + Me+

r(e)$; 0
o~ 0 Wj = 0.1 , Aj, Sj ~ 0 je J
 502

4. Evaluation for Stochastic Case

4.1 Formulations
In order to formulate problem (B 1). the probability of feasible operation given a joint
distribution for e, j(e). this involves the evaluation of the multiple integral

SF(d) = ( j (8) d8 (18)

16:IjI(dP)'>o

where SF(d) is the stochastic flexibility for a given design (see [17], [22]). Note that this
integral must be evaluated over the feasible region projected in e space (see eqn. (12) and
Figure 2). In Figure 2 the circles represent the contours of the joint distribution function j.

Figure 2.SF is evaluated by integration over the shaded area.

For the case when uncertainties are also involved in the equipment, discrete states result
from all the combinations of the vector y. It is convenient to define for each state s the index
sets

(19)

to denote the identity of available and unavailable equipment. Note that state s is defined by a
particular choice of yS which in turn determines the design variables for that state, dS=Dys.
Also. denoting by Pt the probability that equipment I be available. the probability of each state
pes) is given by:

pes) = n n
1eY t
PI
1eYb
(l - P1) s=I, ... 2 L (20)

In this way the probability of feasible operation over both the discrete and continuous
uncertainties (i.e. problem (B2» is given by
2"
E(SF) = I SF(s) p(s) (21)
s=1

where E(SF) is the expected stochastic flexibility as proposed by Straub and Grossmann [22].
 503

4.2 Methods

The solution of problems (18) and (21) poses great computational challenges. Firstly,
because (18) involves a multiple integral over an implicitly defined domain. Secondly, (19)
involves the evaluation of these integrals for 2L states. For this reason solution methods for
these problems have been only reported for the case of linear constraints:

(22)

Pistikopoulos and Mazzuchi [l7] have proposed the computation of bounds for the
stochastic flexibility, SF(d) by assuming that j(a) is a normal distribution. Firstly, expressing
the feasibility function ",k(d,a) as given in (15) through the Lagrangian, this yields for (22) the
linear equation

Ijfk(d, oj =I, A r[c TO + <] (23)

IEJ A

where c;= c? + a Td. Since (23) is linear in 0 and these are normally distributed N(l1a,La),
then the distribution function <I>(ljfk) is also normally disu"ibuted with mean and variance,

I1ljfk = L
jEJ ~
r
A [c Tl1e + <J (24)

(25)

where 1:8 is the variance-covariance matrix for the parameters o.

The probability of feasible operation for the above set k is then given by the one-
dimensional integral

SFk = f <I> (ljfk) dljfk (26)

which can be readily evaluated. For multiple sets the probability is defined as follows (shown
for 2 sets),

(26b)
 504

where IPMVN is the multivariate nonnal probability distribution function.

Lower and upper bounds of the stochastic flexibility SF(d) are then given by
NAS
SF-(d) =L sF' - L SFk< + L SFHm .•. (27)
k=1 kd t<lqJl

SFU (d) = min {

q=i.Q
I1 SFk} (28)
keJA(q)

where JA(q) ~ JA. q=l .... Q. are all possible subsets o'f the inequalities ",k(d.9):50 •
k=l .... NAS. It should be noted that the bounds in (27) and (28) are often quite tight providing
good estimates for the stochastic flexibility.

Straub and Grossmann [22] have proposed a numerical approximation scheme for
arbitrary distribution functions using Gaussian quadrature within the feasible region of the
projected region R(d.9) (see Fig. 3).

Figure 3 Location of Quadrature Points.

The location of the quadrature periods is performed by first projecting the functions 'l'k(d.9).
k=l...NAS into successively lower dimensional spaces in 9; i.e. :[91.92 .... 9Ml -+
[91.9z....9M_Il ...-+[911 This is accomplished by analytically solving the problems r=1.2 .....M-l:

",r+I.. k(d. 91. 92.... 9M-r) = min u

(29)
S.t. ",r.k(d. 91. 92.... ):5 u k=l...NAS(r)

where ",I.k = ",k(d.9) = L fj (d.z.9). and NAS(r) is the number of active sets at the rth state
jeJ ~

of the projection.
 505

In the next step. lower and upper bounds are generated together with the quadrature points
for each 8j component in the order a]"-ta 2... -7a M. This is accomplished by using the analytical
expressions "'r.k(d.8 1.8 2•...• 8 M+r.I) in the order r=M. M-l. ...• to detennine the bounds.
For instance. the bounds 8 1L and 8 1U are determined from the linear inequalities
",M.k(d.8 1)$0. k=I •... NAS(M). The quadrature points 81qlthen are given by:

Vql (8 Y- 8 r) + 8 Y+ 8 r
8i' = 2 ql=l .... QPI (30)

where Vqlo qI=I •.... QP I represent the location of QPI quadrature points in [-1.1]. In the
next step. bounds for 82 are computed for each ail from ",M-I,k(d.al.a2)$0. k=l. ... NAS(M-

1). These bounds are denoted as a~(ail) since they depend on the value of ail. Quadrature

points are then computed as in (30) and the procedure continues until the the bounds

8 ki(8 i'.8i'~ .. 8~i~·'), 8 ~(8i',8i,q, ... 8~'1"') and quadrature points 8~q2 ...qM are detennined.

The numel1cal approximation to (18) is then given by

a U a L QPI [a ~(ai'-a ha (1)]CP2 [a ~(a i',a ~I'l!)

SF(d~ I WfJ2 I WfJ2~---
2 q=l 2 02=1 2
a hai',a ~I'l!] a ~(a i', .. a~~'M·') . a ~(ai' .... a~~'M-')
2'" 2

_I
QPM
xI WqM (a ~1.a~I'l!, ... a~'l!'M) (31)

where wq, are the weights corresponding to each quadrature point.

It should be noted that equiltion ClI) becomes computationally more expensive as the
number of parameters a increases which is not the case with the bounds in (27) and (28).
However. as pointed out before, ClI) can be applied to any distribution function (e.g. nonnal.
beta. log) while the bounds can only be applied to normal distribution. Also. both methods
require the identification of active sets which may become large if the number of constraints is
large.

As for the solution of equation (21) for the expected flexibility. Straub and Grossmann
[22] have developed a bounding scheme that requires the examination of relatively few states
despite the fact that these can become quite large in number. They represent the states through
a network as shown in Fig. 4.
 506

S1={1.2.3}

~
S2={1.2} S3={1.3} S4={2.3}

C><><J
S5={1} S6={2} S7={3}

~
S8={0}
Figure 4 State Network showing different possible sets of active units.

Here the top state has all the units active (i.e. YI = 1). while the bottom state has all units
inactive. Since the states with active units will usually have the higher probability. the
evaluation starts with the top state.

At any point of the search the following index sets are defined:

E={sISF(s) is evaluated}
U={sISF(s) is not evaluated} (32)

The lower and upper bounds are then given as follows:

E(SF)L = 2, SF(s) P(S)

seE
(33)
E(SF)U = 2, SF(s) P(S) + 2, BSF(s) P(S)
seE seU

where BSF(s) are valid upper hounds that are propagated through the subnetwork from higher
states that have been evaluated. Convergence with this scheme for a small tolerance is
normally achieved within 5 to 6 state evaluations (see Figure 5) provided the discrete
probabilities PI >0.5. The significance of this method is that it allows the evaluation of
flexibility and reliability within a single measure accounting for the interactions of the two.
 507

1.0

0.9
Upper Bound

u.
~ 0.8
w

0.7
Lower Bound

2 3 4 5 6 7
Number of States Evaluated

Figure 5 Example of the progression of the bounds.

5. Design Optimization
Most of the previous work ([9), [11)) has only considered the effect of the continuous
uncertain parameters e for the design optimization, and for which the minimization of the
expected value of the cost function has been considered using a two-stage strategy:

(34)

In order to handle infeasibilities in the inner minimization, one approach is to assign penalties
for the violation of constraints (e.g. C(d,z,S)=C if f(d,z,S) >0. This however can lead to
discontinuities. The other approach is to enforce feasibility for a specified flexibility index F
(e.g. [8)) through the parameter set T(F)={SISL -F.1S- $ S $Su+F.1S+, r(S)$O}. In this
case (34) is formulated as

min E [minC(d, z. S)lf(d, z, S):o::O]

d 8ET(F) z
(35)
S.t. max
BET(F)
'V(d, S)$ ()

A particular case of (35) is when only a discrete set of points Sk, k=1..K are specified which
then gives rise to the problem

K
min
d,zl, .. zk
Ii
k= 1
WkC (d, zk, Sk)
(36)
s.t. f(d. zk, Sk) $ () k=1..K
 508

K
where Wk are weights that are assigned to each point ak, and L Wk = 1.
k=l

Problem (36) can be interpreted as a multiperiod design problem which is an important

problem in its own right for the design of flexible chemical plants. However, as shown by
Halemane and Grossmann [8] this problem can also be used to approximate the solution of
(35). This is accomplished by selecting an initial set of points Ok, solving problem (36) and
verifying its feasibility over T(F) by solving problem (AI) as given by (4). If the, design is
feasible the procedure terminates. Otherwise the critical point from (4) is included to the set of
K 0 points and the solution of (36) is repeated. Computational experience has shown that
commonly one or two major iterations must be performed to achieve feasibility with this
method (e.g. see [3]).

While the above procedure can be applied to general linear and nonlinear problems, one
can exploit the structure for specialized cases. For instance, consider the case of constraints
that are linear in d, z, and 0, and where the objective function only involves the design variables
d. This case commonly arises in retrofit design problems.

As shown by Pistikopoulos and Grossmann [13], equation (23) holds for linear
constraints. Therefore, the constraint in (35) can be simplified into NAS inequalities as shown
in the following model:

min C(d)
d
s.t. £."
~ [c JTack + cJ
~ A J. 9 + aT d] < 0
.I- k=l..NAS
jEJ~ (37)

where

The significance of problem (37) is that the optimal design can be obtained through one single
optimization which however requires prior identification of the NAS active sets.
 509

Pistikopoulos and Grossmann [13) have presented an alternative formulation to (37) in

which one can easily derive the trade-off curve of cost versus the flexibility index. The
formulation is given by
min C (dE + ~d)
~d

s.t. Ok ~F \
L k=l, .. NAS (38)
Ok = 0 ~ + ~ cr f ~dl j
~du ::; ~d ::; ~dU , ok ~ 0

k·IS the fl eXI'b'l'

w here 0 E I Ity In , set k at th ease
'dex f'or actIve b d '
eSlgn dE an d crtkOOk O\jlk are
= --k""--d
d\jl 0 l
sensitivity coefficients that can be detelmined explicitly; ~d are design changes with respect to
the existing design dE,

Also, these authors extended the fOImulation in (37) to the case of nonlinear constraints,
Here, the inequalities in (37) are augmented within an iterative procedure similar to the scheme
based on the use of the multiperiod design problem, except that problem (15) is solved for each
active set to determine the critical points and mUltipliers,

Finally, the determination of the optimal degree of flexibility can be formulated for the case
of linear constraints as

max Z = E I max p(z, 8) If(d, z, 8) ::;0 ») - C(~d)

6ET(F) \ z

st, Ok ~F \
L k=l,..NAS
Ok =0 ~ + I cr ~~dt f (39)
l =1

d = dE + ~d , Ok ~ ()

where p(z,8) is a profit function,

Pistikopoulos and Grossmann [14) simplified this problem as maximizing the revenue
subject to minimizing the investment cost; that is (see Fig, 6):
 510

max Z =R(F) - C(F)

F
s.t. C(F) = min C (L1d)
Ok;:: F (40)
L
Ok = 0 ~ + I cr fL1d(
~ =1

and where
R(F) = E (max p (z, e)lf (d, z, e)::; 0 1
/
9ET z (41)
L1d = arg [CCF)]
which is solved by a modified Cartesian integration method. Since problem (40) is expressed
in terms of only the flexibility index F. its optimal value is found by a direct search method.

R(F)

~------------~~F

Figure 6 Curves for Detelmination of Optimal Flexibility

6. Application to Multiproduct Batch Design

The methods presented in the previous section have been only applied to continuous
processes. On the other hand batch processes offer also an interesting application since these
plants are built because of their flexibility for manufacturing several products. Reinhardt and
Rippin [19], [20] have reported a design method when demand uncertainties are described by
distribution functions. Wellons and Reklaitis [29] have developed a design method for staged
expansions for the same type of uncertainties. In this section we will summarize the recent
work by Straub and Grossmann [23J which accounts for uncertainties in the demands
(continuous parameter) and equipment failure (states). This will serve to illustrate some of the
concepts of Section 4 and show how the structure of the problem can be exploited to simplify
the calculations; particularly the optimization of the stochastic flexibility. Consider the model
for the design of multiproduct batch plants with single product campaigns (see [6]):
 511

(42)

Although problem (42) is nonlinear, for fixed design variables Vj (sizes), Nj (number of
parallel units), the feasible region can be desclibed by the linear inequality

NP
LQi'Yi:5:H (43)
i=1

where 'Yi = max {ti/Njl/min IV/Sid·

J J

If we define
NP
HA = L Qi'Yi (44)
i=1

then the problem of calculating the probability of feasible operation for uncertain demands Q,
i=l,N, can be expressed through the one-dimensional integral

(45)

which avoids the direct solution of the multiple integral in (18). Furthermore, the distribution
C/l(HA) can be easily determined if nOlmal distributions are assumed for the product demands
with mean J.lQ and variance cr~. Then proceeding in a similar way as in (24) and (25) the
mean and the variance of C/l(HA> are given by

NP
IlHA = L 'Yi IlQ
i=1
NP (46)
cr~A= L 'Yfcr&
i=1

with which the integral in (45) can be readily evaluated for the stochastic flexibility.
 512

As for the expected stochastic flexibility. let Pj be the probability that a unit in stage j is
available. Also let n j. j=l.M be the number of units that are available for any given state s.
Then it can be shown that the number of feasible states where at least some production can be
M

obtained is given by lFS = n

J=1
(N~. and that the probability of each state is given by

pes) = TI
M NJo!
j=1 (n j)! (Nrn j)!
n'( I-p l(N,·n,')
po'
J J
(47)

In this way the expected stochastic flexibility can be expressed as

1FS
E(SF) = L SF(s) pes) (48)
s=1

where SF(s) and pes) are given by (45) and (47). respectively. The value of E(SF) can then be
obtained by applying the bounding procedure at the end of Section 4 (eqtns. (32) and (33)).

In order to determine the sizes Vj and number of parallel units Nj that maximize the
stochastic flexibility (i.e. only uncertainties in the demands) given a limit for the capital
investment. C. one would have to optimize in plinciple the integral in (45) over the constraint
set in (42). However. this can be avoided in view of the fact that maximizing the normal
deviate .z = (H-IlHAYcrHA. is equivalent to maximizing the integral. Thus by applying
appropriate exponential transformations to (42) to convexify the problem. the optimal design
that maximizes the stochastic flexibility for a limit in the investment cost. can be formulated as
the MINLP (see [23]):

S.l. bi:5; Vj - 10g(Si j) } i~ ~ •.... NP

tu ~ log(tij) -Tlj .1-l... .. M
Tlj=~ Wj r IOg(r)} ._
.1-I. .... M
I Wjr= I
r

L. aj exp(llj +~j Vj) sC

j

~i=tLi-bi i= 1....• NP
I!HA=I exp(~ill!Qj
1 (49)
O~A=I exp(2~j)
1
obi
 513

In(Bb~bi ~In (Bi) i= I ..... NP

In (Vr):::; Vj:::; In (V~) j=l •.... M
In (Tb)~tl i ~In (TEi)
_00 $ ~i $; 00 i=I ..... NP
H. IlHA' OAA ~
In (Vr):::; Vj:::; In (V~) j=l.. ... M
wjr=0.1 j=I ..... M r=I ..... N}l

By solving this MINLP for different values of C one can then determine trade-off curves of the
expected stochastic flexibility versus cost (see Figure 7). Also. note that if the number of
parallel units is fixed (49) reduces to an NLP problem.

As for the optimization of expected flexibility. problem (49) can be extended as a

multiperiod design problem if Nj is fixed. where each period corresponds to a given state s.
There is no need to solve. however. for all the states since a preanalysis can easily establish the
relative magnitudes of (47) and valid upper bounds for SF(s). The optimization of Nj and Vj is
in principle considerably more complicated. However. here Straub and Grossmann [22] have
developed an enumeration procedure that relies on the state network representation and which
minimizes the number of multipeliod optimization problems that need to be examined. The
details of this method can be found in their paper.

1.0

0.8

0.6
u.
f/)

w 0.4

0.2

0.0
120000 145000 170000
Cost ($)

Figure 7 Trade-off Curve

7. General Nonlinear Models

The ideas developed in Straub and Grossmann [22] to evaluate the SF with linear models
have recently been extended to nonlinear models. see Straub and Grossmann [24] and Straub
 514

[21]. In addition the new concepts allow a straightforward extension to the design optimization
problem.
Observe that eqn (18) can be written as follows for a specified design d,
9P 9~J(61) 19U(61.tIJ....IlM-,)
SF= r r
L L .... L j(9) d9M .... d8 2 dSI
]91 ]92(91) 9M(61.t'2. ....IlM-I)
(50)
With nonlinear models in is difficult to analytically determine the constraints 'I' that lead to the
bounds on 8 in the linear case. Thus the bounds on each 8 need to be determined by a
nonlinear programming problem: max 81 U S.t. h=O, g:5:O. The difficulty here is that this may
potentially involve a large number of optimization problems. However, it has been shown that
a single NLP can be used to determine the bounds in place of all of the individual NLP's. The
benefit of the single NLP is that to extend the SF evaluation problem to that of the design
optimization problem one need only consider the design variables as decision variables instead
of parameters. The NLP for design optimization is shown below,

max SF--SP-S~ <{1

£.."
.J!,(S~lql-s~ql) <E
L w,'" J·(SqISql.q2)
I' 2
2 ql=1 2 q2=1
S.t. gt(d.Zl·).Xl·).S~ .Sr) )$0
gP(d.zl·).Xl·l,SP .Sr) )$0
S('=0.5 [s}' (1+V~I)+ sf- (I.vi ' ) J 41=1 •... QI

g2
Lql ql
(d.Z(·).Xl·).S) ,S2
Lql
)$0
Q
41= I.... )
U ql ql llql
g2 (d.zl·).Xl·).SI .S2 )$0 lIl=l •... Ql
si,q,=o.5 [S~' (I+v~')+ s~q, (I.vi') J ql=l .... QI 1I2=1 •... Q2

cost(d)$u
(51)
Sr-$S~$SP$S~AX
sF'$S~ql $S~ ql$Sr AX 1I1=1 •... QI
sF'$sj')$Sr AX
dED

Note that only the inequality conslraints are shown for simplicity in presentation. Also note
that the quadrature approximation to the SF , eqn. (31), is used as the objective function and
the constraints defining the location of the quadrature points are also included. Finally, this
NLP model incorporates explicitly the design variables d with which one can optimize the
stochastic flexibility subject to a cost constraint and in that way generate trade-off curves as the
one in Fig. 7.
The authors in [24] also address the computational difficulties of the program above.
They propose a solution method based on Generalized Benders Decomposition [2]. The
solution technique involves decomposing the large NLP into a series of smaller NLP's that are
 515

easier to solve. The techniques have effectively been applied to a reactor design problem and to
a small flowsheet problem to determine the optimal size of the various pieces of equipment.

8. Conclusions

This paper has given an overview of methods and formulations for evaluating and
optimizing flexibility in chemical processes. As has been shown. deterministic methods have
reached a stage of maturity whereby their wider application. like in process simulators. should
be technically feasible. although not necessarily trivial (e.g. computation of flexibility index
and multi period design problems). Stochastic methods on the other hand are in principle
computationally more expensive. except for few specific cases (e.g. bounds for linear models.
optimization of multiproduct batch plants). However. the major advantage with the stochastic
approach is that it offers the possibility of integrating flexibility and reliability under a common
measure as has been discussed in this paper. It is clear that major advances are required to
make computationally feasible the optimization with the stochastic approach.

Acknowledgments

The authors would like to acknowledge financial support from the Department of Energy under
Grant DOE: DE-FG-02-8SERI3396

References

1. F1oudas. C.A. and I.E. Grossmann: Symhesis of Flexible Heat Exchanger Networks with Uncertain
F10wrates and Temperatures. Compo Chern. Eng. 11.319 (1987).

2. Geoffrion. A.M.: Generalized Benders Decomposition. JOTA 10.237-260 (1972).

3. Grossmann, I.E. and C.A. F1oudas: Active Constraint Strategy for Flexibility Analysis in Chemical
Processes. Compo Chern. Eng. 11.675 (1987).

4. Grossmann, I.E .. K.P. Halemane and R.E. Swaney: Optimization Strategies for Flexible Chemical
Processes. Compo Chern. Eng. 7. 439 (1983).

5. Grossmann, I.E. and M. Morari: Operahility. Resiliency and Flexibility-Process Design Objectives for a
Changing World. Proc. 2nd Inl. Cour. Foundations Computer Aided Process Design (Westerberg and Chien
Eds.). CACHE. 937 (1984).

6. Grossmann, I.E. and R.W.H. Sargent: Optimum Design of Multipurpose Chemical Plants. Ind.Eng.Chem.
Process Design Dev. 18.343-348 (1978).

7. Grossmann, I.E. and D.A. Su-dub: Recent Developments in the Evaluation and Optimization of Flexible
Chemical Processes. Proceedings of COPE-91 (eds. Puigjaner. L. and A. Bpuna). Barcelona. Spain,49-59
(1991)

8. HaJemane. K.P. and Grossmann, I.E.: Optimal Process Design under Uncertainty. AIChE J. 29. 425.
(1983).

9. Johns. W.R .• G. Marketos and D.w.T. Rippin: The Optimal Design of Chemical Plant to Meet Time-
varying Demands in the Presence of Technical and Corrunercial Uncert.1inty. Design Congress 76. Fl.
(1976).
 516

10. Kabatek. U. and R.E. Swaney: Worst-Case Identification in Structured Process Systems. Compo Chern.
Eng. 16. 1063-1072 (1992).

11. Malik. R.K. and R.R. Hughes: Optimal Design of Flexible Chemical Processes. Compo Chern. Eng. 3.
473 (1979).

12. Marketos. G.: The Optimal Design of Chemical Plant Considering Uncertainty and Changing
Circumstances. Doctoral Dissertation No 5f1J7 ETH Zurich. (1975).

13. Pistikopoulos. E.N. and Grossmann. I.E.: Optimal Retrofit Design for Improving Process Flexibility in
Linear Systems. Compo Chern. Engng. 12.719 (1988).

14. Pistikopoulos. E.N. and Grossmann. I.E.: Stocha~tic Optimization of Flexibility in Retrofit Design of
Linear Systems. Compo Chern. Engng. 12. 1215 (1988).

15. Pistikopoulos. E.N. and 1.E.Grossmann: Optinlal Retrofit Design for Improving Process Flexibility in
Nonlinear Systems- J Fixed Degree of Flexibility. Compo Chern. Engng. 13. 1003 (1989).

16. Pistikopoulos. E.N. and Grossmann. I.E.: Optimal Retrofit Design for Improving Process Flexibility in
Nonlinear Systems- II Optimal Level of Flexibility. Compo Chern. Eng. 13. 1087 (1989).

17. Pistikopoulos. EN. and T.A. Mazzuchi. "A Novel Flexibility Analysis Approach for Processes with
Stochastic Parameters." Compo Chern. Eng .. Vol. 14. No.9. pp.991-101O (1990).

18. Pistikopoulos. E.N .• T.A. Mazzuchi and C.F.H. Van Rijn: Flexibility. Reliability. and Availability
Analysis of Manufacturing Processes: A Unified Approach. Compo App. in Chern. Eng. (eds. Bussemaker.
H. Th. and Jedema. p.o.) Elsevier. Amsterdam (1990).

19. Reinhart, HJ. and D.W.T. Rippin: The Design of Flexible Batch Chemical Plants. Annual AJChE
Meeting. Paper 50e. New Orleans (1986).

20. Reinhart. H.J. and D.W.T. Rippin: Design of Flexible Multi-Product Plants A New Procedure for Optimal
Equipment Sizing Under Uncertainty. Annual AIChE Meeting. Paper 92[' New York (1987).

21. Straub. D.A.: Evaluation and Optimization of Process Systems with Discrete and Continuous
Uncertainties. PhD Thesis. Camegie-Mellon University (1992).

22. Straub. D.A. and Grossmann. I.E.: II1legrated Statistical Melfic of Flexibility for Systems with Discrete
State and Continuous Parameter Uncertainties. Compo Chern. Eng.14. 967. (1990).

23. StraUb. D.A. and Grossmann. I.E.: Evaluation and Optimization of Flexibility in Multiproduct Batch
Plants. Compo Chem. Eng. 16.69-87 (1992).

24. Straub. D.A. and Grossmann. I.E.: Design Optimization of Stochastic Flexibility. Compo Chern. Eng .•
17.339-354 (1993).

25. Swaney. R.E. and Grossmann. I.E.: An Index for Operational Flexibility in Chemical Process Design. Part
1 Formulation and Theory. AIChE 1. 31.621 (1985).

26. Swaney. R.E. and Grossmann. I.E.: An Index for Operational Flexibility in Chemical Process Design. Part
2 Computational Algorithms. AIChE J. 31. 631 (1985).

27. Van Rijn. C.F.H.: A Systems Engineering Approach to Reliability. Availability and Mantainance.
Presented at the Conference on Foundations of Computer Aided Operations. FOCAPO. Salt Lake City
(1987).

28. Viswanathan. J. and Grossmann. I.E.: Combined Penalty Function and Outer-Approximation Method for
MINLP Optimization. Compo Chem. Eng. 14.769-782 (1990).

29. Wellons. H.S. and G.V. Reklaitis: TIle Design of Multiproduct Batch Plants under Uncertainty with Staged
Expansion. Compo Chem. Eng.B. 115. (1989).
Artificial Intelligence Techniques in Batch Process Systems
Engineering

Jack W. Ponton

Department of Chemical Engineering, University of Edinburgh King's Buildings, Edinburgh EH 9 3JL, Scotland

Abstract: The discipline of Process Systems Engineering has evolved largely in the area of
continuous chemical processing. However, systematic developments in the batch area have taken
place in a number of centres, and a brief overview is provided ofthe present state of batch PSE
An introduction to some relevant techniques from the field of artificial inte\1igence focusses
on their use in process systems. Most applications, once again, have been in the continuous, but
extrapolation to batch processing is discussed.
Fina11y, some speculative applications of ideas from other areas of manufacturing are outlined.

Keywords: Process systems engineering, batch processing, artificial inte11igence

1 Introduction and Overview

The author i~ uneasy about writing this paper. Firstly, he has little or no experience
of batch processing. Secondly, despite having coauthored a number of papers on the
theme of 'artificial intelligence in process engineering', he has never really used any of the
techniques of AI himself! The work for which he has received vicarious credit has been
performed by his colleagues. What follows is thus necessarily highly derivative, and the
reader who wishes to follow it up must refer to the bibliography.
In the course of writing the author has in fact managed to clarify some of his own
ideas about a discipline of batch process systems engineering. A key difference between
the systems engineering of continuous and of batch processe~ is that the former focusses
on design, and the latter on operation. The serious reader would do well to read an
excellent and wide ranging paper by George Stephanopoulos on AI in plant operations,
[30]. A review of AI in process engineering, which the reader should be warned was not
written for an audience of processes engineers, hut which contains a useful overview and
many references, particularly in the important area of safety, may be found in [1l].
 518

The outline of the paper is as follows:

• An overview of the process systems engineering of batch processes.

• Relevant artificial intelligence techniques and their application:

- Rule based systems

- Qualitative modelling
- Object oriented techniques

• Design environments and databases

• A bit of blue sky

2 The Process Systems Engineeering of Batch Processes

Most work in process systems engineering PSE has concentrated, explicitly or implicitly,
on techniques for continuous processes. Batch processing has had a small, but enthusiastic
following in the areas particularly of scheduling and in the application of techniques from
general manufacturing, see e.g. [29,26,13]. There have been occasional industrial papers,
[27], but it is only recently that a wider range of ideas from the 'mainstream of PSE have
been specifically directed to the batch processing field, e.g. [9, 34, 16, 28].
In this section, I shall look at some characteristics of batch processes and look for
needs and opportunities in the application of PSE techniques.

2.1 Batch versus Continuous

In discussing the choice between batch and continuos processes, Douglas, [8] identifies the
following characteristics for candidate batch processes:

1. Modest production rates of less than 1000 to 10,000 te/year.

2. Seasonal or limited demand for products.

3. Multiproduct operation.

4. 'Difficult' operations, e.g.

• solids or slurries,
• long readion times,
• fouling.
 519

Characteristic a. can probably be regarded as both an opportunity and a consequence

arising from]. and 2., the plant not being fully occupied at all times.
The items under 4. are indicative of the inability to design certain types of operation,
arising largely from a lack of fundamental understanding of these operations. To them
we could add. particularly in the context of biological processes:

• reactions liable to stop or deviate after extended operation, and

• processes sensitive to contamination.

Some properties of batch processes suggested by the above and other observations are
as follows.
Batch plants are in general 'less designed' than continuous ·plants. They tend to be
assembled from standard items of equipment, whose size is often not seen to be particularly
critical. This non criticality is partly because capital equipment charges are typically
a smaller proportion of production cost than for continuous plants due to higher cost
feedstocks and higher labour charges, see below. It is also the case that flexibility may be
introduced by adjusting batch times rather than by changing equipment sizes.
There is thus a. general shift from concentration on the design of the process to its
operation.
Flexibility in operation is often obtained at the expense of increased labour, resulting
from the choice of equipment. For example, batch filters are more flexible than continu-
ous, but involve higher labour costs. There is however an overall increase in the role of
the operator as compared with continuous processes. As well as the desire, or need, to
maintain flexibility, a poorer understanding of basic process principles, e.g. in the han-
dling of wet, sticky solids, or of complex biological reactions, regularly calls for operator
intervention.

2.2 PSE Tools and Benefits in Continuous Processes

For continuous process the following PSE tools and techniques have yielded benefits.

1. Steady state mass and energy balancing programs: better sensitivity analysis and
easier evaluation of alternative designs and operating policies.

2. Large Seale Optimisation: reduced design margins, better and cheaper plants and/or
increased throughput.

3. Thermodynamic analysis: more energy efficient plants.

4. Detailed modelling and dynamic simulation: better understanding and hence more
effective use of key operations, e.g. distillation and reactors.

5. Advanced control: better regulation of certain difficult operations and processes.

6. Methodology of process synthesis and design: better ways of carrying.out design

and developing alternative and new processes.
 520

7. Design and management data systems and databases: faster designs and faster
response to operational needs.

8. Analysis of safety critical systems: the 'Hazop' method and safer plants.

Most of these techniques have brought benefits to both design and operation. However,
the emphasis of most work has been towards design

2.3 Application to Batch Processing

With the exception of the steady state mass and energy balancing, or 'flowsheeting' pro-
gram, all these techniques have already been used effectively in batch processing, although
not necessarily in quite the same way, nor with the same benefits.

1. The steady state 'flowsheeting' program, as a flexible and general purpose design
aid, has no real counterpart in the unsteady state world of batch processing. 1
Discrete event simulators, e.g. [26), developed for use in a variety of other fields,
have never quite filled the same role. This suggests an opportunity, because the
flowsheeting program has arguably the most practically effective tool to have come
from PSE. It is possible that some development of the ideas behind the state-task
network, [16], might lead to an appropriate tool.

2. Optimisation is a powerful tool in batch processing. As noted earlier, operation,

particularly scheduling, tends to feature rather than design, although retrofit design,
[34, 9], is now assuming significant importance.

3. Although some work has been done on thermodynamic analysis of batch plants,
the incentives are less than for continuous processing, energy being usually a less
significant part of production cost. There are also fewer opportunities since there is
no convenient way of moving energy between the time phases of batch processing.

4. Our ability to model, for example solids and biological operations is less advanced
than for fluid and chemical systems. This is less because of mathematical intractabil-
ity than because the basic processes themselves are less well understood. We thus
see empirical or 'black box' models, such as neural nets, used where a fundamental
or mechanistic model is really required.
The importance of the operator in batch processing suggests that we need operator
as well as process models.

5. The control requirements of batch processes are rather different from continuous
processes. The emphasis shifts from regulation to servos, and most particularly to
sequencing. The global control problems of continuous processing are primarily to
do with structure, while those of batch plant are concerned with sequence. This
is not to say that the structure of control systems is unimportant, especially in
1 Flowsheeting programs can be, and indeed are, used to establish time averaged material balances in
batch processing. However, in the continuous domain they are nowadays used as more general design
tools. It is this aspect of their use which has no clear batch counterpart.
 521

multiproduct plants which may require major reconfiguration of the control system
for different products,_ However the structure of a batch processing operation tends
to be mapped into the time domain as well as, and to a large extent, instead of, the
spatial domain.
6. There is no methodology of batch process design which corresponds to any of the
hierarchical schemes, e.g. [8], for continuous processes. In particular there is no cor-
responding hierarchy for operation which the different emphasis of batch processing
suggests may be needed.
7. The use of advanced database techniques in the process industries is still at an early
stage of its development. Some applications to both design and to operation have
been reported. In the latter area there seem to be many opportunities for batch
processing.
8. The Hazop methodology and techniques of quantitative hazard assessment are ap-
plicable to both batch and continuous processes, and indeed outside the chemical
industries. They are already extensively applied to batch processes for both de-
sign and operation. However, Hazop, in particular, is a very demanding and time
consuming tool. Any way of making it easier or more effective would be of great
value.

2.4 Summary
In an attempt to draw some conclusions from these comments and comparisons, I suggest
the following as ideas on which to focus in examining PSE techniques, including those of
AI, for application to batch processes.

• Operation rather than Design

• Sequence instead of Structure
• Ill-defined rather than well understood
• Operators as well as Operations

Some specific areas of likely payoff seem to be:

• A batch process engineering 'flowsheeting' tool

• Techniques for operator modelling

• A hierarchical approach to design of batch plants and operating strategies

• Hazop improvements
• Advanced data systems

These are not exactly neglected areas, and some, e.g. the last two, are equally appli-
cable to continuous processes. As it happens, however, they are areas where ideas from
the AI community may be helpful.
 522

3 Some Relevant AI Techniques

AI has provided us with two things: programming languages and tools, and concepts.
These often a.ppear to be closely linked, in that particular concepts are associated with
tools, but they are in fact independent. AI concepts may be implemented in any language,
even FORTRAN, and AI languages may be used to write more general kinds of program.

3.1 Rule Based Programming

Rule based programming is conceptually little more than the FORTRAN i f construction:
if (condition ) action
This forms the basis of the most familiar AI product, the so-called 'expert' system.
AI languages include special rule based programming tools called expert system 'shells'.
A complete system may be written in such a shell, and consists of a set of rules and an
interpreter which applies some search strategy to see which rules match a given problem.
The final rule which 'fires' is the solution to the problem. Thus a system to analyse some
aspect of process safety might include rules such as the following:

l:if P<P_design then No_pressure_hazard

2:if (C<LEL and C>UEL) then No_flam-hazard
3:if (No_flam_hazard and No_pressure_hazard) then Safe else Unsafe

The system can also 'explain' its conclusions by reporting which rules fired. Such
explanations frequently leave much to be desired. The above set of rules might produce
the explanation:

Unsafe by Rule 3 because not No_flam_hazard by Rule 2 not C<LEL

... which is logical but scarcel.y intelligent!

Although apparently rather successful in some areas of engineering, rule based systems
as such have had limited success in most areas of process engineering, where few concepts
are sufficiently straightforward to be described adequately by reasonably sized sets of.
rules.
One area of success has been in the modelling of operator actions to replace or augment
the plant operator. It would appear that operators work with a set of rules and apply
them, informally and often inconsistently. By formalising the actions of the best process
operators, rule based systems have been able to improve plant operation in some cases
significantly. The same effect could have been obtained, without the use of a computer,
simply by writing down the rules as operating procedures and requiring that all operators
follow them. Automating the rules ensures that they are followed, and is of course much
faster.
An obviolls problem with a book of rules, however implemented, is that it may not
cover all eventualities. There is no 'understanding' of the process and there is thus no
means of extrapolating to new situations. In an attempt to overcome this limitation, rules
have been combined with various sorts of model.
 523

See, for example [24, 17, 18]

It would be contentious to claim that rules form the basis of most other AI techniques,
but to a large extent this is so. There are other ways of writing rules which make them
more tractable for use in different circumstances. Their properties, however, including
their limitations, remain the same.

3.1.1 Blackboard Systems

A blackboard system may be thought of as a communicating collection of separate rule

based systems, each working on a different aspect of the same problem. Such systems
have been succesful, [3,4] in handling problems beyond the scope of any single rule based
system. Difficulties do arise in the coordination of these individual 'experts', none of
which understands the basis on which the others are working. Some means of global
management is required. Unfortunately, there is no obvious way of doing this without
restoring the difficulty, a very large and therefore unmanageable rule base, which caused
the problem to be decomposed in the first place.

3.1.2 Planning

A possible candidate for the blackboard manager is often se~n to be the AI planner.
Extensive work has been carried out in this area by the AI community, and some has been
applied in the process systems area.
Particularly relevant to our interest is work on planning operating procedures, [21,22].
At the end of the day, however, the basis of these planners is once again the use of rules.
The way in which these rules operate may vary. In linear planning the rule interpreter
attempts to progress towards a given goal, backtracking when it fails and trying again
by a different route. Nonlinear planners construct partial solutions and when these fail,
return to a previous partial solution and attempt to modify it to prevent the observed
failure. The nonlinear planner thus works both backwards and forwards. The difference
lies in how efficiently the solution is found, and should not affect what the solution is. A
useful illustration of the operation of these strategies is given in [22].
The area of operating procedure synthesis, with particular reference to startup, and
thus highly relevant to batch operation, has been investigated with these and related
techniques, see also [10]. The extent to which procedures developed in this way have
actually been used is unclear.

3.2 Qualitative Simulation

Within the AI community there has been considerable interest in the ideas of qualitative
simulation, see e.g [14]. The reason for this is, presumably, scientific curiosity, since no
particular claims of utility appear to have been advanced.
Qualitative simulation models the behaviour ofa system using a range of non numerical
measures: HIGH, LOW, RISING, FALLING, etc. Corresponding models can either be
 524

steady state or dynamic, the latter being qualitative differential equations or confluence
equations.
Within the process engineering community there has been some interest in these
techniques, largely in the areas of safety, e.g. [32], and fault diagnosis, e.g. [19]. The
justification for this interest is twofold. Firstly, it may be possible to infer 'universal truths'
which do not depend on specific numerical values. In this context, it is interesting to note
that qualitative simulation has been long established in the process industries in the line-
by-line hazard and operability study, [15], which uses the concepts HIGH, LOW, NONE,
DIFFERENT, etc., applied to process parameters FLOW, PRESSURE, TEMPERATURE, etc.
in what is essentially a simulation procedure carried out by a group of people.
The technique thus clearly has possibilities. Another expectation, relevant particularly
to real time fault diagnosis, is that qualitative models may be easier to construct and to
solve than full qualltitative models. As far as construction is concerned, this is probably
true. It is probably of comparable difficulty to write down equations for both types of
model, but the qualitative model will not require numerical values for all its parameters,
a major saving. However, experience suggests that qualitative models are often more
timeconsuming, sometimes by orders of magnitude, than conventional models!
There appear to have been no applications of qualitative modelling in the batch pro-
cessing field. However, this may represent a missed opportunity, since discrete event
simulation, frequently used to model scheduling, is more 'qualitative' than continuous
dynamic modelling. Adding the concepts BEFORE, DURING a.nd AFTER would provide an
interesting framework for batch process qualitative simulation.
It is likely that qualitative models will be most useful when combined in some way
with conventional quantitative equation models. The two approaches are in a sense com-
plementary. Qualiative models embody the ideas of 'cause and effect', while equations
represent only relationships. It is also possible to refine ideas of qualitative representation
so that they provide almost as much information as numbers, by including qualitative
measures of first and second derivatives [6].

3.3 Object Orientated Programming Techniques and Concepts

One of the most powerful ideas to emanate from the AI and computer science community
is that of Object Oriented programming and the concepts which underly it. Conven-
tional programming languages, such as FORTRAN, provided representational facilities for
a limited number of ideas, initially only numbers, either integra.l or general, and simple
groupings of these into vectors and arrays. Later languages, such as Pascal and Algol'68,
added a few further concepts, character strings and logical variables, as well as more
general groupings of these entities into structures.
Object oriented programs allow the user to define anything as an 'object' and, within
the other constraints of the language, to create and manipulate these objects. Objects may
contain other objects, and by a controllable inhe7·itance mechanism, contained objects may
automatically possess properties of the object which contains them. This leads naturally
to a hierarchical representation which is highly appropriate to physical objects, Huch as
processes which are objects containing other objects, namely streams and units, the latter
possessing parts such as vessels and flanges.
 525

More important, the representation is also appropriate to process engineering proce-

dures, such as the design of a plant in the hierarchical manner, [8J. Unlike the structures
of conventional languages, objects are not just passive. They may contain active com-
puter code which carries out some activity when the object is referenced. For example, a
process unit may be redesigned whenever its specification is changed.
These ideas have been used in process engineering both for physical items and for
conceptual activities. For an extensive example of the first, and an excellent overview of
objected oriented programming, see [31J.
The representation of the design process itself, which in the context of batch processing
would certainly include the design of the operating procedure, is discussed below.

4 Design Environments 'and 'Intelligent' Databases

There are many similarities between the object oriented paradigm and developments in
database technology. Indeed, the coming generation of data bases will be object oriented
databases, [5J. The ability to represent complex objects is clearly crucial to any computer
based system for carrying out process design, whether of continuous or of batch processes.
More subtle is the need to represent the process of design: the designers intentions, the
history of activities which embody the explanation of why design decisions were taken,
the assumptions used by the designer and all the design tools which he used, and so forth.
Many of these are subtle and complex, and techniques for their representation, let alone
manipulation. remain to be developed.
The complexity of the artifacts to be designed, the process of design and the tools used
by the designer has pointed to the need for a unified design environment within which all
activities may be conducted. The environment must provide a means of representing all
the entities described above within what must be, at the very least, an extremely flexible
system.
An example of the use of AI object oriented techniques as a database to keep track
of designers' decisions and to check these for violation of either imposed constraints, e.g.
for safety, or for violation of the designer's original intentions, may be found in [33J. This
particular application is interesting to the batch process systems engineer in that it has
been applied to the design of a batch process, although it was conceived initially in the
context of continuous processing, [20J.
Ideas of modelling the process of design are described in [2J. From that particular
work has grown a prototype design environment, intended for use by a group of designers
working in a cooperative manner on different aspects of the same design. An illustration
of the environment in use is shown in figure 1.
The activities in the numbered windows are as follows.

1. This is the top, system manager, window from which the design was started. It
identifies the Project, the designer and the process, bloek4.

2. This is the design history and hierarchy. The main process is decomposed into a
reaction section reae-seel and a recycle separation section rsep-see1. For the
 ( " 01><1 .) ICIIIS ( ...,)
.,.. .?-=~so-.
ItSlP-SEO ~ s:lI'ftMllOM-SlC .,

(Mtta5c;:hlml ,, ) uESlGMlu I~ Bloat< "" ....llOlH(to

2 m1C-SCl _ _ RlIIC-SltZ

(Schem. '1) Il.DClC4

< -< J·t!~:~t~~~~t:~~;t~;~~i~:;::~!~-

_'_ 8
,.tc;,·Ct-mi'I' I.!.li!:..~
(Jl
I\)
Houn in II ..... Ir !1±?.. EE) 0>

I Numb" of lj",. ' j puComllOllolnt! _,_ EEl

j ...ctl,l.llI'lels: ~--
, U n . "
I ,"v.I141lllr 01 ~lrnm SUites: 1._ EEl
+
I OlltlmluUOfI Method !L'~ _ __
I PI,n ttUl in yn"" !!:.._ ~
I Stalt 'n.. r~I;a1Ion ordw: _'_ E!!:
[UUlHy C~t 1'unctlon: "";""a",",,,,_ __

4 ~ll ty coon o,nlan: :2l!.~!.t.:!.Il.:!~~

II ", -'- .,~ .:..

."
~.
,"
- . .
.• . .. , - ' : .. ..~.:. :~; - ..~ ..

Figure 1. Process Design Envirorunent

 527

latter, a range of alternative designs are being examined and are at different stages
of detail.

3. The Flowsheet Tool contains one of these alternatives which is at the BFD level of
detail. A particular separation task, for example from within this BFD, has been
dropped into CHips ...
4 .. ;. which has synthesised a flowsheet and designed the columns, displaying them
in this flowsheet window. The units are 'mouse sensitive' and can be selected and
examined or changed. If the user were not satisfied with the automatically generated
process it could be dropped into the Flowsheet Tool for manual editing.

5. This menu window was popped up by CHips in the course of its operation. The
options are 'greyed out' as they are not active at this point.

6. The PPDS server is available both to programs, for example, it would have been
called ill for the column design, or for direct access by dropping into it a stream
from a flowsheet, or elsewhere.

7. Finally, this is a epee manager window. It shows a process and a stream object
which can be pulled into compatible tools or windows.

In the course of developing the environment, the need for a high level representation
of process engineering concepts was reinforced. Seeing this as the 'top level' of a database
system not as yet available, an object oriented messaging system was developed. This is
intended to serve as a means of communication within an environment such as that shown
above, but using an object orientated database as the actual 'firmware' communication
mechanism [1).

5 Blue Sky: Robots and Railroads

Finally, here are two ideas quite different from anything above. but perhaps of importance
to batch process engineering.
It is easy to move fluids around between processing steps in either batch or continuous
processes. Solids when dry can be air conveyed, and when slurried are relatively easily
handled. There is a difficult gap between suspended solids, which is how they are usually
formed, and dry solids, as they are generally sold, in which the materials are wet and
sticky. The plate-and-frame batch filter remains the most common method of primary
solid-liquid separation. Unlike most processing operations it is highly labour intensive.
Labour costs are thus significant, but the need to have an operator present at all is the
major drawba.ck, for example in toxic or sterile conditions.
We have recently succeeded, [12] in developing a laser visioned robot operated filter
press. The vision system in particular turns out to be very useful in identifying lumps
of material where they are not wanted, for example in vessel emptying. Techniques from
conventional, non-chemical manufacturing have limited applicability, since these assume
the artifacts being handled have a fixed shape and size. A new methodology of robotics
seems to be required for dealing with amorphous products.
 528

An interesting example of lateral thinking is involved in a batch process, [25], which

instead of moving chemicals between processing units, moves the processing units, on rail
tracks! Initial reaction to this proposal, when photographs of a plant in use in Japan were
shown, was concern about safety. However, the idea is intriguing, and raises the question
as to whether there are any other ways of radically rethinking a the philosophy of chemical
production. While not wishing to dampen enthusiasm for such radical developments, it
is worth noting that, for fluids, pipes do have many fundamental advantages over tank
cars. However, in the area of wet solids processing referred to above, the issue may well
not be so clear cut.

References
I. G Ballinger et al, Epee: A Process Engineering Environment, Proceedings of ESCAPE 3, Graz, Austria, 1993
2. R Banares-Alcantara, Representing the Engineering Design Process: Two Hypotheses, prize winning paper at AI
in Design'9l, proceedings, ed J Gero, published by Butterworth Heinemann, 199\. Also in Computer-Aided
Design, 595-603, 23, 91991
3. R Banares-Alcantara, EI Ko, AW Westerberg and MD Rychener, DECADE: a Hybrid Expert System for Catalyst
Selection: Part I, Expert System Considerations, Comp and ChemEng 11,267-277,1987
4 ... Part II- Final Architecture and Resul\S, ibid, 12,923-938, 1988
5. MR Blaha, J Mehta and RL Motard, Structure and Methodology in Engineering Information Management, AIChE
Fall Meeting, 137e, Los Angeles, 1991
6. IT -Y Cheung and G Stephanopoulos, Representation of Process Trends: Parts I and II, Comp and Chern Eng 14,
495-510 and 511-540,1990
7 BJ Cott and S Machietto, An Integrated Approach to Computer Aided Operation of Batch Chemical Plants, Comp
and Chem. Eng. 413, II112, 1263-1273, 1989
8. 1M Douglas, The Conceptual Design of Chemical Processes, McGraw-Hili, 1988
9. R Fletcher, JAJ Hall and WR Johns, Flexible Retrofit Design of Multiproduct Batch Plants. Comp and Chem., Eng
15,12,843,852,1991
10. RH Fusillo and GJ Powers, Operating Procedure Synthesis using Local Models and Distributed Goals, ibid, 12,
1023-1034,1988
I\, D Hutton, JW Ponton and A Waters, AI Applications in the Process Industries, The Knowledge Engineering
Review, 5, 2 69-95,1990
12. A Jaffrey, N Macleod and JW Ponton, The development of a laser guided robotic discharge system for filter presses,
Proc IChernE Research Event, 143,564-566, 1992
13. LS Karimi and GV Reklaitis, Intermediate Storage in Noncontinuous Processing, in 'Foundations of Computer
Aided Process Design', eds. Westerberg and Chien, CACHE publications, 1984
14. J de Kleer and JS Brown, A Qualitative Physics Based on Confluences, Artificial Intelligence, 24, 7-83, 1984
15. T Kletz, Bazap and Hazan, IChemE, London, 1986
16. E Kondili, CC Pantelides and RWH Sargent, A General Algorithm for Scheduling Batch Operations, Proc PSE'88,
62-75, Sydney, 1988
17. MA Kramer, Malfunction Diagnosis Using Quantitative models and non-Boolean Reasoning in Expert Systems,
A1CHE J 33, 130, 1987
18. MA Kramer, Expert Systems for Fault Diagnosis: a General Framework;, in Foundations of Computer Aided
Process Operations', eds Reklaitis and Spriggs, CACHE, Austin, Texas, 1987
19. MA Kramer and BL Palowitch, A Rule based Approach to Fault Diagnosis Using the Signed Directed Graph,
AICHEJ,33,1067-1078,1987
20. EM Lakie, Evaluation of a Tool for the Representation and Checking of Design Constraints by its Continuous Use
in a Student Process Design Project, Project Report, Department of Chemical Engineering, University of Edinburgh,
June 1989
21. A Lakshrnanan and G Stephanopoulos, Synthesis of Operating Procedures for Complete Chemical Plants. I:
Hierarchical Structured Modeling for Nonlinear Planning, Comp and Chem. Eng. 12, 985-1002, 1988
22 ... II: a Nonlinear Planning Methodology, ibid 1003-1021
 529

23. J Love and B Shiel and N Drakos, A Declarative Approach to Applications Software for Batch Process Control,
proceedings oflChemE Research Event 680-682, IChemE Rugby, 1992
24. DR Myers, JF Davis and OJ Hennan, A Task Oriented Approach to Knowledge Based Systems for Process
Engineering Design, Comp and Chern. Eng. 12, 959-972, 1988
25. T Niwa, Transferable Vessel Type Multipurpose Batch Plant, Proc PSE'91, IV, 2.1-2.15,1991
26. BW Overturf, G. V. Reklaitis, 1M Woods, GASP-IV and the Simulation of Batch and Semicontinuous Operations:
Single Train Processes, lEC Proc Des Dev, 17, 161, 1978
27. ML Preston and DH Cheny, Batch Process Scheduling - ICI's BatchMASTER, in "Process Systems Engineering:
PSE'85", Institution of Chemical Engineers, Rugby, 1985
28. DWT Rippin, Simulation of Single and A Multiproduct Batch Chemical Plants for Optimal Design and Operation,
Comp and Chern. Eng. 7, 3,137-156,1983
29. RE Sparrow, GJ Forder and DWT Rippin, The Choice of Equipment Sizes for Multiproduct Batch Plants:
Heuristics vs Branch and Bound, IEC Proc Des Dev 14,3, 197, 1975
30. G Stephanopoulos, The Scope of Artificial Intelligence in Plant- Wide Operations, in 'Foundations of Computer
Aided Process Operations', 505-555, eds Reklaitis and Spriggs, CACHE, Austin, Texas, 1987
31. G Stephanopoulos et ai, Design Kit: an Object Orientated Environment for Process Engineering, Comp and
Chern,. Eng. 6311,6,655-674, 1987
32. A Waters and JW Ponton, Qualitative Simulation and Fault Propagation in Process Plants, ChERD 67, 407-422,
1989
33. A Waters and JW Ponton, Managing Constraints in Design: Using an AI Toolkit as a DBMS, Computers and
Chemical Engineering, 16, 10111, 987-1006, 1992
34. JA Vaselnak, IE, Grossman and AW Westerberg, Optimal Retrofit Design in Multiproduct Batch Plants, lEC Res
26, 718-726, 1987
Elements of Knowledge Based Systems
Representation and Inference

Kristian M. Lien

Department of Chemical Engineering, The University of Trondheim, Norway

Abstract: Over the last decade, the use of Knowledge Based techniques have become popular in
Process Systems Engineering. This paper will give an overview of some of the main issues
involved in the application of such techniques. It does not emphasize batch systems in particular,
but does rather present a collection of useful general concepts in search, representation, inference
and systems architectures, applicable as well in batch process systems as in e.g. synthesis of
process flowsheet. After all, there is quite a number of similarities between batch process systems
engineering and process synthesis, since both may be viewed as configuration problems: In batch
processing, a major task is to configure a processing plan configuration of a set of actions in time.
Search, representation formalisms, inference mechanisms and problem solving systems
architectures are the main ingredients of this paper, where the latter of these three is particularly
focused towards blackboard systems. The last part of the paper describes the development of a
particular blackboard system - AKORN D, its successor AKORN Dr and an application system
for synthesis of separation systems built on top of AKORN Dr.

Keywords: Knowledge Based Systems, Search, Representation, Inference

Search

Search is a fundamental concept in computer problem solving. From an initial situation - the
initial state- it is desired to get to another situation- the gool state. A set of operators can be
applied to states. States are transformed into new states when operators are applied to them;
operators facilitate state transitions. In order to determine whether the search should continue or
 531

be stopped when a new state is reached, a termination criterion is applied to every new state
generated. If the criterion is satisfied- for example if the new state is a goal state - the search
terminates. Otherwise, the search continues.
State +-- initial state;
Until State satisfies Termination-Criterion do
begin
Select an operator Op from the set of operators applicable to State;
State +-- r~sult of applying Op to State;
end
A sequence of applied operators resulting in a goal state constitute a solution to the search
problem - a solution path. The set of possible states that can be reached is termed the problem's
search space.
There are numerous ways the search among the possible states can be performed. The two
major kinds of search - search regimes - that are distinguished are termed irrevocable and
tentative search regimes. In an irrevocable search regime, operators are selected and applied
without provision for reconsideration later. There may be alternative operators applicable to a
particular state, but one is chosen, and the state it is applied to may never be returned to later.
The pseudo algorithm above is of this type. Since states are never reconsidered in irrevocable
search regimes, these do not need to save descriptions of former states. In tentative search
regimes, previous states may be returned to so that alternative sequences of operators may be
applied. This implies that states must be saved so that they may be returned to at some point
later in the search process.
When may an irrevocable search regime be used, and when is a tentative regime to be
preferred? That depends on the nature of the search space: If all sequences of operators lead to
goal states - no partial sequence ends in a blind aIley - and any sequence is as good as any other,
then irrevocable regimes may be good enough. Application of hill climbing techniques, e.g.
'steepest ascent' methods, to determine local maxima of mathematical functions is an example in
case. However, if the global maximum is desired and the function of concern has more than one
local maximum, the hill climbing procedure may easily get stuck on one of these, failing to find
the global maximum solution.
Tentative search regimes may be classified as backtracking or graph search regimes.
Backtracking search regimes insist on exploring every possible path from the most recently
 532

generated state before tracking back to previously generated states to explore alternative paths
from there. Backtracking regimes need therefore only to store the states on the current path.
Graph search regimes do not put such an emphasis on finding extensions to the current path.
Several partial paths may coexist, and one may switch among these.

Evaluation of Solutions and Paths

There must of course be a rationale behind switching from one partial path to another; a belief
that one partial path is better than another in some sense. In order to determine the goodness of a
path, some goodness criterion must be present. A path from the start to a goal state may often be
assigned a value or cost, but for the goodness criterion to be useful in guiding the search process-
deciding which of the existing partial paths to extend - it must be applicable to partial solutions
as well as to completed ones.
The value or cost of a partial path from start to an already generated state S is usually the
easiest one to calculate - either, if the cost or value is additive, by summation of the cost of each
operator application on the path- or, if the cost is not additive, by an evaluation of the cost or
value associated with the state S. The problem lies in the estimation of what the cost or value of
extending the path from S to a goal state is going to be, since that part of the path is yet
unknown; it remains to be generated.
To illustrate the problem, suppose that the problem is to find a path to travel from New York
City (NYC) to Los Angeles (LA). Being in New York City there is the choice between going to
Boston and going to Chicago. ("Go to Boston" and "Go to Chicago" are the operators applicable
to the state "Being in NYC") If it is not known where Boston or Chicago is relative to LA, the
goal state, one may as well go to Boston first, because that is the closer of the two, less
expensive than going to Chicago. Being in Boston, there is the choice between going to Montreal
and going to Cleveland. Not knowing where Cleveland or Montreal is relative to LA, one may as
well go to Montreal first, because that is closer than Cleveland. All the way, the least expensive
next move is selected, minimizing local cost, but the overall result may be getting nowhere near
the goal state (LA) until all other possibilities are explored. If it is possible to estimate remaining
distances from the next possible location to the goal, efficient routes may be found with much
less effort: Go to Chicago first, because the distance
 533

NYC ~ Chicago + Chicago ~ LA

is less than
NYC ~ Boston + Boston ~ LA.
n may be shown that if the following three conditions are satisfied:
a) The goodness criterion is additive:
Cost of going from A to B through S
Cost of going from A to S
+ Cost of going from S to B
b) The cost ofthe remaining part of any partial path is never overestimated, that is, the real
cost will never tum out to be lower than the estimated cost.
c) The path extensions are always selected so that the path which has the lowest combined
cost (Cost of already generated part ofthe path + Estimated cost of remaining distance to
the goal) is expanded next.
then the first complete path to the goal state will comprise an optimal path: No other path
leading to the goal state generated subsequently can have a lower cost than the path already
found.
The A >I< search algorithm is an example of an optimal search algorithm. A description of A >I<
and an informal proof of its optimality may be found in Nillson[48]. If it cannot be guaranteed
that the estimated cost is always an underestimate, it cannot, in general, be guaranteed that an
optimal path has been found until all alternatives have been explored.
Suppose that there are two alternative estimation functions for remaining paths, FJ and F2.
Neither overestimate remaining cost, but F J consistently estimates higher costs than F2. Then, it
may be shown that a larger number of states will be generated with search procedures using F2
than with search procedures using Fl. It may also be shown that any state generated using FJ
will also be generated using F2. This does not necessarily mean, however, that a search
procedure using F J will be more efficient than one using F2, because efficiency also will have to
include how long it takes to evaluate the estimation function on a state or partial path. But, the
search tree generated using F J will be less "bushy" than the search tree generated using F2.
The best conceivable estimation function is one that estimates remaining costs exactly. If such
a function can be devised, every state generated in the search will be on the optimal path.
The above statements about optimality and heuristic power (heuristic power is a term
reflecting the "bushiness" of the search tree generated: the less bushy the tree, the more
 534

heuristically powerful the underlying search procedure.) hold only if the underlying assumptions
hold; that the goodness function is additive, that estimated costs never exceed real costs, and that
the path with lowest estimated total cost is always extended next.
The above description assumes that there is one goal state or that any goal state is as good as
any other. But often the problem is not to find an optimal path to a goal but to find a path to an
optimal goal. (Referring to the traveling example, the problem is not to find the optimal path to
LA but rather to find the city on the west coast with the best opportunities for getting into the
movie business. At first it seems unimportant how long it takes to get there, because being there
is what counts. In reality that is not entirely true. You may spend so much time looking for the
place that when you finally get there you are too old to become an actor.) Optimal goal problems
can often be addressed by the same means as optimal path problems. This is particularly true for
configuration or assembly problems: The goal state is described in terms of functionality, the
function that the configuration must be able to perform. The search space is described in terms of
parts that mayor may not be part of a solution configuration. The problem is to select a subset of
those parts and combine them so that the combination of parts performs the specified function
better than any other combination of parts present in the search space with respect to some
goodness criterion.
It is of course implicit in this formulation that optimal solutions are optimal in a context: the
present search space and the present goodness function. Another goodness criterion or search
space might have termed a totally different solution optimal.
An underlying assumption here is that it is possible to decompose the specified goal
functionality into subfunctionalities that can be implemented by devices present in the search
space that can be combined to a device that performs the goal function. The naive approach to
solving the problem would be to try every possible configuration of the search space elements,
evaluate the goodness of each and pick the best. This would be naive because only very small
search spaces could be handled that way. To illustrate, assume that the search space contains M
parts. The optimal solution to the problem could contain one of those parts, two of them, three ..
or at most all M. Assuming that parts can only be configured serially, one part after the other, a
solution comprising N parts can be configured in
 535

(~) * N! =-(M-~-~-)-!

M
ways from the M parts in the search space: There are ( N ) ways to pick N different parts from

a set ofM, and each ofthese can be configured in N ! ways if the configurations are restricted to
linear strings of parts. The total number of string-like configurations to consider (of which only a
small fraction will constitute valid solutions, most likely, but that cannot be known before they
are analyzed to check validity) will thus be

MI 1
"M . -MI*,\,M_- 2 *MI
L...i=l (M - i)! - . L..;=l i! = .

For M = 5, this amounts to approximately 250 possible combinations, which is not an

unreasonably large number, but already when M = 10 the number of possibilities approaches one
million ! And, even this is a gross underestimation, since only string -like configurations are
allowed.
This does of course not mean that all these alternatives are going to be explored, the example
is included only to demonstrate how hopelessly large search problems become in the absence of
more specific knowledge about the problem. Problem domain specific decompositions,
classifications and constraint logics enable the configurer to avoid taking the naive approach:
The goal functionality can be decomposed into subfunctionalities. A few of the search space
elements can be classified as being able to perform a given subfunction (that is why they were
included in the search space in the first place; to be alternatives for performing a subfunction that
likely would have to be performed given the goal functionality) so that the search space can be
viewed as partitioned into functionality groups. And within and between the functionality groups
choices are often constrained by domain considerations: Inclusion of one element into a partial
solution makes it unnecessary to consider certain other elements or groups. For example, if a
particular pump has been selected to increase pressure and the pump has sufficient capacity to
increase the pressure enough, it would not make any sense to select another pump to increase the
pressure since that functionality already has been implemented. Nor would it make any sense to
 536

include a pressure decreasing device - say, a valve or expander - into the partial solution
immediately before or after the selected pump.
An optimal search approach may be taken towards the configuration problem if it is possible
to estimate the goodness of a partial solution by calculation of the cost of parts already included
in the solution plus an estimated cost of remaining parts to include that is not an overestimate.
Unfortunately, it is often hard to find good underestimator. However, it may still be possible
to use a modified approach: Partition the cost of the parts into two classes, a class for which the
optimal search assumption holds and a "residue". Then, use optimal search on the part of the
costs for which the assumption holds.

Example
The dominant costs associated with a chemical plant are investment costs and energy costs. The
investment costs can be considered to be additive, but energy cost is a function of how well
energy integrated the plant is. If the energy consumption of each piece of equipment is summed,
one may get an overestimate of the energy consumption of the plant. Then, if the investment cost
is the dominant cost factor, ignore the energy cost and use optimal search on the investment part.
Alternatively, if it is possible to associate an energy cost with a piece of equipment that does not
reflect the entire energy consumption of the unit, but is close to what its final share in the total
energy cost is going to be, include that part of the cost into the modified optimal search
procedure.
The first modification has the consequence that the search procedure loses some of its
heuristic power as a result of the systematic underestimation of real costs. This will result in a
more 'bushy' search tree, but the procedure will finally terminate with a solution that is optimal
with respect to investment since no overestimation takes place. If investment costs are much
larger than energy costs, then this solution will also be close to the optimum for combined
investment and energy costs.
The second modification may lead to overestimation of costs. Harris [24] has shown that if the
maximum possible overestimation k of an estimation function can be determined, it can be
guaranteed that no solution will be generated costing more than k more than the optimal solution
before the optimal solution is generated. It may thus be tempting to try to use an overestimating
function, keep track of all estimated costs, compare the estimated costs with the real cost when
the first solution is found and claim that the solution found is no more than a more expensive
 537

than the optimal solution, where a is the maximum overestimation of all the cost estimates on the
way to the found solution. However, this does not work unless it can be demonstrated that the
function overestimates maximally on the way to that particular solution, that is, a = k for that
particular solution.
The basic problem in finding the optimal solution in general is that the cost of the optimal
solution is unknown initially. So in principle one may risk an exhaustive search before the optimal
solution can be recognized. In practice, this is often avoidable: Sometimes one may predict an
attainable lower bound on the cost of an optimal solution, sometimes one may find predictive
cost functions whose maximum possible overestimation can be determined, and sometimes both
these sources of information may be exploited:
a) If a lower bound L on the optimal solution can be predicted, then no partial solutions
with (partial) cost above L need to be explored.
b) If a predictive cost function has k as its maximum over-estimation, only the partial
solutions with the lowest estimated total cost and those partial solution with an estimated
total cost within k above the lowest one need to be explored.
c) If an attainable lower bound L on the cost of an optimal solution can be predicted and a
predictive cost function exists with a maximum overestimate k, no partial solutions need
to be explored that has an estimated total cost above L + k.
Case c) stated above allows the expansion of partial solutions up to the point where the cost of
already generated parts plus estimated remaining cost exceed L + k . Case a) allows the
expansion of partial solutions up to the point where the cost of the already generated part of the
solutions exceed L. When estimated remaining costs exceed k, unfruitful partial solutions will
thus be kept longer with a) than with c); c) is a stronger exclusion criterion than a), even though
it first may seem to be the other way around.
Optimal solutions may be hard to find. Often a major search effort is required to find the one
best solution, particularly if many solutions cost little more than the optimal one. However, when
cost differences are minute, finding the best solution may be more of an academic exercise than
an issue of real importance. Often solutions 'reasonably close' to optimum will be accepted. In
most cases, the search is performed within a search space that only approximates the set of
alternatives available in reality. Furthermore, the criteria used to evaluate solutions are frequently
simplified versions of real world considerations. Issues like human factors, politics, company
policies, etc., are rarely easily quantifiable.
 538

The context in which the search is perfonned is important. Even though optimal search
procedures may be devised within a given search space, it is hardly a too strong claim that the
goodness of sets of alternatives to explore and evaluation criteria to use strongly depend on
familiarity with characteristics of the problem. Unless search space and rating function are
carefully selected, the optimal solution may be no more than an excellent answer to the wrong
question.

Representation

In the previous section of this paper, problem solving was viewed as search within a space of
possibilities. Each possible situation on the path from initial problem to solution was tenned a
state. Operators could change one situation or state into another. This chapter addresses
representational issues; how to describe states and how to describe operators so that computers
can transfonn initial problem state descriptions into solution state descriptions.
In computer problem solving, states may be viewed as data and operators as procedures. The
operator descriptions can be given procedural interpretations, the interpretation of a state
description relies entirely on the correspondence between the symbols comprising it and the
external world the symbols represent.
A state may be described in terms of the objects present in the state, their properties and the
relations existing among the present objects. A collection of symbols can be said to be a valid
representation of a state if every distinguishable object, property and relation of concern is
represented by a unique symbol in the state representation. A collection of symbols can be said to
be a valid representation of an operator or state transition between two states S1 and S2 iff the
collection of symbols has a procedural interpretation generating behavior capable of symbol
manipulation characterized by
a) leaving every symbol corresponding to objects, properties and relations existing both in
S1 and in S2 unchanged,
b) destroying every symbol corresponding to objects, properties and relations present in S1
and absent in S2, and
c) creating new symbols uniquely corresponding to any object, property and relation of
concern absent in S1 and present in S2.
 539

Representations of valid solutions may be mechanically produced by application of operators

to state representations when such symbol structures representing states and state transitions can
be devised. However, note that the meaning is in the transformations: The interpretation of such
a collection of symbols relies entirely on the procedures manipulating it.
There is more than one approach to representing application of a set of operators to a set of
states: One might explicitly code the operators in some procedural computer language as a set of
functions that take a state description as input, where every such function explicitly is called to
check whether it is applicable to the current state description, for example representation as a
global block of data. If it is, execute the function and let its result be the new current state. Iterate
explicitly until no function is applicable, or an explicit exit from the loop is encountered.
Alternatively, one could imagine the operator application cycle partitioned into
a) a background processing mechanism that 'invisibly' selects 'operators proper' for
execution and handles the details of state description management and update, and
b) a higher level description of the 'operators proper' that is interpreted and executed by the
background mechanism.

Rule Based Programming

Rule-based programming systems incorporate such a partition. Frame-based or object oriented

systems are also examples of similar partitions of representation and processing into layers, as are
many process simulators, equation solvers, network communication protocols and robot
programming languages, to name a few. A rule-based program consists of three parts:
1) A global collection of symbols serving as state representation. This collection of data is
also frequently termed the system's 'dynamic database' or 'working memory'.
2) A collection of conditional statements or rules serving as operators representation. A rule
is on the form:
If <antecedent statements>
Then <consequent statements>
3) A background process or interpreter. The interpreter, which is invisible in the explicit
code comprising the rule based program, iteratively
a) finds the rules that can be applied to the current state
 540

b) selects a rule to apply among these

c) applies the rule.
Application of a rule may change the current state description, enabling some rules not applicable
in a given iteration cycle n to be applied in iteration n + 1.
Rule-based systems are often classified as being either fOlWard chaining or backward
chaining. A forward chaining system compares the antecedent statements with the current state
description, and if antecedent statements and state description statements are compatible, given
some compatibility criterion, the rule is applicable and may be selected. If it gets selected, the
consequent statements are processed. The processing of the consequent statements change the
state description. In other words, the forward chaining system produces symbol structures that
should be understood as consequences of applying a sequence of rules to an initial state
description.
A backward chaining system focuses on a particular statement and tries to establish a path
backwards through rule applications to statements present in the state description. It tries to
prove that this particular statement will be a consequence of application of a sequence of rules
from the current state. The statement of concern, frequently termed the goal statement, is
compared to consequent statements of rules. All rules that have consequent statements
compatible with the goal statement are considered to be on the path of rule applications from the
initial state to states containing the goal statement. One of these rules is selected and now the
antecedent statements of this rule become new statements of concern, new goal statements -
subgoals. This is illustrated below:
Database ~ initial state description;
Goal ~ statement of concern;
RuJes ~ the set of rules;
Result ~ Backchain (Goal, Database, Rules).
Function Backchain (Goal, Database, RuJes);
ResuJt ~ Success;
If Goal is not matched by any statement in Database, then
If a rule R exists in Rules with a consequent statement matching Goal, then
For every antecedent statement A in R OR until Result = Fail do
Result ~ Backchain (A, Database, Rules);
 541

else Result ~ Fail;

end if;
Backchain ~ Result;
The function Backchain as described above employs an irrevocable search regime for
identification of a rule sequence that will generate the initial goal statement as a consequence of
the statements in the initial state description. This is done for the sake of simplicity in the
presentation; a tentative search regime may of course be used, so that when a rule R is selected
from Rules that is a blind alley, alternative rules may be chosen until one succeeds or all
alternatives are explored.
It should be emphasized at this point that rules can serve two distinct purposes as
representational elements. Rules as representations of state transitions has already been described.
However, rules may also serve as a representational mechanism for explication of implicit
statements of state descriptions. Consider the statements:
C1 is a Big Compressor.
C1 is Expensive Equipment.
C2 is a Big Compressor.
C2 is Expensive Equipment.
Using rules a more compact description is possible:
C1 is a Big Compressor.
C2 is a Big Compressor.
If X is a Big Compressor, then X is Expensive Equipment.
Clearly, this rule does not represent any state transition. The world the statements describe does
not change depending on whether the rule is applied or not. The only thing that changes is the
amount of explicit information about that same world. Making decision (state transitions) and
deducing logical implication are totally different activities and should be treated accordingly in
any rule-based system:
Logic programming = logical implications.
Production systems programming = state transitions.
Logic programming is a rule-based approach to programming principally concerned with
deducing logical implications of state descriptions. Production systems programming is another
rule-based approach to programming oriented more towards decision making. These two
 542

approaches will now be presented and exemplified by a logic programming language; PROLOG,
and a production. system language; OPS.

Logic Programming and Mechanical Theorem Proving

In the early days of Artificial Intelligence, there was a dominant view focusing on formal logic as
the foundation for computerization of human intelligence. Mechanical derivation of theorems and
proofs from a set of initial axioms by means of one or a few formal rules of inference has
therefore had significant attention in this research area. The foundation of much of this work has
been the Predicate Calculus, a formal representation of classical logic.
To represent a natural language (here English) sentence in the predicate calculus, focus on the
relations and entities the sentence describes. There are many possible ways to represent an
English sentence in terms of predicate calculus formulae:
"The topstream output from column 1 is the feed to column 2"
may, for example, be represented as:
Same-stream (topstream (Coil), feedstream (Co12»
or as:
Same-stream (stream (Coil, Top), stream (Co12, Feed»
or as two connected formulae:
Top-stream (Streaml, Coil) n Feed-stream (Streami, Co12)
etc.
Atomicformulae are the building blocks of the predicate calculus. Connectives, such as 'n'
(logical'and~, 'u' (logical'o~ and '~' (logical implication), are used to combine these into more
complex formulae. '-,' (logical negation) is also often termed a connective, even though it is really
not used to combine formulae, but to negate them. Formulae connected by 'n' (and) are termed
conjunctions, formulae connected by'u' (or) are termed disjunctions, and formulae prefixed by
'-,' (not) are termed negations. In addition, we do also have implications, e.g.:
"IF columnl is a distillation column
and it doesn't have a partial condenser
Then it has a total condenser"
which may be represented as:
 543

Distillation-column (Coil)
("\-,Condensertype (Coil, Partial)
~Condensertype (ColI, Total)
The part of the last fonnula preceding the '~' is tenned its antecedent. It is composed of
antecedent statements. The part of the fonnula succeeding the '~' is tenned its consequent. It is
composed of consequent statements. It should be noted that implications may be expressed in
another equivalent fonn. "A implies B" is e.g. logically equivalent to "Not A or B".
When variables occur in a predicate calculus fonnula, there are alternative ways to interpret
the fonnula: It may be true for whatever object one may imagine in the domain of discourse
(universal quantification), or it may be true for at least one such object. (existential
quantification). The symbol 'v' denotes universal quantification, and the symbol '3' denotes
existential quantification. The meaning of
V (x) [Distillation-column (x)
~ Condenser-type X, Partial)]
("All distillation columns have partial condensers")
is obviously different from the meaning of
3 (x) [Distillation-column (x)
~ Condenser-type (x, Partial)]
(" Some distillation columns have partial condensers").

Resolution-A Widely Used Role ofInference

In the predicate calculus, rules of inference may be applied to fonnulae to produce new
formulae mechanically. Resolution is a widely used rule of inference on predicate calculus
fonnulae for mechanical reasoning purposes [51]. Resolution may only be applied to a certain
class of fonnulae called clauses, defined as fonnulae consisting of disjunctions of atomic (or
negated atomic) fonnulae only. But this is no real restriction, since any fonnula in the predicate
calculus may be converted to a set of clauses. To explain how is a fairly lengthy process, so for
brevity the reader is referred to NIlsson [48].
Resolution may be described as the combination of a pair ofparent clauses into a new clause
termed the resolvent of the parent clauses. The resolvent is computed by taking the disjunction of
 544

the two parent clauses, eliminating any complementary pair. (A complementary pair is the
disjunction of a fonnula and the same fonnula negated, for example P u ,P)
Example Parent clauses:
P
-,PuQ
Resolvent:
Pu[-,PuQJ
== [pu-,P] u Q
==Q
Since P u -,P is a complementary pair (with truth value always true, independent ofP's truth
value), eliminate the pair. It should be noted that the expression -,P u Q is logically equivalent
to the expression P ~ Q.
Atomic fonnulae can make up complementary pairs even when they contain variables. For this
to be true, the variables in the two fonnulae must match, their bindings must be made to be the
same.

Example
(A is a constant symbol denoting some named object, X, yand z are variable symbols.)
Parent clauses:
P(x, A, y)
,P(A, x, w) u Q(x, w)
Resolvent:
Q(A, z)
where x = A, w = Z, y = z
The process of finding variable substitutions to make expressions identical is tenned
unification. Ntlsson gives a description of unification [48]. Resolution refutation systems are a
particularly widespread type of resolution system. The basic difference between a resolution
system and a resolution refutation system is that instead of inferring the fonnulae inferable from
the initial set of fonnulae (the axiom fonnulae), one attempts to demonstrate that from the
negation of a particular fonnula - the hypothesis formula - combined with the set of axiom
 545

formulae, logical contradiction will be inferred by means of resolution, thereby demonstrating

that the negated hypothesis is inconsistent with the set of axiom formulae.

The Programming Language PROLOG

PROLOG is a logic programming language of increasing popularity [9]. The language is a

resolution refutation system restricted to a particular class of clauses - Hom clauses. While
general clauses are arbitrary disjunctions of atomic predicate calculus formulae, Hom clauses are
restricted to having at most one unnegated atomic formula per clause. This does not restrict the
expressive power of Hom clauses, however, since any unrestricted clause can be equivalently
represented by a set of Hom clauses. An example of a Prolog statement follows:
Simple-column (col):-
Nr-of-feeds(col, I)
Nr-of-topstreams (col, 1 ),
Nr-of-bottomstreams (col, 1) .
This corresponds to the Hom clause:
Simple-column (col) \J -, Nr-of-feeds(col, 1)
\J -, Nr-of-topstreams (co~ 1) \J -, Nr-of-bottomstreams(col, 1)
which may be phrased in English in several ways:
1) Either col is a simple column
or it does not have one feed
or it does not have one topstream
or it does not have one bottomstream.
2) If col has one feed, one topstream and one bottomstream
then it is a simple column.
3) In order to prove that col is a simple column,
It must be demonstrated that:
col has one feed,
col has one topstream, and that
col has one bottomstream.
The last paraphrase is the one closest resembling the way Prolog works.
 546

Resolution refutation systems may in principle consider every possible pair of clauses in the
role as parent clauses in each resolution step. If there are m initial clauses, and n resolution steps
are performed, there will be m + n clauses at the end of the process, and the total number of pairs
of clauses considered as parent clauses during the entire resolution process will be:

since each of the m + i clauses present after i resolution steps at least in principle can be paired
with any of the other m + i -1 clauses present.
In order to reduce the number of pairs of clauses to consider, Prolog uses a variant ofa search
strategy called linear input resolution. Using a linear input resolution strategy, the hypothesis to
be proved is initially resolved with one of the initial axioms resulting in a new clause, the
resolvent. In the next resolution step, this particular resolvent clause is resolved with another
present clause, and so on. This implies that in every resolution step one of the parent clauses is
already selected, only the other parent clause needs to be determined. The number of possible
pairs to be considered is consequently reduced to

L~=o(m +i -1)= mn +n(n -1)/ 2

The use of such a strategy imposes a strictly hypothesis-oriented, goal-driven nature on the
behavior of the search process: There are no means of inferring consequences of the axioms
beside those explicitly asked for. The variant of linear input resolution Prolog uses is further
restricted in two ways.
1) The search proceeds in a depth first manner: If there are several clauses {S} in a given
resolution step i that can be resolved with the resolvent from the previous step one of
these will be attempted to produce a new resolvent. Only after every possible search path
from the new resolvent has failed to demonstrate logical contradiction, another clause
from {S} will be attempted. This is handled automatically by the chronological
backtracking search mechanism built into the Prolog interpreter.
2) Alternative clauses to be resolved with the resolvent from the previous resolution step
will be attempted in the order they occur textually in the Prolog program. The ordering
 547

of clauses in a Prolog program may therefore have a substantial effect on the time it takes
to find a proof
A Prolog program consists of a set of conditional or unconditional axioms and an hypothesis
to be shown to be a logical consequence ofthe axioms. The schematic example below is without
variables for brevity. Prolog's termination criterion is the production of the empty clause- a clause
with no formulae in it.

Prolog form Hom Clause form

1) A A
2) B. B
3) C:-A,B. C u-.[AnS] '" Cu-.Au-,B
4) D:-C. Du-,C
1) and 2) are unconditional axioms stating that A and B are true. 3) and 4) are conditional
axioms. 3) states that C is true if A is true andB is true. 4) states that D is true ifC is true.
Suppose now that we want to demonstrate the truth of D; the logical consequence of the
statements 1) through 4): Add the negation ofD to the set of clauses and attempt to demonstrate
that this is inconsistent with the axioms present:
5) :-D.
Once this is done, the resolution refutation process starts searching for the empty clause, denoted
':-' in Prolog form, 'NIL' in Hom clause form. (The 'or' denotes the logical disjunctive
combination of two Prolog form Hom clauses, and the '=>' is followed by the resulting resolvent,
the new clause with complementary pairs removed): First, Combine 5) and 4):
:-D. or D:-C. --,D u [D u -,C]
=>6) :-C. -,C
Then, combine 6) and 3):
:-C. or C:-A, B. -,[C u [C u -,[AnB]] '" [-,C u C] u -,[AnB]
=> 7) :-A,B. -,[AnB]
Thereafter, combine 7) and 1):
:-A,B. or A -,[AnB] uA", [-,Au A] u-,B
=> 8) :-B. -,B
Finally, combine 8) and 2):
 548

:-B. orB. -.B uB

=>9) :- NIL
The empty clause is produced after four resolution steps: -J) is shown to be inconsistent with the
database 1) through 4) of clauses. D may thus be said to be a consequence of the initial
statements present in the program.

Production Systems

Production system programming is another widespread approach to rule based programming.

This approach differs substantially from logic programming approaches:
The information processing is focused towards responding to the data present in the current
database rather than towards proving that a current hypothesis may be deduced from these
data.
There is no formal logic underlying the manipulation of data representations; the style of
problem solving is more associative than deductive.
While a logic program typically involves one rule 'calling' another in order to prove
sub-hypotheses of the current hypothesis, there is no explicit rule interaction in production
systems; rules do not 'call' one another, but should be viewed as independent, interacting
through the current data in the database only.
Allen Newell [39] has explained the concept that later turned into the realization of production
systems as follows:
"Metaphorically we can think of a set ofworkers, all looking at the same blackboard: Each
is able to read everything that is on it, and to judge when he has something worthwhile to
add to it. This concept is .... a set ofdemons, each independently looking at the total
situation and shrieking in proportions to what they see fit their natures [39]."
Newell was concerned with the structure of the existing problem solving programs when he
wrote this. The programs existing at that time were mostly organized along a generate-and-test
search model, and the primary difficulties with this organization were inflexible control of the
search and restricted data accessibility. The blackboard solution cited above that Newell
proposed was later termed production system. In a production system, each 'worker' or 'demon'
is represented as a condition-action rule, and data are globally accessible in a working memory.
 549

One of the many 'shrieking demons' (those which have their conditions satisfied) is selected
through a conflict resolution process emulating the selection of the loudest shrieking demon.
As an illustration of the production system concept, consider the following informal simple
example.

Initial working memory

1) current temperature is 300 K
2) boiling point ofn-C6 is 342 K
3) boiling point ofn-C7 is 371 K
Rules
a) If <x> is liquid, and <y> is liquid,
Then consider distillation to separate <x> from <y>.
b) If <x> is vapor, and <y> is vapor,
Then consider adsorption to separate <x> from <y>.
c) If <x> has its boiling point above the current temperature,
Then <x> is liquid.
d) If <x> has its boiling point below the current temperature,
Then <x> is vapor.
Given the above working memory and set of rules the following steps may take place:
Step 1: Rule (c) matches working memory elements (1) and (2).
Insert '4) n-C6 is liquid' into working memory.
Step 2: Rule (c) matches working memory elements (I) and (3).
Insert '5) n-C7 is liquid' into working memory.
Step 3: Rule (a) matches working memory elements (4) and (5).
Insert '6) consider distillation to separate n-C6 and n-C7'
into working memory.
After step 3, working memory includes statements (1) through (6).
The execution of a production system is an iterative process consisting of three steps: Match,
Select and Execute. In the Match step, every rule is inspected (its condition part is) to see
whether the rule is applicable. The set of applicable rules from this step is termed 'the conflict set'.
The Select step decides which of the rules in the conflict set to apply next. Obviously this choice
makes a difference in the behavior of the system, and a wide variety of mechanisms have been
 550

provided in different systems. Such rule selection mechanisms are termed 'conflict resolution
strategies'. In the Execute step, the selected rule is executed - or more precisely - the statements
comprising its action part are. Rule action statements typically specifY patterns of change to take
place in the working memory, changes that will enable rules not present in the current conflict
set to be present in the conflict set in the next cycle. This way a forward chain of rule
applications take place.
In the following section, the main features of OPS5, the currently most widely used
production system programming language, will be described.

OPS5, A Production System Language

OPS5 programs consist of working memory elements (data) and productions (rules). A working
memory element is, to some extent, similar to a Record in Pascal. Working memory elements
(wmes) may be of different types (belong to different wme classes), where each type has a
number of predefined attributes (record fields) where attribute values may be stored. For
example the sentence "Coil is a distillation column with partial condenser" may be represented
inOPS5 as:
(distillation-column tname colI tcondenser-type partial)
Wmes cannot (as Pascal Records can) be explicitly linked with pointers to form complex data
structures. In this respect, wmes are more like entries in a relational database table: Links are
present only implicitly, through shared attribute values.

Example
"Coil is a distillation column with condenser condI, whose surface area is A." may be
represented as:
(distillation-column t name colI t condenser condI)
(condenser t name cond I t surface-area A)
The link between the two wmes is implicit in that the former and latter have the same
symbol-condI-referred to as value of one of their attributes. The symbol NIL is the default value
for all attributes in OPS5 working memory elements.
 551

Each wme has associated with it a time-tag. This is a unique identifier, any wme may be
identified by its timetag. Whenever a wme is modified, it gets a new time-tag. OPS5 productions
operate on the database of wmes. It should be noted that the entire set of wmes and productions
is memory-resident. A production's If-part - termed its left hand side (Ihs) - may be viewed as a
query, using database terminology, into the database of wmes. This query is in the form of a
pattern, and it is matched against the contents of the wme database. The Then-part - termed the
right hand side (rhs) of the production - suggests patterns of change to be performed on subsets
ofwmes that match the left hand side query.

Example
"If the distillation column colI has a condenser condl and cor.til is not explicitly associated with
any column then associate condl with colI"
may be represented as the production:
(p create-inverse-link-from-condl-to-coll
distillation-column tname coIl tcondenser condl)
condenser tname condl tcolumn NIL)
~

(modifY 2 tcolumn coil»

The above production (where 2 is a cursor to the 2nd wme referenced in the production's Ihs)
is not particularly useful, it contains no variables in its Ihs, so the pattern matching involved here
is a simple equality test. A more useful and general production is possible if variables are
substituted for constant symbols:
(p create-inverse-link-from-a-condenser-to-the-column-it-belongs-to
(distillation-column tname <col> tcondenser <cond»
(condenser t name <cond> t column NIL)
~

(modifY 2 tcolumn <col»)

OPS5 productions may also match against missing data - verifY the absence of a piece of data
in a particular data pattern. It should be noted that this is not the same as the logical negation of
the data pattern.
 552

Example
"If there is a distillation column C associated with no particular condenser, and there is no
condenser which is associated with this particular column,
Then create a new unique symbol X to denote the name of the condenser, associate X with C
and C with X. "
(p make-new-condenser-wme-and-associate-bidirectionally-with-its-column
(distillation-column tname <c> tcondenserNIL)
(condenser t column <C»
~

(bind <x»
(make condenser t name <x> t column <C»
(modifY 1 t condenser <x»)
As any production system, OPS5 works in cycles, each involving the three steps Match,
Select and Execute. Step one, Match, identifies all production instantiations matching the current
wme database: Multiple sets of data may match a production consistently (a variable must be
consistently matched throughout the entire production; all its occurrences must denote the same
constant symbol throughout the production), and these sets must be distinguished among. The
pair of such a set of wmes and the identity of the production (its name) comprise a production
instantiation. The set of production instantiations comprises the current cycle's conflict set.
In conflict resolution - the Select step following the conflict set identification step- one
production instantiation is selected as being dominating according to OPS5's conflict resolution
strategy. The main principles in OPS's strategy are Refraction, Recency and Specificity.
Refraction implies that if a production instantiation has been selected before, it cannot be
selected again. This is to avoid infinite iterative execution of the same production instantiation.
OPS5 identifies instantiations by means ofthe wmes' time-tags, so if a wme gets a new time-tag,
it and all production instantiations it is part of are considered to be new, and may thus be
allowed to execute again. Recency specifies that if a particular wme referred to by a production
instantiation Z is more recently created or modified than any other wme referred to by any other
production instantiation in the conflict set, then Z will dominate in conflict resolution. If several
instantiations have the same recency rating- they refer to the same set of wmes - then the one
 553

associated with the production with the largest number of condition elements will dominate. This
is termed the Specificity principle in OPSS.
The rationale underlying the Recency principle is an attempt to extend the 'current line of
processing' - a focusing of attention: Give preference to the processes operating on the most
recently created or modified data. The rationale underlying the Specificity principle is: The more
conditions a production Ihs contains, the more specific is the situation the production is intended
to address likely to be, and specific prescriptions are given priority over more general ones.
When the dominating production instantiation is identified, the rhs of the production it refers
to is executed with variable bindings identical to the bindings from its Ihs. (These were stored as
part of the production instantiation.)
OPSS is an entire programming language, this brief description of its basic characteristics is
only an overview. More thorough descriptions may be found in the OPSS manual [19] and in the
textbook by Brownston et. a1. [6]. Two points should, however, be emphasized with respect to
production systems programming:
1) The textual occurrence of a rule within a production system program is insignificant.
2) There is no explicit sign of a rule's interaction with other rules present in the system, no
direct references to any rule from any other.
These features have been marketed as a blessing, since they make it possible to introduce new
pieces of program code into an existing program without having to insert it in a particular place
in the program code. No references to the new piece of code is necessary, since productions do
not 'call one another' as in procedural languages or as in Prolog, so in principle the existing
program does not have to be revised. However, in practice this may be a mixed blessing: The
new rule may execute in situations where it was not intended to, or it may fail to execute in
places where it was supposed to. This is due to the fact that it is not sufficient for a rule to
execute that it is present in the current conflict set - that its conditions for execution are satisfied.
It must also be the winner of the conflict resolution step - achieve the highest rating in the conflict
set as calculated by the conflict resolution strategy.
In principle, it should be possible to formulate a rule so that it specifies what should be done in
a given situation without concern for when this situation might occur in the execution course of
the program. However, in reality the situation description triggering this rule, call it rule A, may
coexist with a situation description triggering another rule, rule B, whose action part disables the
execution of rule A. (Formally, rule B may negate at least one of A's condition statements).
 554

If the production system programmer does not foresee this situation, and rule B has a higher
rating than A according to the conflict resolution mechanism in charge, then rule A will not get a
chance to execute. However, in order to foresee such a situation the programmer must have the
knowledge that there is a rule B that may prevent rule A from executing: He must be concerned
with what other rules might execute in the situation where rule A is intended to, and he may also
have to change some of the existing rules when new ones are added to prevent the old ones from
executing in situations that were unforeseen before the new rule was added to the program.
A rule may be explicitly prevented from executing by adding statements to its condition part
that limit the range of situations where it is satisfied. (Rule specialization) It is important that this
is done in such a way that the added statements can reasonably be related to the domain of
discourse. Otherwise, the rule may easily become a mixture of a rule representing domain
knowledge and an execution control element, its interpretation becomes unclear, and the
modularity of domain knowledge representation that a production system offers may become
very difficult to maintain as the system grows. Flexibility turns into chaos.

Structured Objects Representations; Frames

In 1975 Marvin Minsky wrote one of the more influential papers on knowledge representation.
The paper introduced the notion of 'frames', one ofthe most popular representation styles in the
years to follow [37]. Minsky's basic idea was that:
"The ingredients of most theories both in Artificial Intelligence and in Psychology have been
on the whole too minute, local, and unstructured to account - either practically or - for the
effectiveness of commonsense thought. The 'chunks' of reasoning, language, memory, and
perception ought to be larger and more structured; their factual and procedural contents
must be more intimately connected in order to explain the apparent power and speed of
mental activities [38}."
Minsky pointed to other recent works too, as
"moving awcry from the traditional attempts . . . in trying to represent knowledge as
collections if separate, simple fragments." (an implicit reference to expressions in predicate
calculus)
Brachman and Levesque later pointed to the core of Minsky's argumentation:
 555

"That the pursuit of formal logic in AI has been very misleading, and that some of the
demands of logic, such as consistency and completeness, are probably not even desirable in
a representation language [4j. "
The extraordinary influence of frame representations led Patrick Hayes in 1979 to take a
closer look at what this movement had produced since Minsky's original work in 1975:
"A frame is a data structure ... intended to represent a 'stereotypical situation'. It contains
named 'slots', which can be filled with other expressions - 'fillers', which may themselves be
frames, or presumably simple names or identifiers (which may themselves be somehow
associated with other frames,) .... For example, we might have a frame representing a typical
house, with slots called kitchen, bathroom, bedrooms . . . A particular house is then to be
represented by an instance of this house frame, obtained by filling in the slots with
specifications of the corresponding parts of the particular house, so that, for example, the
kitchen slot may be filled with an instance of the frame 'contemporary-kitchen' which has
slots 'cooker', 'floorcovering', 'sink', 'cleanliness', etc... Not all slots in an instance need be
filled, so that we can express doubt. .. Looked at this way,jrames are essentially bundles of
properties.... Thus far, then, working only at a very intuitive level, it seems that frames are
simply an alternative syntax for expressing relationships between individuals, i.e. for predicate
calculus. But we should be careful, since although the meanings may appear to be the same,
the inferences sanctioned by frames may differ in some crucial way from those sanctioned
by logic. In order to get more insight into what frames are supposed to mean we should
examine the ways in which it is suggested that they be used [25j. "
Minsky argues that

"traditional formal logic is a technical tool for discussing either everything that can be deduced
from some data or whether a certain consequence can be deduced; it cannot discuss at all
what ought to be deduced under ordinary circumstances [38]".
This emphasis on control of the inference process in frame systems points to what Hayes
terms 'the heuristic interpretation':
"that frames are a computational device for organizing stored representations in computer
memory, and perhaps also for organizing the processes of retrieval and inference which
manipulate these stored representations [25]."
Frame systems have been said to involve two fundamental issues that distinguish them from
other representational frameworks like for example the predicate calculus. The first is the use of
 556

procedural attachment in the retrieval and update processes, enabling very specialized inference
procedures to be used efficiently, because they are explicitly connected to the sets of data they
are meant to be applied to. The second is the ability to represent and manipulate descriptions of
and inferences on the representation itself frames, slots and values may all have attached
meta-information, frames describing the frames/slots/values. Hayes has emphasized this latter
point, pointing out that frames may be used as a medium for refleXive reasoning: reasoning
involving description of the self. In fact, Hayes claims that the emphasis on the analysis of
processes of reflexive reasoning is the only thing that is conceptually new. Apart from this he
claims that the frames idea's real force is not on the representational level at all, but on the
implementationallevel:
''How to organize large memories; that assertions should be stored in namable 'bundles'
which can be retrieved via some kind of indexing mechanism on their names. In fact, that
assertions should be stored in non-clausal form [25]".
The following section describes some central features of a frame-based representation system
called CRL - Carnegie Representation Language [7].

CRL -A Frame-Based Representation Language

CRL is a frame representation language developed at Carnegie Mellon University / Carnegie

Group Inc. It is, like most AI knowledge representation languages, based on Lisp. (Common
Lisp) Frames are termed schemata in CRL, and the schema is a particular type of Lisp object.
Schemata have slots (attributes), and slots have values. Slot values may be arbitrary Lisp objects
unless otherwise explicitly specified.
Schemata will here be displayed in the form:
{{schema-name
slot-1: value (or values; multiple values are allowed by default)
slot-2: value or values

slot-n: value} }
 557

Example
"Solution Xl consists of the distillation columns Cl and C2, and it has a high rating." may be
represented in CRL as:
{{solution} }
{{Xl
instance: solution
solution-elements: Cl C2
rating: high}}
{{distillation-column} }
{{Cl
instance: distillation-column} }
{{C2
instance: distillation-column} }
A set of schemata comprises a schema database, and this database may be queried or
manipulated using access-functions, such as:
(get-value schema slot) , which returns the value of slot in schema
(new-value schema slot value) , which sets the value of slot in schema
(add-value schema slot value), which adds the value to already existing ones
(delete-value schema slot value) , which removes value from slot in schema
etc.
Values may be inherited. Suppose that the database is:
{{separator
purpose: separate}}
{{distillation-column
is-a: separator}}
{{CI
instance: distillation-column
number-of-feeds: I}}
Then, the following query results may be obtained:
(get-value 'Cl'number-of-feeds) => 1
 558

(get-value 'Cl'purpose) => separate

The value 1 is the resident value of slot number-()f1eeds in schema Cl, but the value separate
is inherited from schema separator. Inheritance is possible because the special slots is-a and
instance belong to a set of special slots in CRL tenned relations.
The behavior of any slot in a CRL schema database may be specified using a slot control
schema for the slot. This is a schema which has the same name as the slot, where additional
information about it is stored. It is not necessary for a slot to have a slot control schema, but if it
has one, it typically contains the following information:
{{<slot-name>
domain: some value or expression
range: some value or expression>
cardinality: two integers
demon: a sequence of names of demon schemata
...... }}
The domain slot of the slot control schema makes it possible to restrict the use of the slot to a
set of schemata by use of a system-defined restriction grammar. The range slot makes it
accordingly possible to define the set of legal values a slot may have. Cardinality specifies an
upper and lower bound on the number of values the slot is allowed to have. If the demon slot is
used, it specifies pre-and/or post-processing procedures that will be executed automatically,
without being explicitly called (thus, the name) when the slot is accessed in a particular way. The
values of the demon slot must be schemata that describe the nature of the demons: There must be
one schema for every demon, and such a schema must contain the information:
{{<demon's schema-name>
instance: demon
access: the name of the access function(s) triggering this demon
when: before or after the access-function executes?
action: name of the Lisp function that constitutes the 'demon proper'
...... }}
If the same kind of pre - or post-processing is needed every time a particular slot is accessed
using a particular access-function - for example, delete-value - then the use of demons is a
convenient way of hiding behavior specifications from main program code.
 559

Example
{{vital-infonnation
demon: you-blew-it} }
{{you-blew-it
instance: demon
access: delete-value
when: after action: (lambda (schema slot value)
(tell-user schema slot value "is gone. Too bad!"))}}
This is a sketch of a demon executed every time a value in slot vital-information is deleted by
means of access-function delete-value. The demon executes after the value is deleted, and its
purpose is to tell the user that the value is gone.
If a slot is to be used as a relation, its slot control schema must be is-a or instance linked to
the system-defined schema relation, where the necessary 'relation-handling machinery' is stored
and may be inherited from.
Relations do frequently have additional slots in their slot control schemata. These - termed
inheritance specifications - allow specification of how to inherit information over the relation
link: Inclusion-specs specifY information that is allowed to be inherited. Exclusion-specs specifY
information that is not allowed inherited. Introduction-specs may introduce new information into
a schema when the relation slot is created in that schema. Map-specs specifY information
transformations during inheritance (English-metric unit conversion, for example).
The slot control schema is a global definition of slot behavior, the definition applies to all
occurrences of the slot in the entire schema database. But local specialization of slot behavior is
allowed using meta-slots: Any entity in a CRL database may have meta-information attached;
information describing information. A schema may have a meta-schema attached, a particular slot
occurrence in a schema may have a meta-slot attached (this is NOT the same as the slot control
schema, which has a global scope), and any slot value may have a meta-value attached. Both
meta-schemata, meta-slots and meta-values are represented as schemata, they may have an
arbitrary number of slots and values, and these may again have meta-infonnation attached;
arbitrary levels of meta-information is representable.
Meta-slots may have the same slots as slot control schemata - domain, range, demon, etc., and
if they have, these function as local overrides of the slot control schema. A meta-slot may be
 560

entirely local to the slot occurrence it is a meta-slot for, or it may be inherited down through a
hierarchy of schemata, thus making is possible to define localized slot behavior for arbitrary
classes of slot occurrences.
The value(s) of a slot in a CRL schema database may be arbitrary Lisp objects (unless
otherwise specified in the slot's slot control schema or meta-slot). Functions are valid Lisp
objects, so functions or function names are also valid slot values: Pieces of Lisp code may be
stored in CRL schemata, and results of CRL database retrieval may therefore be executed as
programs:
{{any-schema
slot-with-function-name-as-value: plus}}
(funcall
(get-value 'any-schema 'slot-with-function-name-as-vaIue)
2
2)
=>4
A CRL schema database may be partitioned in a tree structure using contexts. This is an
advanced feature of CRL that may be used to implement hypothetical reasoning (multiple
worlds) in a convenient way: Suppose that a decision point is reached in the problem solving
where several alternative decisions should be considered. Then a new context may be created for
each of these, and changes in the CRL representation of these decisions and their consequences
are stored in their respective contexts separated from each other. Only changes are stored,
incrementally: Unchanged information is inherited in any context C from the set of contexts
comprising a path from C up to the root-context, so that any such path comprises a consistent
wor/dview.

Object Oriented Programming

Object-oriented programming is a programming concept closely related to frame-based

representation systems. (Actually, it is the other way around; historically the concept of object
oriented languages was there almost a decade before Minsky's influential paper on frames:
Simula [10], the first object oriented language, was designed and implemented as early as in the
 561

mid-sixties by the NOlwegian researchers O.I. Dahl and K. Nygaard.) Today, e++ and Smalltalk
are probably the most widespread object oriented languages.
An object oriented program consists of a collection of objects and methods. Objects are

similar to frames; they have attributes with values, they are organized in hierarchies, and attribute
values may be inherited down through the hierarchy as described for frame systems. Methods are
procedures attached to particular objects' attributes. They may be equated with procedural
attachment to slots in frame systems. Methods are inheritable like attribute values.
Objects communicate by message passing. A message is a data structure consisting of three
parts; the verb, the receiver object and optional arguments. When the message is received by the
receiver object, the method associated with the verb is retrieved and executed. (Either there is a
procedure attached to the verb attribute in the receiver object, or the receiver object inherits such
a procedure from higher up in the object hierarchy it is part of) Sending messages to other
objects is typically part of any method's execution.
The philosophy underlying object oriented programming is that as much as possible of the
inner workings of an object should be hidden to the outside world. The outside world must
know that the object exists, that certain operations (verbs) can be performed on it, but the details
of how that is done is of no concern. It is the object's own responsibility to manage itself This
approach to programming makes it convenient to write very compact and high-level code once
the building blocks (the collection of objects and methods) are established. Object oriented
programming makes procedural abstraction very convenient and natural: Several distinct
procedures may be said to do conceptually 'the same thing', but to different types of objects and
consequently in different ways, without having to specifY such details in code calling on the
execution of these procedures. As an example, assume that the chemical components A and B
may be separated using either distillation, extraction or absorption. The goal is to simulate the
separation, and the simulation procedures 'dist-sep-simulation', 'extr-sep-simulation' and
'abs-sep-simulation' exist. Traditionally, one would then say:
call dist-sep-simulation (A, B, --specifications data), or
call extr-sep-simulation (A, B, --specifications data), or
call abs-sep-simulation (A, B, -specifications data).
Adopting an object oriented approach instead, one would create objects representing prototypical
distillation-columns, etc., and attach the procedures to these objects as methods, for example
with method name 'Simulate separation'. Then, one would create particular instances of these
 562

objects, with attribute values reflecting e.g. what solvent is used in the extraction unit to be
simulated, its size, operating pressure and temperature, etc. Finally, to trigger the simulation, one
would send the message:
Simulate-separation «instance>, A, B),
where <instance> would be the name of an instance of a distillation column object, an extraction
unit object or any other object that has a procedure associated with the verb (method) 'Simulate
separation'.
The primary distinction between object-oriented programming systems and frame-based
representation systems seems to lie more in how the systems are used than in how they differ in
their inner workings: While people who have used structured representation techniques as an
augmentation to procedural programming have termed their systems 'object-oriented', people
with a rooting in representation and inference have tended to use the phrase 'frame-based'.
The current tendency is integration of different styles of programming; logic-based,
production-based, procedural and structured-representation programming all present in the same
system. In such a context, it seems like objects and frames are evolving in the same direction.

Problem Solving Systems Architectures

This section addresses organizational frameworks for 'Expert-level' problem solving computer
systems - different logical architectures.
Hayes-Roth et aI. provide a categorization of Expert Systems applications [28]. It is proposed
that each of their categories have particular problem class characteristics, and that these
characteristics should be reflected in the organization of problem solving systems attacking these
classes of problems. Their classification is shown below:

Expert System Application Types

Interpretation Inferring situations from descriptions.
Prediction Inferring likely consequences of given situations.
Diagnosis Inferring system malfunctions from observables.
Design Configuring objects under constraints.
Planning Designing actions.
 563

Monitoring Comparing observations to plan vulnerabilities.

Debugging Prescribing remedies for malfunctions.
Repair Executing a plan to administer a prescribed remedy.
Instruction Diagnosis, debugging and repair of student behavior.
Control Interpreting, predicting, repairing and monitoring system behaviors.
However, systems architectures developed within one of these application areas may very well be
suitable for application in other areas. The blackboard architecture is an example in case. This
architecture was originally designed for speech interpretation [17], but over the last years
variants of the blackboard architecture have been used also in other domains; crystal structure
determination [57], planning [27], design .
While the early problem solving systems tended to be built from scratch, a number of system
building tools exist today, making it possible to start system design and implementation on a
higher lever than before. Ranging from simple PC-based rule-interpreters to powerful hybrid
representation and reasoning frameworks, it may often be hard to determine which system to use
for the problem at hand. Particularly, because flexible hybrid systems often will be harder to learn
and use than simpler, but more restricted ones. To guide system designers in their selection, a
number of 'Guides to Expert Systems' have been published [22,23,31,32,60].
The architecture of a system should reflect the characteristics of the problem it addresses.
While simple, uniform architectures may suffice for relatively limited problems, more complex
and heterogeneous architectures may be required when the problem becomes less structured and
less understood. (Apart from this, it should be stated, present system architectures naturally also
reflect the way the system designer believes the problem should be solved; subjective preferences
for particular problem solving strategies and formalisms.)
Stefik et a1. (in [28] Ch. 4) give an overview of Expert Systems architectures. They classifY
problems along three dimensions-problem size, reliability of data, and constancy of data- and
they suggest systems architecture guidelines for each of eleven classes of problems.

Problem Size; Decomposition, Precedence Ordering And Constraints:

Large problems must be decomposed to be solved. If one does not decompose such a problem,
then simple statistical arguments [55] demonstrate that they will be almost impossible to solve.
 564

There are many ways to decompose a problem. Functional decompositions decompose the
problem into subproblems that address different functional aspects of the problem. For example,
the problem 'design a chemical plant' may be decomposed into 'design the reactor system', 'design
the separation system', 'design the heat exchanger network', etc.
Decomposition into different levels of abstraction is another frequently used decomposition
strategy. For example, the problem 'design a chemical plant' may be decomposed into 'design the
Process Flow Diagram', 'design the Piping & Instrumentation Diagram {P&ID)', 'perform the
mechanical and thermal design', etc.
A third frequently used decomposition is decomposition into views, for example, 'a distillation
column as viewed by mechanical engineers', 'a distillation column as viewed by chemical
engineers', 'a reactor viewed as a heat source', , a reactor viewed as a chemical components
converter', etc. Real world decompositions often involve these three types, and probably others
as well, although this may not be explicitly recognized, because some of the decompositions are
implicit in the educational and professional training of human problem solvers. As an illustration,
when a chemical engineer is given a problem of the type 'design a process for production of
methanol from natural gas', he will typically come up with a Process Flow Diagram (PFD) and
consider this to be a solution to the problem stated. However, if you give this solution to an
instrumentation engineer and he will view it as part of a problem specification. (" One's roof is
another's floor", as the idiom goes.) It is rare that any given decomposition decomposes the
problem into completely independent subproblems; subproblems usually interact, weakly or
strongly. When subproblems interact, there are constraints on either the solution of the problem,
the solution process itself: or both, depending on the nature of the interactions: A P&ID requires
that a PFD exists - there is a strict precedence ordering constraint on the design of PFDs and
P&IDs, and the results of the design on the PFD level forms part of the problem specification for
the design on the P&ID level. In other cases, there is perhaps no strict precedence ordering, but
the order in which the subproblems are addressed has effect on the final solutions: If a car is
designed by designing the engine first and then the body, it is very likely that the results is going
to be different from what it would have been if the body were designed first.
The initial problem has an initial number of degrees of freedom. When subproblems are solved
they consume some of these, leaving fewer degrees of freedom to the remaining yet unsolved
subproblems. The consequence is that the remaining subproblems have to be solved within a
limited space of alternatives. Ifthe precedence ordering is selected poorly, there may not even be
 565

a single valid alternative left for the last subproblems, or there are only poor solutions left for
them. (If the validity of this claim is doubted, one should try to put on one's tightest pair of jeans
after having put on one's biggest pair of boots.)
The intuitive heuristic on precedence orderings coming out of this discussion is that if the
problem can be decomposed into subproblems of varying importance to solution of the overall
problem, then the most important ones should be addressed first, before the space of alternatives
gets too constrained by already made decisions.

Uncertainty

The reliability of data and inferences are important parameters in the design of system
architectures. In classical logic, statements may only be either true or false, and consequently
inferences on such statements are assumed to be either true or false, too. When premises are said
to be more or less probably true, and inferences more or less likely follow from the premises,
classical logic is no longer sufficient.
Mycin, an early medical diagnosis Expert System [53] and Prospector, an early geological
data interpretation Expert System [14], are examples of systems designed to work with
uncertain data and probabilistic inferences.
Formal methods exist for the solution of such problems - Bayes statistics- but Expert Systems
tend to use other less formal certainty factors and certainty propagation procedures. The
rationale for this has been that the experts providing the information both factual and judgmental
- present in these systems "do not use the information comparable to implemented standard
statistical methods. However, the concept of certainty factors did appear to fit .. . their judgment
of how they weighted factors, strong of weak, in decision making" [3]. Such subjective
probabilistic approaches to problem solving have become popular particularly in diagnosis and
interpretation applications. The main characteristic of such problems is that:
1) There exists a finite number of predetermined hypotheses, a set of more or less certain
observed data, and a set of probabilistic conditional statements (rules).
2) The task is to identify how probable the hypotheses are, given the data and their certainty
This is done by quasi-statistical (Bayes-like) inference combination of data and rules
with an update of probability- or rather degree of subjective belief - similar to the
 566

statistical calculation of posterior probabilities from observed events (data) and a priori
probabilities on conditional statements (rules) involving these observed data.
Besides giving an estimate of how strongly an hypothesis is believed be true, certainty factors
are also used as control parameters in systems taking this approach. During the inference process,
the various inferences produced by the system get their certainty factors updated. These are
propagated through the system so that the hypotheses' degree of belief also will change. There
are typically many hypotheses to be considered, and a heuristic frequently used is to focus on the
one with the highest current degree of belief (Rule applications may increase on decrease this
degree of belief). Shortcliffe [53], Davis [11] and Duda [14] describe such certainty propagation
procedures. Fuzzy logic is one of several other approaches to reasoning with uncertainty [63].

Assumptions and Expectations

Logical deduction is a necessary but not sufficient condition for problem solving. 'Educated
guessing' often plays a major role - making the appropriate assumptions and solving the problem
assuming these hold. Assumptions may of course tum out to be wrong, and in that case
experienced human problem solvers will be able to steps back, review and modifY assumptions,
retract the parts of the solution relying on invalidated premises and start over again, but not from
scratch; even the solutions based on faulty assumptions have given some insight into the
problem.
Problem solving is performed within a framework of assumptions and beliefs, and the sooner
the appropriate framework can be established the more efficient will the problem solving be;
assumption making is indeed another form of problem decomposition.
Assumptions may be verified up front by analysis or experiments; 'assume that no azeotrope
can be formed from a mixture of components A and B', they may be unverifiable in principle,
which is the case e.g. with predictions and forecasts; 'assume that the product market price over
the next five years will be X', or they may be verifiable only after decisions are made and
solutions are analyzed; 'assume that all the distillation columns may be run using low pressure
steam'.
If an assumption is verifiable, that does not necessarily mean that it should be verified
immediately. That depends on how strongly the assumption is believed to hold, and it depends on
the effect an assumption violation would have on the solution. The important thing is that they
 567

should be visible as assumptions as long as they are unverified - explicitly represented so that
they may be inspected later if possible and desirable. Dependencies should also be explicitly
represented when reasoning is performed on assumed premises: A derived results that depends
on an assumption may loose its foundation when the assumption is retracted - the result itself
might have to be retracted. In order to retract it when necessary, it must be explicitly represented
how it was derived. For a discussion of dependencies see [8].
So-called hypothetical reasoning - What if .. .'- reasoning, also termed 'Multiple worlds'-
reasoning, is an approach to reasoning with uncertainty, based on assumptions and expectations
instead of certainty factors and statements that are true to a certain degree. Systems
incorporating hypothetical reasoning allow inconsistent statements to coexist in different worlds:
In one world, a statement may be considered to be true, and inferences depending on the truth of
this statement are derived in this world. In another coexisting world the same statement may be
considered to be false. This world will therefore produce other inferences than the former.
It is the present author's opinion that decision making systems (as opposed to systems where
the truth of an hypothesis is to be determined) lend themselves more naturally to the hypothetical
reasoning approach than to the certainty factor approach. The justification for this position is that
decision systems are primarily concerned with consequence and utility - not truth: A decision is
not primarily true or false - it may be more or less appropriate, better or worse. A medical doctor
may be 20% or 70"10 certain that Mr. Smith has a particular disease, and this degree of belief will
hopefully affect the treatment prescribed for him. However, when it comes to acting on the
belief - deciding what kind of treatment to give, there are only two alternatives - either assume
that he has the disease, or assume that he dose not. And, a chemical engineer may be 30% or
60% certain that a particular separation task in a petrochemical process may be performed using
a simple flash drum instead of a more expensive distillation column, but in order to verifY or
falsifY the belief, an analysis must be carried out where either the drum or the column is present.

Meta-Level Problem Solving

In the section addressing search, it was emphasized that several competing operators are
typically applicable to a particular state. Selection among these could be guided by the evaluation
of a rating function. The rating function was therefore said to control the search. However, in
complex decision making problem domains, typically involving uncertainty and assumption
 568

making, where analysis and decision making both are part of the problem solving process, it is
frequently hard to identifY a single simple rating function capable of rating properly all the
actions that could conceivably be carried out as part of the problem solving process: Alternative
decision making operators must compete among themselves for priority, but they must also
share resources with analysis operators - operators capable of identification, verification or
falsification of critical assumptions. The control problem - determination of what to do next in a
complex problem solving system - may therefore itself become a complex task.
Operators may be divided into distinct base-level and meta-level categories [20], where
base-level operators are those which when strung together solve the domain problem the system
is intended to solve, and meta-level operators are those involved in deciding which base-level
operators to apply next. Genesereth [20] emphasizes the importance of keeping base-level and
meta-level components as 'separate agents'; not letting an operator have both these roles.
Inclusion of a meta-operator layer makes it possible to bridge the gap between a general purpose
interpreter and domain specific base-operators without confusing control and domain knowledge
in the base-operators.
Why Is It Important Not To Confuse Control And Domain Knowledge? The main issues are
modularity and flexibility. Complex systems tend to grow incrementally, operators are added one
at a time. If a particular set of operators contain information assuming the presence or absence of
other operators, addition of new operators or removal of existing ones may have wide-ranging
effects on those left unchanged.
Hayes-Roth et. al. argue that "An intelligent system reasons about its actions", and that "In
order to perform effectively, an intelligent system must have knowledge about its actions" [30].
Such reasoning on knowledge about actions implies that the system must have -or be able to
generate - information describing the purpose and consequences of operator applications. The
consequences of applying an operator should of course not have to be known in full detail in
advance; if that were the case it would not be necessary to apply the operator, it is superfluous.
However, basic characteristics of it could be described so that a meta-level layer of operators
would be able to forecast an approximate potential utility of applying it, like for example the
typical time it needs to execute, the expected quality, accuracy and level of detail of the results it
produces, etc.
 569

Goal- and State- Driven Integration of Independent Operators

The collection of operators comprising a system reflects an implicit decomposition of the

problem the system is intended to solve- an operator addresses a task that is part of the problem.
The purpose of the operator is to perform the task. How the different tasks in a problem are
related can be represented either explicitly or implicitly and with or without subgoaling.
Subgoaling is a mechanism where operators respond to goal statements present in the current
state description, where a goal corresponds to a request for a task to be performed, and the
operator responding to the goal statement either performs the task to completion or posts
subgoals, thereby letting operators capable of performing subtasks of its task execute to their
completion before completing its own task. Every operator must have encoded in it what kind of
tasks it may solve.
With such a subgoaling scheme, there is no need to represent the purpose of an operator
explicitly. Its purpose will be to respond to a goal statement, and if goal statements are labeled
with the identity of the goal statement activating the operator posting these goals, there will be a
link upwards through all currently suspended operators from the currently active one all the way
up to the top level goal.
It is important to note the interaction between goal statement, task and operator. The operator
does not refer to any particular sub-operator. It posts a goal, and once the goal is posted the
sub-operator 'sees' that the goal refers to a task that it can respond to. It is therefore in principle
not necessary that anything but goals are communicated among operators in this type of system.
There may of course be many operators capable of solving the same type of task. In that case,
one of these will have to be selected among those applicable. This is typically the case in
problems where there are alternative solution methods to solve a problem, possibly with varying
accuracy, detail and computational resource requirements. In such cases, the goal statement may
be used to specify restrictions on the current range these parameters are allowed to take. The
operator performing the goal-task matching these restrictions better than any of the others
performing the goal-task may than be selected for execution and executed. Meta-level operators
may be used to determine what constitutes a 'better match'.
The above assumes that there is only one active goal at any time-it is a depth-first
goal-oriented architecture. Real world decision problems are frequently more complex: Many
 570

different tasks may be perfonned as the next one, leading to a situation where many goals should
be considered as the one to pursue next. Now, the relevance of achieving a particular goal to the
overall solution of the problem will have to be considered, too: Which of the currently active
goals will promote a good solution and an efficient solution process the most?
Two approaches may be taken: try to first establish which of the uninstantiated goals to
pursue, and after that decide which operator is best suited to achieve the selected goal.
Alternatively collapse these two decision phases into one: Decide directly which applicable
operator should be selected without any separate concern for the goal it achieves.
It may be argued against the latter that this looks a lot like adding apples and oranges. To
some extent that is a valid objection, but in the alternative approach- select goals first and then
ways to achieve them - infonnation available on the expected characteristics of potentially
applicable operators is not considered; infonnation that could have an impact on the selection of
the next goal to pursue is not used.
Architectures without subgoaling are also possible: Problems can be decomposed so that an
operator that requires subgoaling is decomposed into
a) the part of the operator that is executed before subgoaling would occur.
b) the part of the operator that is executed after.
It is not even necessary to represent the subgoaling explicitly: The state description at the
instance where subgoaling would occur may be used instead of the goal statement as the
statement telling the sub-operator that it should signal its applicability. So instead of looking out
for a goal statement to appear, the operators look out for specified data patterns in the state
description. This is tenned data-driven processing: The state description determines the set of
applicable operators.
The major difference between goal- and data-driven architectures may be phrased as follows:
In a goal-driven, scheme operators performing particular tasks are requested. In a data-driven
scheme, there are no requests, operators 'volunteer' their bids for execution.
The terms goal- and data-driven are frequently confused with the terms backward- and
forward-chaining, but these two concept pairs are orthogonal: Backward-and forward-chaining
are concepts describing how a system implementation is operated, while goal-and data-driven are
concepts describing how the problem is decomposed into operators and whether operators are
invoked voluntarily or on request. The stereotypical view is that data-driven systems chain
 571

fOlWard while goal-driven systems chain backwards. However, goal-driven systems may also
chain fOlWard, as in:
"If Goal X is currently active
then do something procedural ...
and post subgoal Y
and do something more procedural ... "
In both cases - data-driven or goal-driven - the applicable operators should submit a signal,
globally visible to the system, that they are applicable. They should not be allowed to execute
before the system realizes system wide that they are applicable. If they were, there would be no
way eventual meta-operators could submit their 'opinion' on whether an operator should be
allowed to execute or not. Such a globally visible signal - a bid for execution - should contain as
much information about the operator submitting it as possible, because the more the system gets
to know about the operators the better is the basis for deciding which to choose.
Goal-driven system architectures, where several goals are allowed to be active at any
instance, may be viewed as a special case of a data-driven system architecture. The rationale for
this claim is that the relevance of any active goal needs to be computed in order to determine
which of the alternative goals to pursue as the next. The goal is there, but this does not any
longer have an imperative interpretation in the sense that one in the set of operators able to
address this particular goal is guaranteed to execute next. The set of operators have to compete
with other sets of operators trying to achieve other active goals.

Hearsay H - An Early Example of Integration of Goal-and Data Driven Processing

The Hearsay IT speech understanding system [17] was an early system addressing this kind of
problem. Hearsay IT's domain was 'close-to-real-time' interpretation of connected speech natural
language sentences in a restricted vocabulary containing approximately 1000 words. The system
consisted of three parts; a 'Blackboard' -the global data structure holding the representation of
the current state, a set of 'Knowledge Sources' (KSes) - independent pieces of code performing
the different domain tasks the system could perform, and a scheduler-a mechanism for selection
of Knowledge Sources to execute next.
The Hearsay IT blackboard was partitioned into a 2-dimensional grid, one dimension
comprising the different levels of abstraction that the speech understanding system had to work
 572

on, the other dimension representing the time interval during which the sentence was uttered.
Most of the Hearsay II KSes worked on the border between two of the Blackboard levels,
fonning hypotheses on the upper level from the data on the level below or generating data on the
lower level under the assumption that hypotheses on the upper level were true, extending the set
of hypotheses along the time dimension and comparing and rating hypotheses originating from
different KSes. In other words, both the KSes and the Blackboard data structure reflected a
particular decomposition of problem and solution methods.
Hearsay II can neither be classified as goal-driven nor as data-driven. It can best be described
as applying an opportunistic reasoning model - applying pieces of knowledge in the direction
towards either goals or data in an opportune manner. The input data to Hearsay 11-
parameterized voice signals- is the basis of hypotheses formed on the syllable level. Signal
patterns may indicate that a particular syllable has been uttered in a given time interval in the
sentence. Combinations of recognized syllables may further indicate that a particular word has
been uttered, etc. This kind of processing is data-driven; data patterns suggest hypotheses to
verifY. When a number of words have been hypothesized as possible explanations of segments of
the voice signal, further processing may be done on alternative interpretations of such a segment,
establishing a relative likelihood that a particular hypothesized word is the correct one by
comparison of an expected signal pattern (such expected patterns are stored within the system
for each word it is able to recognize) with the observed pattern. This is goal-driven processing;
attempting to identifY which in a set of alternative hypotheses to believe the most. Alternatively,
combinations of words may be combined into word phrases, for example a verb phase if the
hypothesized word is a verb, word phrases may be combined into full sentences, and from the
assumption that every utterance conforms to a specified grammar, a grammatical analysis may
then be applied to detennine whether an interpretation of a segment of the utterance constitutes
a legal statement in the language described by this grammar. Formation of hypothesized word
phases and sentences are data-driven activities, analysis and verification of these are goal-driven
activities. Hearsay II switches opportunistically between these two modes of reasoning, all the
time trying to:
1) move up the levels from signal description (bottom level) to semantic content (top level)
ofthe sentence,
2) expand the time interval of the currently most strongly supported hypothesized sentences,
 573

3) discriminate among alternative hypothesized interpretations.

The meta-level or control reasoning in Hearsay IT- deciding which KS to execute next- is all
contained in the systems' scheduler, a single, complex fixed piece of procedural code whose task
it is to predict where on the blackboard grid it is most profitable to focus, what kind of
processing is needed and which KS is the best one suited to do this.
The scheduling mechanism is described in [26], and Hearsay IT as a structure for problem
solving is described in [15]. A description of the implementation may be found in [17].
Hearsay IT originated a new class of problem solving systems - blackboard systems. Nti
[46,47] gives an overview of "The blackboard Model of problem solving and the Evolution of
Blackboard Architectures." She states that "Subsequently, many application programs have been
implemented whose solutions were formulated using the blackboard model. Because of the
different characteristics of the application problems and because the interpretation of the
blackboard model varied, the design ofthese programs differed considerably." The HASP/SlAP
system (military system for identification of various types of ocean vessels from digitized
hydrophone data; [45]), the Chrysalis system (structure interpretation of X-ray diffraction data
for protein crystals; [57] and OPM (simulation of human errand running protocols; [25]) are
examples in case.
The essence of the control (meta-level) organization in OPM has been extended and modified
into the domain independent blackboard system BB 1 [29]. The core idea underlying BB 1 is that
domain and control problem solving are intertwined, and therefore the domain and the control
problem should be solved simultaneously. It is suggested that the control problem should be
approached as a real-time planning problem.
The basic control scheme in BB I is very simple; three small and simple basic control KSes are
activated iteratively. These three basic KSes update the agenda of KSes that may be executed,
select one for execution, and when selected, interpret it. The reason that the scheduling
mechanism may be kept so simple is that additional control knowledge may be encoded in more
or less domain specific control KSes that, when selected, alter the criteria for KS selection in a
way reflecting general or domain specific problem solving strategies. The control problem and
the domain problem are solved simultaneously, but BB 1 specifies two separate blackboards - a
control blackboard and a domain blackboard. The structure, contents and vocabulary of the
domain blackboard will vary from one application to another, but the structure, contents and
vocabulary of the control blackboard are determined by BB 1. A thorough (but unfortunately not
 574

too readily comprehensible) description of the control blackboard may be found in [29]- it is too
lengthy and complex to be accounted for in any detail here. The key point is that the control
blackboard specifies a control plan for the solution of the domain problem on several levels of
abstraction and detail:
Desirable actions or classes of actions according to some explicitly represented (named)
strategy.
The current focus of problem solving.
Past and future foci.
Policy decisions determining weighs in a multi objective rating function.
The rating function itself
The set of actions that may be immediately performed and their variable bindings.
The KS currently being interpreted.
All this information is represented declaratively on the control blackboard, and any piece of it
may be modified by control KSes. This makes the BB 1 architecture extremely flexible and
potentially applicable to a wide range of problem domains.
The generality of BB 1 makes it a complex tool to use, and instead of building application
systems directly on top of BBl, work has later been focused towards providing higher level
system building tools on a level between BB 1 and specific applications.
ACCORD is one such problem solving framework for arrangement assembly tasks built on
top ofBBl [30]. It is hypothesized that assembly problems constitute a distinguishable class of
problems that may be approached specifically with a set of control structures and
'semi-domain-independent' concepts tailored to problems of this class. The concept of planning
islands, a problem solving strategy particularly emphasized in Stefik's system MOLGEN [56] is
underlying the design of the ACCORD framework. ACCORD also provides generic verbs to use
in the description of problems modeled as arrangement assembly tasks, and the system 'knows'
how to interpret these, thereby providing a higher level description language than the underlying
BB 1 representation that ACCORD statements map into.
The major benefit of having such a layered structure is that control and representation
paradigms proven to be useful in a problem are made available to similar problems without
having to redesign and reimplement it all down to the lowest level of the system building tool.
 575

AKORN D -A Blackboard Model For Process Synthesis.

A problem solving system may be viewed from at least two points of view: The structure and
representation of the domain specific information the system is supposed to manipulate, and the
structure and representation of the manipulating mechanism itself- the system architecture. It is
naturally desired that the problem solving architecture to a large extent should reflect the basic
characteristics of the domain problem to be solved. One of the basic characteristics of process
design and synthesis problems is their open-endedness, there is hardly ever one correct solution,
but a lot of better and worse ones. Another basic characteristic is that such problems are usually
configurative; the solutions of the problem do not pre-exist in form of a fixed set of hypotheses
to be verified or falsified, it is a major part of the problem solving process to generate them. A
third characteristic is that the overall problem solving strategy seems sensitive to special
characteristics of the domain problem, for example the identification of bottlenecks in the design
- dominant features of the problem pointing to essential decompositions and precedence
orderings once the features are recognized. A fourth characteristic is the mixed use of analysis
and assumption making used to limit the search space. A fifth characteristic is the use of multiple
approaches to the problem- more or less accurate analysis methods, heuristics based on one
point of view or another, and several levels of abstraction on which the problem may be
described.
These characteristics are typical for problems that have been approached using blackboard
models. It was therefore a natural choice to attempt to use a blackboard model approach here,
too. The proposed architecture, which was given the name AKORN D for obscure reasons,
attempts to be general and flexible but at the same time simple. Concepts were borrowed from
BBI, but AKORN D distinguishes itself from BBI in that it is much less complex: In BBI, a
substantial effort is put into representing declaratively extensive control plans on multiple levels
of abstraction independent of application problem domain. In the design of AKORN D, it was
felt that such complexity may, when it comes to representing domain control knowledge, be
counterproductive - it requires substantial insights in the domain problem to be able to formulate
and represent extended control plans, and it may not even be fruitful to any worthwhile extent to
do so in an early development phase. So the opposite approach was taken - it was assumed to be
sufficient to be able to decide what to do next in a qualified manner. Crucial prerequisites to be
able to do this are:
 576

The system has a system-wide awareness ofwhat can be done next.

It is able to apply control knawledge to select among the options available.
As in BBI, it was decided to parameterize the basic KS selection function and to keep the
parameters represented on a separate control blackboard distinguishable from a domain solution
blackboard, where emerging domain solutions are represented. It was also decided to retain the
BB I concept of separate domain and control KSes.
The KSes may be viewed as (simulated) independent processes, each constantly monitoring a
common data area looking for data patterns indicating their relevance and responding to these.
When a KS finds a pattern it can respond to, it posts a data element called a bid on the control
blackboard. The bid is a signal to the rest of the system that a particular KS considers itself to be
useful. It contains information about the characteristics of the KS posting it and about this KS's
own opinion of the importance of being executed in the current situation.

Knowledge Sources: In AKORN D, KS consists ofthree parts:

I) A monitor looking for characteristic patterns: The monitor submits a bid when it detects
'its' pattern. This bid is put onto the control blackboard.
2) A declarative representation of the task the KS performs and characteristics of the way
the KS achieves it.
3) An action part that when executed performs the actual task and typically manipulates
one of the blackboards.
Domain KSes monitor and manipulate the domain blackboard only. They comprise the
available domain actions- the set of actions that may be performed next to promote or extend an
emerging solution represented on the domain blackboard. Control KSes essentially monitor and
manipulate the control blackboard. They do also have monitoring access to the domain
blackboard, but they are not allowed to manipulate it.

Scheduler: The scheduler is a low-priority independent process whose purpose it is to select a

bid from the collection of bids present on the control blackboard and hand over execution
control to the action part of the KS that posted this winning bid. The winning bid is selected by
means of a selection function and the parameters currently present on the control blackboard.
Figure I illustrates this information flow.
 577

Domain-
knowledge-
source

Para-
melers

Control-
blackboard

Control-
know ledge-
source

Scheduler

Pallem malching
Bid submission
C1I:) Trigger

Modification of blackboard
Basic loop c=) Action

FJgUJ"e 1. AlCORN D information flow

AKORN D operates in cycles. In one cycle,

a) All KSes that detect their potential for application submit a bid which is placed on the
control blackboard.
b) When the bid submission halts the low priority scheduling process takes control and
selects one of the bids on the blackboard.
c) Execution control is handed over to the action, part of the KS that submitted the selected
bid.
When the selected action is completed a new cycle is started. In the meantime, a blackboard has
typica1ly been modified so that there are new data patterns to respond to for the set of KSes.
 578

Execution may be tenninated either explicitly by a KS executing a halt command or by the

scheduling process if no bids are present on the blackboard.
AKORN D as described above may be termed an extended production system:
It is based on pattern matching
It operates in cycles
There is no direct communication among rules (KSes). They communicate solely through
working memory (the blackboards)
In every cycle a 'conflict set' is formed and 'conflict resolution' is performed on it in the sense
that bids are submitted and KS selection is based on information contained in these bids.
Yet, AKORN D is different from a conventional production system: The conflict set (the set
of bids) is an integral part of the problem state because it is explicitly represented in working
memory (on the control blackboard This is an important feature in that it allows the system to
respond to signals about what it considers to do before actually doing it, and does so
independently of proposing it. For example, a control KS containing knowledge about situation
where one task is preferred over another may trigger when these two task have bids signaling
their potential application on the blackboard. This is readily done without having the domain
KSes contain references to the conflict situations or to each other. Domain knowledge (which
options are available) and control knowledge (which actions are preferred in the current
situation) is clearly separated.
Using previously defined terms, domain KSes may be viewed as base operators, while control
KSes are meta-level operators (or, if a control KS triggers on the presence of a control bid-
meta-level, etc. ). An interesting point is that all levels manifest themselves in the same basic
cycle.

Conflict Resolution in AKORN D: Nothing has yet been said about the conflict resolution
mechanism in AKORN D- the basic mechanism for selection of a bid and thereby of the next
action to perform. This basic mechanism is a parameterized selection function, where the
parameters are explicitly represented on the control blackboard and thus modifiable by control
KS actions. The selection function parameters describe the current problem solving focus of
attention in a domain independent way. The following are the key points justifying their form and
existence: The design of AKORN D attempts to allow several partial solution to the problem at
the same time on the blackboard, so the system has to decide which of the present solutions to
 579

work on. The KSes thus have to be explicit about which solution they propose to extend. Every
solution must be provided with a rating describing how promising it is currently believed to be.
The system may also decide to focus on a particular set of task. Since it is desirable to be able to
add new task without having to modifY existing code, there is no explicit reference to task
names. It is instead assumed that the problem task can be said to exist on a level, so that a
reference to the level is an implicit reference to a set of task. The KSes have to know their own
level. The bids submitted by a KS on a particular level will thus represent potential actions on
that level. Within a level there may exist alternative actions. A domain KS recognizing its
applicability will also to some extent be able to predict the goodness of the action it proposes.
Some problems have more strongly predefined precedence ordering than others. The amount
of search needed to find a good solution depends very much on the problem characteristics, so it
should be possible to modify the width of the search, both in terms of moving among alternative
solutions and in terms of moving around on different levels in any particular solution. AKORN
D distinguishes among three different classes of bids: Decision bids are bids from KSes that
make decision: A partial solution gets closer to completion by application of decision KSes.
Analysis bids are bids from KSes that perform analyses (generate information) without making
any decision. Control bids are bids from control KSes. These propose to modifY selection
function parameters. Control KSes do not exist on any level.
The selection function parameters present on the control blackboard are:
Current-ps: The partial solution currently in focus. A partial solution name.
Current-level: The level currently in focus. A level name (a positive integer).
Current-rigor: (Applicable to analysis bids only)
The current importance of accurate analyses.
(a number in the range from 0 (unimportant) to 1 (important))
Current-speed: (Applicable to analysis bids only)
The current importance of fast analysis.
(a number in the range from 0 (unimportant) to 1 )
Ps-dop: Degree of focusing on the current partial solution.
(number ranging from 0 (no focusing) to 1 (strong focusing))
Level-dop: Degree offocusing on the current level.
(number ranging from 0 (no focusing) to 1 (strong focusing»
 580

Four additional parameters are used by the scheduler to compute bid priorities:
Ps-support: A binary switch with value 1 if the bid addresses the current partial solution,
-1 otherwise
Level-support: A binary switch with value 1 if the bid addresses the current level, -1
otherwise
Opgood: The 'subjective opinion' of the KS submitting the bid on the potential value
of its application.
(number ranging from 0 (not very promising) to 1 (very promising»
Psgood: The 'subjective opinion' of a partial solution on its relative value.
(number ranging from 0 (not very promising) to 1 (very promising»
Opgood and Psgood will clearly be strongly domain dependent. AKORN D does therefore
not specify how to compute these numbers, but a mechanism is specified for their determination:
It is assumed that every partial solution and KS is represented as an object with a method
attached or inheritable which computes these numbers. These methods are allowed to have
read-access to the control blackboard.
Domain bids (decision or analysis bids) are rated according to the following selection function:
DBP =[2 + Ps-dop * Ps-support]
* -2 + Level-dop * min {Level-support, Ps-support * Level-support}]
* 1 + Opgood]
* 1 + Psgood]
DBP is a number in the range from 1 to 36. It should be pointed out that level here becomes
subordinate to partial solution: It rarely makes any sense to promote actions on solutions out of
focus even if they are on the current level. Level focusing is thus viewed as a refinement of
solution focusing. Control bids are rated according to:

CBP = 1 + 6(1 + Opgood)

CBP is a number in the range from 7 to 37.

The rationale for these rating functions is an attempt to represent desired cross-effects of the
parameters involved: Control KSes represent control heuristics which may address more or less
specific situations. Although any KS signals its application through an active assessment of the
situation - it does not suggest itself unless it identifies the current situation as interesting, there
 581

may be situations where both general knowledge and more specialized knowledge compete for
application. The general guideline is that the more specific set of situations a KS is intended to
handle, the higher should the Opgood values of the bids it submits be. If high rated domain bids
are present-the actions they represent seem locally very promising, they address a promising
solution, this solution is currently in focus, the level is the current one and the focusing is very
strong-then only very specialized control knowledge is allowed to interfere. (for example to
discriminate between two particular (named) very promising actions, if such control knowledge
is present in the system.)
When the rating of domain bids drop, a larger set of control KSes become competitive-that is,
if the situation is interesting to them; they have submitted bids. But still the more specialized of
them will dominate. There are several reasons for domain bids to get low rating. Often these
reasons may themselves indicate the proper control actions to get back to high ratings:
There may not be any more good bids on the current level in the current solution.
If good bids are present on another level in the same solution and the solution is believed to be
good then this suggests to focus on this other level.
If good bids are present in another solution on a particular level and this other solution is
believed to be comparable with the current one, then this suggests to switch focus to the
solution and level where the good bid is.
If it is unclear what level and solution to focus on, this may indicate that search is needed to
identifY where to focus: Reduce the emphasis on focus-lower the values for Ps-dop and/or
Level-dop and work where it locally seems to be good until a more specific control pattern is
recognized and acted upon.
Bids representing defocusing control actions should in general probably have lower Opgood
values than bids representing focusing ones. This because focusing aims to reduce search effort,
while defocusing increases it.
AKORN D attempts to provide an additional degree of freedom in representing domain
knowledge as compared with a traditional production system: A KS may have its conditions for
execution partitioned among
1) the KS trigger, which determines whether the current situation is interesting at all to this
particular KS,
2) the selection function and its current parameters, which decide whether this KS in the
current situation is interesting enough to allow its execution,
 582

3) (an arbitrary number of) control KSes which are capable of changing the current situation
(the focusing on a particular solution, leve~ and also the degree of such focusing) so that
a locally interesting action may become interesting enough to be selected.
The original AKORN D architecture was is implemented using the development system
Knowledge Craft as a set ofCRL schemata, a set of Lisp functions and an OPS5 rule driving the
basic loop. An in-depth description of the implementation, as well as an example of how the
architecture was used to generate heat integrated distillation sequences using mainly qualitative
information, may be found in [35].

S6 - Automatic Synthesis of Separation Sequences with a Blackboard Based System

combined with a Mathematical Programming Model

The S6 program combines Knowledge Based and Operations Research techniques in the
generation of solutions to sharp split heat integrated distillation sequence problems. The KBS
parts oj the system construct superstructures tultomaJically. These superstructures are
subsequently pruned down to sets of cost optimal solutions by combinatorial optimization.
KBS techniques are furthermore used to represent complete and emerging solutions, as well as
to allow some extent of user interaction in the problem solving process.
The overall structure of S6 is illustrated in Figure 2. The 'knowledge based' parts of the
program are implemented on a Symbolics Lisp machine running Knowledge Craft, while other
parts of the program (e.g. parts of GAMS input files and the GAMS optimization system itself)
reside on a SUN workstation connected to the Symbolics via Ethernet. The knowledge based
parts of the program are implemented using the AKORN DT blackboard architecture, an
extension of AKORN D. The parts of the program residing on the SUN workstation are invoked
on demand from the Symbolics computer. Communication between the two machines is
controlled by a separate communication module. Some user interaction is allowed via a graphical
interface.
AKORN DT is an extension and reimplementation of the AKORN D blackboard architecture.
The 'T indicates that tentative search has been included in the new version of the architecture. Both
AKORN DT and its predecessor are opportunistic control architectures, meaning that there is no
explicit representation of a hierarchy of goals to be fulfilled. Any KS may in principle be activated at
any time, providing a 'broad focus of attention'. The bids submitted onto the control blackboard
 583

serve as a globally visible and explicitly represented set of alternative next moves. This set
resembles the conflict set in production systems, but its explicit representation as part of the
problem solving state provides improved opportunities for the use of domain specific information in
the conflict resolution.

SYMBOLICS

Domain-
knowledge-
source

Numerical program File Trigger Action

Figure 2. S6 overall structure

Some basic modifications had to be made to AKORN D in order to facilitate tentative search:
Incremental representation of partial solutions, combined with explicit representation of
backtracking points and alternative next moves. In AKORN Dr, incremental representation of
partial solutions has been achieved by use of the Context Mechanism provided in Knowledge
Craft: When the action part of a KS is activated, the results generated by this KS are stored in a
new context, which is the 'child' context of the context where the KS triggered. Only new or
 584

modified information is stored in the new context, unaltered information is made available
through inheritance from ancestor contexts. This is a representation form which makes it easy to
preserve consistency during tentative search, and the set of possible backtracking points may be
defined in a very general way: Every context is a potential backtracking point. Furthermore, if
previously unused bids from domain KSes are stored with an association to the context where
they were submitted, then the possible next moves from any context may also be defined in a
very general way: Every unused bid associated with a context represents a possible next move
from that context. Wahl presents an in-depth description of AKORN DT [58].
On a high level of abstraction, the tasks performed by S6 may be described as follows:
1) A separation problem to be solved is specified by the user on the Symbolics machine.
2) A first solution to the problem is generated automatically by the program and displayed.
3) The user may decide that the solution is the desired one, or that parts of the proposed
solution should be modified. (The modifications are limited to modifications of the
superstructure from which optimal solutions are extracted)
4) If the user requests it, an alternative solution may be generated automatically, and the
new solution is displayed. Repeat from step 3) until the user accepts the solution or no
more solutions can be generated.
Automatic generation of the first solution involves the following steps:
a) Generate the set of possible separation tasks needed to separate the mixture described in
the separation problem specification. Store this in a semantic network (e.g. a frame
structure) on the domain blackboard.
b) Simulate, size and cost each of the separation tasks. Augment the semantic network with
this information.
c) Generate a logical superstructure consisting of separation columns on different levels.
Store this superstructure in the semantic network.
d) Construct an input file to the combinatorial optimizer GAMS from the information
contained in the semantic network on the domain blackboard. Perform the optimization,
extract a description of the optimal solution from the GAMS output file and store this
information in the semantic network on the domain blackboard.
As currently implemented, S6 may use two different mathematical programming formulations of
the problem - the Andrecowich Transshipment formulation or the corresponding Transportation
formulation.
 585

All infonnation about the emerging design solutions is stored in a semantic network (e.g. a
frame structure) on the domain blackboard. The network consists of schemata (the Knowledge
Craft notion for 'frames' or 'objects') describing entities such as columns, streams, and hot and
cold utilities, or representing decisions / alternatives and relations among these. A simplified
graphical presentation of the network is given in Figure 3.

top-
produc
bottom-
product

Figure 3. Part of the semantic network in AKORN DT

The knowledge sources currently implemented in S6 are basic knowledge sources which are
general and rather lean on heuristic information. Other knowledge sources implementing specific
heuristics can be added to compete with or supplement the basic ones.
The domain knowledge source Generate-Columns generates a network of streams and
conceptual columns. This knowledge source does not include any heuristic knowledge, it
generates all possibilities. Later, other knowledge sources may be added that either create
another network or modify this network. This reflects a basic principle adapted by us, believed to
 586

be rather important when the purpose of the knowledge based parts of a system is to provide a
superstructure that is subsequently going to be pruned down to one or a small set of solutions by
OR methods: Heuristic elimination of a subset of the set of logically possible alternatives will
usually be based on a weaker foundation than a cost based elimination in the subsequent
optimization of the superstructure. Therefore, within the limits imposed by computational
tractability (e.g. the size of the superstructure), the superstructure should to the extent possible
contain the entire set oflogically possible solutions to the problem.
The domain knowledge source Simulation-Conditions determines the pressure and flow
conditions for the simulation of the columns. Four simulations of each column are required
-simulations at two different pressures for two different amounts of total feed flow to the column.
Simulate-columns performs the simulation of the columns for the different simulation conditions
determined from the previous described knowledge source.
The knowledge sources Suggest-HI- Transh-MILP and Suggest-HI-Transp-MILP are
conceptually equal: "If there is a separation problem which does not have any models associated
with it and there exist columns that have been simulated, then suggest to we the transportation
(transshipment) modelfor the problem."
Use-Char-Temp-Alg uses the characteristic temperature algorithm to create a superstructure
of columns [1]. The knowledge source generates an input file to the characteristic temperature
algorithm. The contents of the input file is created on the Symbolics machine and transferred to
the SUN machine, where the characteristic temperature algorithm is executed. The result file is
transferred back to the Symbolics, where it is transformed to another format, to be stored in the
semantic network.
The two domain knowledge sources Use-Gams-Hi-Transh-Milp and Use-Gams-Hi-
Transp-Milp use the optimizer GAMS to extract the optimal solution from the ones embedded
in the superstructure. They are almost equal, so they can be explained together ."If there is a
separation problem with the transportation/transshipment model and there exists a set of
superstructure columns and no solution exists, then suggest to we GAMS to solve the problem."
First, this knowledge source creates the 'dynamic' parts of the GAMS input file, including e.g.
sizes of sets, values of parameters and so on. 'Static' parts of the input, e.g. indexed model
equations, are stored in other files. A series of Unix commands for the SUN machine are then
invoked from the Symbolics. These commands merge the different input files, start GAMS, and
extract the relevant information from the GAMS output file. The extraction of information from
 587

the output file is done using AWI<, an efficient UNIX pattern matching utility program. These
results are retrieved to the semantic network and the solution is displayed.
It is up to the user to decide what to do with the displayed solution - four alternatives are
available through a menu in the user interface. Each of the four choices are carried out by
activation of a control knowledge source: User-Response-Halt stops further execution of the
system. It is used when the user indicates satisfaction with the solution. User-Response-Ok "If
there is a user response with the value OK and there exists a unused bid (other altematives are
possible), then suggest to explore this alternative bid." The knowledge source finds the context
where this bid was submitted and makes it the default context (i.e. moves to this context) and
lets the bid execute. User-Response-Next adds an integer cut - the current solution is not allowed
in the next attempt. This causes the second best solution to be generated. User-Response-Edit
signals that the user has chosen to edit the superstructure.
The domain knowledge source Edit-Formulation is triggered when the user has signaled that
he wants to edit the current superstructure. It pops up a graphical superstructure editor, where
the user may move the temperature / pressure level a superstructure column is defined on, delete
an existing column, or add a new column.
The Graphical User Interface: The main purpose of the graphical interface is to display the
solutions and the superstructures in a convenient format when the user has the possibilities to
interact with the system. The graphical interface has been implemented using the Knowledge
Craft WindOW/Graphics System. The screen is divided into different areas (viewports and
windows). In the upper left comer is a list of icons, in the upper right comer is the name, at the
bottom there is a command window. The rest of the screen consists ofa viewport for display of
costs for the selected solution, one for display of the superstructure, one for display of the labels
used in the superstructure, one for display of the heat integration pattern, and one for display of
the mass flow pattern. This is illustrated in Figure 4 and described in some detail below:
The Command Window allows the user to type in Lisp commands instead of using icons and
menus. The Icon Window consists of five icons: Quit, Problem Specification, Run, Display, and
Edit. All of theses are activated by clicking the mouse bringing up a menu. Quit makes it possible
to pause (temporarily leave and return) or to exit (leave with no return). Problem Specification
brings up a slot filling editor where the specifications for the separation problem can be filled in.
Alternatively, one has the option to read the problem specification from a file. Run starts or restarts
the system (loading of files + a few commands). Display gives a menu for the display of the
 588

viewports COst, Superstrocture, Heat Integration and Mass Flaw. Edit is used for modifying the
superstructure. The possibilities are: Delete Column, which removes a superstructure column, Edit
Column, which allows the user to modify the temperatures for the condenser and reboiler by the
use of a slot filling editor, and Create Column, which allows the user to generate another
superstructure column by use of the same slot filling editor.

S5
~Iulpl .. 5wp a"IH b','.lI."UD &lqlltDle S,)Dliuul
~
'fI":'It Tntf' ";'1'1nn

A ""<)1'_ Tolal :21827 ....

Ii 1-fJUTANE.
C ~ANE.
Col ••n 18ntn .•
o 2-o.ETHYU!lJTANE.
E _fANE. Heet ExchangerIS4S •••••
Operation 115173 •• '

------,~-------~-

" .

ft... 5 "- U.11,31 pew a. cu. ...

Figure 4. S6 user interface

The Cost VleHpol1 displays some cost variables returned from GAMS annualized cost, which
is the objective function of the math. programming mode~ investment cost for the columns and
the heat exchangers, and operating costs cold and hot utilities). The purpose of this window is to
 589

give the user infonnation about the distribution ofthe cost and thus hopefully make it easier to
evaluate the returned solution or decide whether to modifY the superstructure or not. The
Superstructure Viewport displays the superstructure generated by the characteristic temperature
algorithm. The superstructure columns are represented as white boxes. When a solution has
been generated, the selected superstructure columns are displayed as black boxes with an index.
Utilities are shown as horizontal lines -hot utilities are solid and the cold utilities are dashed. The
splits are labeled by letters for compactness in the graphical representation, and displayed labels
are described in an own window above the superstructure window.
In the Mass Flow Viewport, each column is represented as a black box with an index (which
also occurs in the superstructure viewport) and the name of the light and heavy key. The Heat
Flow Viewport describes the heat integration of the selected solution. Columns are here
represented as trapezoids in a temperature-energy coordinate system, resembling the
Andrecowich box representation. (Trapezoids are used instead of rectangles here, because a
column's reboiler and condenser duties are not necessari1y equal.) When the transportation
fonnulation is used, one trapezoid may be displayed above another, to represent graphically the
heat integration of the corresponding columns. When the transshipment fonnulation is used, on
the other hand, the optimization results do not explicitly tell which columns to integrate. The
graphical representation will in these cases not display any trapezoid on top of another.
There are also two other windows, which will nonnally be buried by the others: One for the
OPS process (OPS is used for generation of bids and for triggering of the AKORN DT
scheduler) and other messages from the Symbolics system, and one terminal window for
processes running on the remote Unix: machine. These are shown in Figure 5. The fonner is
useful for debugging purposes. The latter may be useful e.g. for the monitoring of progress in
time-consuming optimization runs.
Chemical components are organized in a hierarchical network structure with multiple
inheritance. Components and component classes are implemented as schemata (objects) where
each slot can have values or methods attached. This gives rise to possibilities for object oriented
programming. Physical property parameters for each compound are stored in each schema.
Mainly the slots of the schemata are filled with quantitative data taken from [50]. Physical
properties, e.g. vapor pressures, may be computed by alternative equations, where a particular
 590

S6
11111"1,1. sta.. ,. SpUI 5.,.u40lla.I',G ••,e Sratbn',

~~.,~..".~.. ~.!ii.~..~~~.IJ!!P ( ...... I

)pIiiiiiiiiiilmii:iiiiiiiiii~!ii_~_~;ij__ ii!
. '.h, '.- . I'''.,ICI'-U,...."., - .U(,., )44MiN1." .
,, : I1fJaUL
__ ..... ~I~t .... II ... . H .,. . -...d I"", _ ~ C~~~: ~t~.""",::' : .~'~ ~~~~ ... : ~ . '~U )J
'I_._,- t.,,,-
• ! 'CLI."9C:
C : cy~rAHf. . 1h • • .arJo .... t.II..~ :
" ._.l. I"c.. ...-k ...... CI\.~
. .. ~,, - ,~ .• t"' . ~'" t..c. ... 111 . . . C

. ·:'r.
..... "" '-r.'~-"'- ···.,i. :.:::.
I; r ... , . C.l_ c:....t.

Figure 5. Hidden trace and debugging/monitoring windows

equation is better suited for some components or components classes than others. In S6, each of
the components (or component classes) have stored references to which of these methods to
apply, so that the calculation method needs not be known 'outside' the component. This gives the
calculations a flavor of object oriented programming: When the program needs e.g. the vapor
pressure of a component, a message is sent to the component, saying that the vapor pressure at a
given temperature is needed. The choice of method is then detennined by the component 'itself.
The functions used in S6 for the simulation and cost estimation of the columns and heat
exchangers are to a large extent the shortcut methods reported in [13]. Interfacing of S6 to a
commercial simulator has been considered, but since the main purpose of the program is to
demonstrate the integration of AI and OR techniques, we were not willing to spend the required
effort. The required minimum set of shortcut methods was instead implemented in Lisp in the
 591

version of the system reported here. It should however be mentioned that in an even more recent
reduced implementation of S6 in C++ [59]. The program has been interfaced to the commercial
simulator ASPEN+ for generation of simulation, sizing and costing data.

Final Remarks

This paper has presented an overview of some important elements of knowledge based systems.
The overview has by no means been intended to be complete. Instead, we have tried to focus on
what the current author believes to be the central issues. It is expected that knowledge based
systems techniques will continue to play an important role in the research and development of
useful computer aids for process engineers.
Two key issues described here -pattern recognition and search, are expected to have a
profound influence on the development of such systems, but it is difficult to predict what form
this will take. Our - and others' -experience with pattern directed knowledge based methods in
the form of rule based systems indicate that there is a need for more powerful pattern
recognition methods than those offered by symbol oriented rule based languages, e.g. pattern
detection in multivariable continuous variable spaces. Currently, such isolated pattern matching
features are offered both by e.g. Neural Networks and by several more established statistics
based pattern recognition algorithms. It may therefore be expected that the future may bring
hybrid systems where symbolic problem solving and search is integrated with this kind of
advanced pattern matchers. This would in turn have a profound effect on representation
techniques: The various representation techniques described in this paper are essentially symbol
oriented, and this is most like going to be insufficient if more advanced pattern matching
techniques are going to be used. What is needed under these circumstances is a subsymbolic
representation - a representation where explicit symbols are not necessarily explicit in the
representation but may possibly be generated as needed as a kind of abstraction.
In search, it may be expected that knowledge based search techniques may be combined with
the more powerful analysis methods of mathematical programming, because mathematical
programming allows a wide range of powerful tools from mathematics to be applied. Object
oriented representations will naturally also play a role in this development process, but in the
design of object oriented representations one should be aware of the fact that the information
 592

hiding characteristic of object oriented programming may in fact in many cases make advanced
pattern matching more difficult rather than easier, simply because the data to match are hidden
and not globally visible to the pattern matcher.

Acknowledgments

The work reported here has been funded in parts by the Norwegian Science Foundation (NFR),
Norsk Hydro, Statoil and the Nordic Petroleum Research Program. Thanks to Prof A. W.
Westerberg, Carnegie Mellon Univ., for many inspiring discussions during the development of
the AKORN blackboard architectures, and to Dr. P.E. Wahl who developed the AKORN DT
architecture and S6 as parts of his Ph.D.

References

1. Andrecowich, M: Synthesis of Heat Integrated Distillation Sequences, Ph.D. Dissertation Thesis, Chern. Eng.
Dept., Carnegie-Mellon Univ., PA, (1983).
2. Andrecowich, M. and WesteIberg, A. W.: A Simple Synthesis Method based on Utility Bounding for
Heat-Integrated Distillation Sequences, AlChE 1. 31, p.363, (1985).
3. Barr, F. and Feigenbaum, E.: The Handbook of Artificial Intelligence, W. Kaufinann, Los Altos, CA, (1982).
4. Brachman, R and Levesque, EI.: Readings in Knowledge Representation, Morgan Kaufinann, Los Altos, CA
(1985).
5. Brooke, A., D. Kendrick and A. Meeraus: "GAMS - A Usel's Guide", The Scientific Press, Redwood City, CA
(1988).
6. Brownston, L. et a1.: Programming Expert Systems in OPS5, Addison-Wesley, Reading, MA, (1985).
7. carnegie Group: Knowledge Craft 3.1 Reference Manual, Carnegie Group Inc., Pittsburgh, PA, (1986).
8. Chamiak, E., Rie&leck, C. K. and McDermott, D. W.: Artificial Intelligence Programming, Lawrence
Eribaum Assoc., Hillsdale, NJ, (1980).
9. Clocksin, W. F. and Mellish, C. S.: Programming in Prolog, Springer-Verlag, Berlin, (1981).
10. Dahl, O. 1. and Nygaard, K.: Simul-an Algol-Based Simulation Language, ACM Comm. 9, p.671, (1966).
11. Davis, R and Lena!, D.: Knowledge-Based Systems in Artificial Intelligence, McGraw-Hill, NY, (1982).
12. Douglas, 1.: A Hierarchical Decomposition Procedure for Process Synthesis AlChE 1. 31, p.353, (1981).
13. Douglas, 1. M: "Conceptual Design of Chemical Processes", McGraw-Hill, NY (1988).
14. Duda, R O. et al.: Final Report, SRI Projects 5821 and 6415, SRI International Inc., Meulo Park, CA, (1978).
15. Erman, L. D. and Lesser, V. R: A Multi-Level Organization Cor Problem Solving using many, diverse,
cooperating Sources oCKnowledge, Proc. 4th IJCAl, TSibili, USSR (1975).
16. Erman, L. D. and Lesser, V. R: System Engineering Techniques for Artificial Intelligence Systems, in
Hanson, A. and Riseman, E.: Computer VISion Systems, Academic Press, NY, (1977).
17. Erman, L. D. and Lesser, V. R: The Hearsay-II System: A Tutorial, in Lea, W. A. (ed): Trends in Speech
Recognition, Prentice-Hall, Englewood Cliffs, NJ, (1978).
18. Forgy, C. L. and McDermott, 1.-: OPS, a domain independent production system language, Proc. 5th IJCAI,
MIT AI Lab., Cambridge, MA (1977).
19. Forgy, C. L.: OPS5 UseI'sManual, Tech. Rep., Dept ofComp. Sci., carnegie Mellon Univ., PA, (1981).
20. Genesereth, M: An overviews of meta-level architecture, Proc. 3rd Ann. Cont: on AI, Los Altos, CA (1983).
21. Goldberg, A. and Robson, D.: Sma1Italk-80: The Language and its Implementation, Addison-Wesley,
Reading, MA, (1980).
 593

22. Goodall, A: The Guide to Expert Systems, Learned Infonnation Ltd., Oxford, England, (1987).
23. Hannon, P. and King, D.: Artificial Intelligence in Business. Expert Systems., John Wiley, NY, (1985).
24. Harris, L. R: The heuristic search under conditions ofelTOr, Artificial Intelligence, 5, No.3, p.217, (1974).
25. Hayes, P. 1.: The Logic ofFI1IIIIes, in Metzing, D. (ed): Frame Conceptions and Text Understanding, Walter
de Gruyter & Co., Berlin (1979).
26. Hayes-Roth, F. and Lesser, V. R: Focus of Attention in a Distn"buted-Logic Speech Understanding System,
Tech. Rep., Comp. Sci. Dept, Carnegie-Mellon Univ., PA (1977).
27. Hayes-Roth, B.: Modeling Planning as an Incremental, Opportunistic Process, Proc. 6th UCAl, Los Altos CA,
( 1979).
28. Hayes-Roth, F. (ed): Building Expert Systems, Addison-Wesley, Reading, MA, (1983).
29. Hayes-Roth, 8.: A Blackboard Architecture for Control, Artificial Intelligence, 26, p.251, (1985).
30. Hayes-Roth, 8. el. al.: A Modular and Layered Environment for Reasoning about Action, Tech. Rep.
KSL-U-38, Stanford Univ., CA (1987).
31. Hewett, J. and Sasson, R: Expert Systems 1986, Vol 1: the USA and Canada, Ovwn, (1986).
32. Hewett, J. and Sasson, R: Commercial Expert Systems in Europe, Ovwn, (1986).
33. Kraft, A: XCON: An Expert Configuration System at Digital Equipment Corporation, in Winston, P. and
Prendergast, K A (Eds): The AI Bwiness: The Commercial Uses of Artificial Intelligence, MIT Press,
Cambridge, MA (1984).
34. Lien, K Suzuki, G. and Westerberg, A W.: The Role of Expert Systems Technology in Design, Chern. Eng.
Sci., 42, No.5, p.l049, (1987).
35. Lien, K M: "Expert Systems Technology in Synthesis ofDistillation Sequences", Dr.Ing. Thesis, The
Norwegian Institute of Technology, Trondheim, Norway (1988).
36. Lien, K M: "A Framework for Opportunistic Problem Solving, Computers and Chemical Engineering, vol.
13,00. 4/5 (1989).
37. Minsky, M: A Framework for Representing Knowledge, in Winston, P. (ed): The Psychology of Computer
Vision, McGraw-Hill, NY, (1975).
38. Minsky, M, in Haugeland, J. (ed): Mind Design, The MIT Press, Cambridge, MA, (1981).
39. Newell, A: Some Problems of Basic Organization of Problem Solving Programs, in Yovits, M C. (ed): Proc.
Cont: on SelC-Organizing Systems, Wash. D.C., (1%2).
40. Newell, A, Shaw,1. C. and Simon, H.: Empirical Explorations of the Logic Theory Machine: A case study in
heuristics, in Feigenbaum, E. and Feldman, 1. (Eds): Computers and Thought, McGraw-HilI, NY (1%3).
41. Newell, A, et. al.: Information Processing Language V Manual, Prentice-Hall, NY, (1964).
42. Newell, A, and Simon, H.: Human Problem Solving, Prentice-Hall, NY, (1972).
43. Newell, A: Production Systems: Models of Control Structures, in Chase, W. C. (Ed): Visuallnfonnation
Processing, Academic Press, NY, (1973).
44. Newell, A: The Knowledge Level, Artificial Intelligence, 18 No. I, p.87 (1982).
45. Nii, H. P.: HASP 1SlAP Case Study, The AI Mag., Spring, p.23 (1982).
46. Nii, H. P.: Blackboard Systems: The Blackboard Model of Problem Solving and the Evolution of Blackboard
Architectures, The AI Mag., Summer, p.38, (1986).
47. Nii, E. P.: Blackboard Application Systems, Blackboard Systems from a Knowledge Engineering Perspective,
The AI Mag., August p.82, (1986).
48. Nilsson, N.: Principles of Artificial Intelligence, TiogaPubI. Co., Palo Alto, CA, (1980).
49. Pearl, 1.: Heuristics, Addison-Wesley, Reading, MA, (1984).
50. Reid, R C., 1. M Prausnitz and B. E. Poling: "The Properties of Gases and Liquids", McGraw-Hili, NY
(1987).
51. Robinson, 1. A: A Machine Oriented Logic Based on the Resolution Principle, JACM, 12, No. I, p.23, (1%5).
52. SanGiovanni, 1. P. and Romans, H. C.: Expert Systems in Industry: A Survey, Chern. Eng. Progr., Sept-87,
p.52 (1987).
53. Shortcliffe, E. H.: Computer Based Medical Consultations, Elsevier, NY, (1976).
54. Simulation Sciences Inc.: PROCESS Reference Manual, Simulation Sciences Inc., Fullerton, CA, (1984).
55. Simon, E.: The Sciences of the Artificial, MIT Press, MA, (1%8).
56. Stefik, M: Planning with Constraints, Ph.D. dissertation Thesis, Comp. Sci. Dept, Stanford Univ., CA,
(1980).
57. Terry, A: The Chrysalis Project-Hierarchical Control of Production Systems, Tech. Rep. HPP-83-19, Stanford
Univ., CA, (1983).
 594

58. Wahl, P. E. "Synthesis ofHeat Integrated Distillation Systems - Approaches Combining Artificial Intelligence
and Operations Research Methods", Dr. Ing. Thesis, The NOIwegian Institute of Technology, The University
ofTrondheim, Trondheim, Norway (1991).
59. Wahl P. E. and K M Lien: "S6 - A Computer Progrnm for Automated Synthesis of Heat Integrated
Distillation Systems", SintefTech. Report, The SintefGroup, Trondheim, Norway (1992).
60. Waterman, D. A: A Guide to Expert Systems, Addison-Wesley, Reading, MA. (1986).
61. Wehe, R. Lien, K and Westerberg, A W.: Control Architecture Considerations for a Separations Systems
Design Expert, Proc. NFS-AAAJ Wolkshop on Artificial Intelligence in Process Engineering, Columbia Univ.,
NY, March-87 (1987).
62. Winston, P. and Eorn, B. K P.: LISP, Addison-Wesley, Reading, MA. (1984).
63. Winston, P.H.: "Artificial Intelligence", Addison-Wesley, Reading, MA (1984).
64. Zadeh, L.: A Theo'Y of Approximate Reasoning, in Michie, D. (Ed): Machine Intelligence Vol. 9, Ellis
Horwood, Chichester, (1979).
Selected Topics in Artificial Intelligence for Planning and
Scheduling Problems, Knowledge Acquisition, and
Machine Learning

Aydin K. Sunol, Muzaffer Kapanoglu, Praveen K. MogiJi

College of Engineering, University of South Florida Tampa, FL 33620, USA

Abstract: Application of artificial intelligence techniques to planning and scheduling problems are
briefly reviewed, issues involved in knowledge acquisition are discussed, and methods for machine
learning are introduced.
More in depth treatise of selected approaches, with processing systems engineering examples,
are used to illustrate some of the issues involved. The examples utilize genetic programming for
batch scheduling and design problems, and a hybrid of symbolic and connectionist approaches to
automated knowledge acquisition (machine learning).
Concepts and issues involved in Genetic Programming are discussed within the context of
batch processing systems examples. A scheduling example is used to illustrate discrete variable
based decision models and associated terminology while a design problem with continuous
decision variables and constraints is also optimized. Possibility of using genetic programming as
an integrating tool for computer integrated manufacturing is also discussed.
A novel instance based learning algorithm that allows symbolic information to be encoded into
a connectionist representation is introduced. Selection of complex distillation column sequencing
designs is used as the example. However, the approach is ideally suited for fault diagnosis and
structured selection problems. The knowledge/rules extracted from the learning algorithm are
very similar to the design heuristics proposed in literature. The performance of the Symbolic
Connectionist network (SC-net) is studied based on its knowledge extraction capabilities and the
classification accuracy in the test case.

Keywords: Machine learning, planning, genetic programming, artificial intelligence, batch process
scheduling, optimal design of batch processes
 596

Planning And Scheduling

Introduction

Planning is a difficult task for both natural and artificial intelligence. A non-exhaustive list of its
various dimensions, with some overlap of the items, include: achieving goals, satisfying
constraints, minimizing cost, maximizing profit, resolving conflicting goals, handling undesirable
side-effects, scheduling, learning from failure, generalizing and reusing old plans, and planning in
a dynamic environment. These dimensions bring forth the complex and heterogenous nature of
planning which is a NP-hard problem where the components are usually tackled hierarchically,
in part through a portfolio of techniques.
Batch production or manufacturing problem is often tackled in two phases, process planning
and scheduling. In the planning phase, a partial ordering of batch production operations or
activities is generated to achieve one or more goals. The scheduling problem is that of assigning
resources and times for each activity in a plan such that the ordering relations imposed by the plan
and the capacity limitations of shared resources are met.
The planning and scheduling problem could be tackled through optimization based, heuristic
based, and knowledge based approaches. Each class of methods has inherent advantages and
shortcomings over the other two. Optimization-based schedulers guarantee optimality in a static
way. Combinatorial optimization techniques and graph-theoretic approaches are effective for
small-size schedules. However, as the search space increases, heuristic methods become necessary
for reduction of the search space and result in a tractable problem. Unfortunately, the heuristic
methods do not guarantee the best solution, or indeed any solution. For complex problems, testing
heuristic methods is difficult and only comparative tests are possible. In the relevant literature,
including the articles in this book, many recent surveys of the optimization and heuristic methods
are available for the manufacturing scheduling problem. Knowledge based schedulers use expertise
in dealing with uncertainty and complexity. The approach has comparative advantages in
capturing human expertise and intermeshing efficiently both the heuristic and the algorithmic
methods. Knowledge acquisition step appears to be a serious bottleneck in AI-based approaches.
Recent surveys of the knowledge based approaches to scheduling include a review by Noronha
and Sarma [25]. The current thinking on use of artificial intelligence techniques in manufacturing
may be best found in a book [8] which is an outgrowth of several workshops held by the American
Association of Artificial Intelligence (AAAI) special interest group on manufacturing. For
 597

researchers in Artificial Intelligence in Planning, a recent monograph that contains a collection of

articles [22] may also be of interest.

Knowledge Based Approaches to Planning And Scheduling

A comparison of discrete manufacturing systems is shown in Figure 1. At one extreme, we have

mass production which is characterized by high production rate, high production volume, reduced
cost, special tooling, minimal labor skill, product flow type, fixed automation, and very limited
product variety. At the other extreme, we have job-shop manufacturing with low production rate,
low production volume, higher cost, general tooling, high-level labor skill, process-based plant
layout, programmable automation, and high product variety. The batch production may be
considered an intermediate within these extremes. The trade-offs between the approaches at the
extremes may be addressed in part through Flexible Manufacturing Systems (FMS). Modular
Manufacturing Systems (MMS) have been conceptualized to further expand the flexibility [! 5].
The motivation here is to be able address the scheduling and planning problems ofMMS, which
in tum would hopefully facilitate effective solution of the batch scheduling and design problems.

Low S ecial Toolin High

High

General
Job-Shop Manufactwing

Programmable
Automation
t MODULAR MANUFACTUR!NG
SYSTEMS

Flexible Manufactwing
E quipment
Part
Variety Labor Batch Production
Specialization

...
MODULAR
- Fixed Automation
Mass Production
MANUFACTUR!NG SYSTEMS.

..
Special

Low
...
Process P /ant layout Product Flow

Low Production Volume High

Figure 1. Characteristics of Discrete Manufacturing Systems

 598

General Approaches and Considerations

There are many issues to consider in the development ofan AI-based scheduler. One dimension
of scheduler classification [I6] include Expert Systems [17,33], Deeper Method-Based systems,
and Interactive schedulers [4,33]. Batch-Kit [18] presented in this NATO ASI falls between the
categories of Deeper Method-Based and Interactive.
Another dimension is whether the scheduler is predictive (deliberative) or reactive. Most
optimization methods and simulation provide static (predictive) schedules that may quickly
become obsolete over time. A completely different approach to the aforementioned deliberative
planning techniques is the reactive systems [3]. The idea here is to avoid planning altogether, and
instead use the observable situation as a clue to which one must react. The system must have
access to a knowledge base in order to know what actions to take under which circumstances. The
reactive approach is instrumental in interleaving planning and plan execution. Deliberative planners
that learn to be reactive over time seems to be the direction in which practical planning systems
may emerge [23].
Equipment break-down, operator availability, raw material delivery, continual arrival of new
orders, and product quality assurance are typical elements that contribute to the dynamic
environment of a batch plant. As the dynamism of the environment increases, it becomes more
difficult to distinguish the difference between the predictive and reactive behavior, so the need for
real-time systems arise [35].
Another consideration involves, distributed versus centralized batch scheduling systems. The
advantages of the former to the latter include [8]:
a) The distributed systems address the problem of coordinating a number oflocal schedulers
which may each be responsible for scheduling a part of the work of a larger proj ect better.
b) The underlying scheduling model's organization, if represented as a network of scheduling
agents, can easily be changed dynamically to adapt to changes on the plant (e.g. addition
or removal or rearrangement of the equipment).
The solution to planning problems often involve decomposition. Hierarchical planning
involves first solving the problem completely, considering only preconditions whose criticality
value ( a measure of expected difficulty of satisfying preconditions) and augmenting the plan with
 599

operators that satisfy the preconditions. It could be viewed as a length-first search. As the
computational power keeps increasing, popularity of concurrency and aggregated approaches will
grow.
One shortcoming of rule based systems is the context dependency of the production rules and
definitions, limiting the effectiveness of the approach to narrow domain applications. Case Based
Reasoning aim to remedy the aforementioned shortcomings. The application of the idea in
scheduling problems would be Case-Based Planning [12] where the old plans are re-used to
make new ones.
In the planning stage, the design and production information is translated into instructions for
production of either intermediates (parts) or end-products through a combination of human
planners and software packages. Since these plans are not always executed as planned, the system
must therefore monitor the execution of the plans, learn from the successful as well as
unsuccessful plans, and replan the process. These issues, knowledge acquisition and machine
learning are discussed in more depth in the second part of this paper.

Search Methods

In order to solve most non-trivial problems, it is necessary to combine some of the basic problem
solving strategies (search methods) with one or more of the knowledge representation
mechanisms. Although search and knowledge representation issues are discussed as another paper
in this book, applicable search related practice will be briefly discussed.
Once the problem is defined as a state space search problem [27], search algorithms such as
A· could be used to find a solution. A· provides a way of conducting best first search through a
graph representation of the problem. Each node that is examined by the algorithm represents a
description of a complete problem state, and each operator describes a way of changing the total
state description. For more complicated problems, decomposition into smaller subproblems and
ways of avoiding recomputing entire problem state, as it changes, is necessary. AD· algorithm
[24] provides a way to decompose the problem when it is completely separable.
One of the earliest planning methods is goal-stack planning where individual goals are solved
one at a time in order. This approach can not handle cases where goals interact. Constraint
Posting is an alternative for such nonlinear planning problems. The basic idea here is to build
up a plan by incrementally hypothesizing operators, partial orderings between the operators, and
 600

bindings of variables within the operators. The differences between state space search and
constraint posting could be found in most text books on AI techniques [31].
In order to solve more complicated problems, it is important to be able to eliminate some of
the detail of the problem until a solution that addresses the main issues is found. Subsequently, the
appropriate details are filled in. These hierarchical planning approaches differ in the way the
search problem is decomposed and the operators used.
Although many knowledge-based approaches and mathematical programming techniques have
been suggested to solve scheduling problems, a single universally acceptable technique that could
solve the problem is yet to emerge. In the recent years, Genetic Algorithms attracted considerable
interest [9,10,11,14,19]. It seems to be a quite robust search technique and an effective tool for
Machine Learning in Knowledge Based Systems [13]. The literature in GA is quite scarce,
particularly in Chemical Engineering literature. The recent interest to the field, its promise for
scheduling problems, the robustness of the approach as a search technique, and utility in machine
learning are responsible in the extended coverage of the subject.

Genetic Programming

Genetic Programming (GP) is composed of genetic modelling (GM) and genetic algorithms (GA).
Genetic Modeling involves description of a system using genetic structures, such as genes,
chromosomes, alleles and so on. For instance, any number between 0 and 7 can be modeled as a
chromosome with three genes by using a binary alphabet (0 and 1). If the number is 6, it could
then be genetically modeled by using the following chromosome:

Q...llJ2
The differences between genetic algorithms and conventional search methods could be
summarized as follows [6]:
a) GAs work with a coding of the parameter set, not the parameters themselves.

b) GAs search from a population of points not a single point.

c) GAs use payoff (objective function) information, not gradients or other auxiliary
knowledge.
d) GAs use probabilistic transition rules, not deterministic rules.
Genetic Algorithms manipulate the genetic models to obtain optimal solutions to the problem
represented by this model. This manipulation utilizes genetic operators such as crossover,
 601

mutation, mating, and reproduction.

Since GA was invented to mimic some of the processes observed in natural evaluation, its
description is inevitably based on the biological vocabulary. Table I provides the equivalence of
traditional and GA search terminology. The flow-chart of a simple genetic algorithm is presented
as Figure 2.

Table 1. Comparison of the TernUnologies of Genetic Progranunmg and Operations Research

GENETIC PROGRAMMING MAmEMATICAL PROGRAMMING

(Decision Theory)

Genetic Model (Genotype, Structure) Mathematical Model-Decision Model

Genetic Algorithm (Reproduction Mathematical A1gorithm- Solution

Technique) Technique, Decision Making

String (Chromosome) A Mathematical Statement- A Constraint, an

Objective Function, etc.

Gene (Feature, Character, Detector) Mathematical Variable- Decision Variable

Allele (Feature Value) Value of a Mathematical Variable- Value of

a Decision Variable

Locus (String Position) -

A Solution of Genetic Model (phenotype, a A Feasible Solution of the Model- A
Decoded Structure) Decision

Generation Iteration- Next (Improved) Decision

Epistasis Nonlinearity, Multi-modality- Complex

Decision

Alphabet Variable Types- Decision Types

Population A Set of Feasible Solutions

The number of genetic operators that participate in reproducing an organism determines the
complexity of that organism. Hence, more sophisticated problems require more advanced genetic
operators. This leads to finer models and, thus, better solutions (Dominance, diploidy, haploidy).
In this study, simple genetic programming, hence, basic genetic operators are introduced and
 602

Figure 2. Flowchart of a Simple Genetic Algorithm

discussed in the context of batch process systems engineering.

Selection: Selection process is perfonned based on fitness values (a low fitness value means
poor adaptability to the environment or large deviation from the goal) of chromosomes
(structures, alternative policies, ... etc.).
Reproduction: This operation determines the number of replicates of each selected
chromosome used in the new generation (iteration). A higher chance to contribute to
reproduction is given to those chromosomes with higher fitness values (lower cost, higher profit,
briefly, better adaptability to the environment). Usually, reproduction and selection are
implemented at the same computational stage.
Example: It is very common to use roulette-wheel to allot off-spring strings.

Parent # Fitness Expected # Actual Count (next generation)

1 6 1.8 2
2 1 0.30 0
3 3 0.90 1
Total 10 3 3
Average 3.33
 603

Crossover: The crossover operator provides the discovery mechanism for GAs. Therefore,
when modeling a problem, the type of crossover operator to be applied should be considered
carefully. A comprehensive coverage of the possible crossover types will not be attempted, a
rather simple one-point crossover will be given for illustration.
Parent #1 B B B B B B Parent #2 A A A A A A
Initially, a crossover point is selected randomly, all chromosomes are then cut and spliced. If the
crossover points are in the middle, then following siblings are obtained:
Child # 1 A A A B B B Child #2 B B B A A A
This operator provides information exchange between strings. Crossover enables combination
of building blocks of better solutions from parent chromosomes. Since the new population is
generated from superior parents (or bit strings with higher fitness value), crossover allows
information exchange among the elite. It is expected that some superior parents may have many
features in common. These common characteristics (similarities) of some parents are reflected on
similarity templates. Each similarity template is called a schema. Holland's Schema Theorem
asserts that genetic algorithms manipulate schemata during its execution [14]. Since GAs allow
only better parents to reproduce, one can predict the relative increase or decrease of a certain
schema in the next generation. The expected number of occurrences of schema H in the next
generation of the population is n*r/a minus disruptions caused by mutation and crossover [2].
Here, r is the average fitness of all chromosomes in the population containing the schema H; n is
the number of chromosomes; and "a" is the average fitness of all chromosomes.
The reproduction ends with the growth and decay of the several schemata contained in a
population. How the growth and decay of some schemata may lead to an improvement is
illustrated next. The bit string [ 0 1 1 I 1 1 0 0 0 1 0 0 1 1 0 0 0 1 0 1 ] which can symbolically
be written as A = [ a 1 a 2 ... a 20 ] consists of genes (a j, i = 1, 2, ... 20 ), each of which represent
a single binary feature, that may take a value of either I or o. A j's (j = 1, 2, ... n ) represent
individual strings and A(t) symbolizes the population of ~'s at time t.
For the description of a schema, one may use a wild card ">I<" which may take a value of either
1 or 0 at a given particular position. If one considers the first ten digits of the bit string given
above, the schema H = [ 0 1 1 >I< >I< >I< 00 >I< 1 ] represents the modified bit string A' = [ 0 1 1 1 1
1 0001]. It is apparent that there are more schemata than bit strings. More specifically, there are
3L schemata defined over a binary string oflength L.
There are two important concepts yet to be defined, schema order and its defining length. The
 604

order ofa schema is the number of fixed positions in the template. For example, the order of the
schema H given above, is 0 (II) = 6. The defining length of a schema is the distance between the
first and last specific position (i.e. the defining length of the schema H, d (II) = 10 - 1 = 9 ).
Schemata and their properties provide the means for analyzing the net effect of reproduction and
genetic operators on building blocks of the population [6]. A particular schema grows as the ratio
of the average fitness of the schema to the average fitness of the population increase. Of course,
one can not generate any new (better) schema by copying old strings without any change.
Therefore, at this stage, crossover may come into picture. Crossover is nothing but information
exchange between the newly reproduced strings. Since the number of undesirable strings decrease
during the reproduction process, one may expect some improvement with this exchange. There
is no guarantee that crossover will improve the fitness of the pool of bit strings. Therefore, since
each string is representative of some schemata, the effect of crossover on the schemata has to be
analyzed. For the bit string and schemata given below:
A =[01011101010101001000]
H I = [01011***********1**0]
H2 = [* * * * * * * 0 1 0 1 0 1 * * * 1 * * 0 ]
String A can be represented by either schemata HI or ~. However, in order to mate and
crossover A with another string, HI and ~ has to be disrupted in a different fashion. For a
uniformly distributed random number between 1 and 19, the probability of destruction for schema

Pd = d (HI) fL-l = 19/19 = 1.0

whereas, same probability for the schema H2 is
Pd = d (H2) IL-l = 12/19 = 0.6315
This example can be generalized as, "the schema with a shorter defining length has a higher
probability to survive". Then, for any schema If, the probability of survival in the next generation
is given as
Ps =1- Pd
and the expected number of schema in the next generation becomes
n (t+ 1) = (n(t)*r/a)*Ps.
So far, the crossover rate (Pc) did not come into the picture. When Pc= 1, all strings are mated
and the crossover operation is performed. The crossover rate controls the frequency with which
the crossover operator is applied. For each new population, crossover is applied on Pc*N strings.
 605

As the crossover rate increase, the rate at which new individuals are introduced into the population
increases. Crossover rate detennines the exploration rate. While higher crossover rates may
disrupt the fittest chromosomes, lower Pc may reduce the improvement rate. Usually, a crossover
rate of 0.6 is considered appropriate for many applications.
Mutation: Mutation is a secondary operator. After selection, each bit position of each string
in the new population is applied a random change with a probability equal to the mutation rate Pm'
Consequently, approximately Pm "'N"'L mutations occur at each generation, where N is the
number of strings and L is the number of bits at each string. Mutation operator helps to recover
lost strings that had high-performance. However, high mutation rate causes an increased level of
randomness. For Pm=.5, GA becomes a random search technique. With A = [0 1 0 1 1 1 0 1 0
1 0 1 0 1 001000], if mutation is performed on the third gene of the chromosome, then one
gets A=[ 0 1 1 1 1 1 0 1 0 1 0 1 0 1 0 0 1 0 0 0 ].
Performance Measures: In an effort to quantify the effectiveness of GAs, DeJong devised
two measures, one to monitor convergence characteristics and the other for evaluation of ongoing
performance [7]. These measures are called off-line (convergence) and on-line (ongoing)
performance, respectively. The off-line performance is a running average of the best performance
values over a particular time period and is given as

• l~T.
Xe (s) =- L.Jl fe (t),
T

The on-line performance is an average of all function evaluations up to the current trial and
is given as

X(S)=-
e
lET
T
If(t).
e

Genetic Programming for Batch Process Systems Engineering

The application of Genetic Algorithms to Batch Process Systems Engineering are fairly recent and
the authors are aware of only a set of internal reports [29,30]. Simulated annealing, which is often
considered an alternative to GA, received more attention for both design [26] and scheduling [5]
of Batch Processes. The following example illustrates the use of GA to solve a scheduling problem
is included primarily for its pedagogical value.
 606

Application 1. A Scheduling Problem [20]:

This is a 4-Product 3-Stage Flowshop problem which will be used to illustrate the genetic
algorithm for a case that involves discrete variables. The processing times for the problem are
given in the following table and is followed by its genetic model:

Stage Product A ProductB Product C ProductD

I 2 4 5 6

2 4 4 2 4

3 6 4 5 2

Chromosome Syntax: Each product is represented with a gene where alleles are the product
types (A,B,C,D).
Chromosome Interpretation: The locus of each gene corresponds to the order of that
specific product. The part corresponding to the most left gene is processed first.
Chromosome Evaluation: Population contains four chromosomes. Makespan time is used
to evaluate the fitness of each chromosome. Selections are made through roulette-wheel.
Operators: Order-based crossover is used on all chromosomes (P c=1.0). Order-based
mutation probability (Pm) is 0.03. Chromosomes are initialized randomly (Generation # 0).

Population # Population Fitness (makespan) Actual Count

I ABDC 24 2

2 DABC 28 (die oft) 0

3 CBAD 25 1

4 BACD 25 1

On-Line (total) 102 4

Off-line (best) 24
 607

According to actual count values, mating and crossover operators are used on the following
four parent strings:
PI = [A B· D· C], P2 = [ A· B· D C], P3 = [C BAD], P4 = [B A C D]
Initially, PI-P4 and P2-P3 are mated. While applying order-based crossover, a set oflocations
is randomly picked and this sequence is imposed on the other parent. Mutation is not used at this
iteration.
P2-P3 has the following two children C#1 = [A C D B] and C# 2 = [C A B D]
P 1-P4 has the following two children C#3 = [A B D C] and C# 4 = [B A CD]

Population # Population Fitness Actual Count

1 ACDB 26 (die oft) 0

2 CABD 24 2

3 ABDC 24 1

4 BACD 25 1

Total 99 4

On-line fitness = (102+99)/2 =101.5 Total 99

Improvement 3

Off-line 3 Improvement 0
Performance

After reproduction, mating and crossover operators applied between Pl&P3 and P2&P4
P#1 = [c" AB D·] P#2 = [C A· B· D] P#3 = [A B D C] P#4= [B ACD]
the following siblings are obtained
C#1 = [AB CD] C#3 = [ACBD] C#2 = [A CB D] C#4=[ABC D]
Suppose that mutation is applied to the first child of the generation, to explore some
overlooked or lost sequences and to avoid premature convergence. If the location is selected
randomly, then C#l = [A B c" D· ] becomes C#1 = [A B D C] and the following new
generation ( #2 ) is obtained:
 608

Population # Population Fitness

C#1 ABDC 24

C#2 ACBD 23

C#3 ACBD 23

C#4 ABCD 23

Thus, two alternative solutions, [A C D B] and [A BCD], are obtained. The optimum
makespan time is 23.

Criteria Perfonnance Improvement

Total fitness 93 6

On-line 97.5 4

Off-line 23.6 0.6

Best 23 1

An exhaustive search would have required evaluation of 4!=24 sequences to obtain the
optimal, instead of 8 different sequences were evaluated with GA.

Application 2. A Batch Plant Design Problem [1]

The followin~ problem has been widely used in the batch plant design literature for testing the
efficiency of various algorithms. The problem is fonnulated as follows while the design that is
modeled below is depicted in Figure 3 and the parameter values used are given in Table 2. :

O. The Objective Function

1. The batch equipment capacity constraint

 609

Product 2

From
Storage To
Feed Packaging
Pump Reactor Pump Heat
Tank
(Batch) Exchanger Tray Dryer
(Batch)

Figure 3. A Batch Plant

2. The continuous equipment capacity constraint

D·kB.
R ~-'--' keK(j) , jeS
k t .. '
'J

3. The cycle time constraints

L Q.T.
~_ ~!.H
B.
,-1
,

4. The production time constraints

jeS

jeB,

5. The equipment capacity bounds

 610

Table 2. Data for Application 2

Cost Coefficients and Exponents

al 592 b2 250 "I 0.65 P2 0.40

~ 582 b3 210 ~ 0.39 P3 0.62

a3 1,200 b. 250 "3 0.52 P. 0.40

bl 370 bs 200 PI 0.22 Ps 0.85

Batch Equipment Processing Times for each product, potj (hours)

Products 0
Ptj (hours)

1 3 1 4

2 6 - 8

3 2 2 4

Production Requirements (Q) and Size Factors (Sij)

Products Q 1 (lhs/year) Size Factors (Sij )

1 400,000 1.2 1.4 1.0

2 300,000 1.5 - 1.0

3 100,000 1.1 1.2 1.0

Duty Factors (D ik )

Product Duty Factors

1 1.0 1.2 1.2 1.4 1.4

2 1.0 1.5 1.5 - 1.5

3 1.0 1.1 1.1 1.2 1.2

After relaxation of some of the constraints, the revised model reduces to 23 decision
variables and 30 constraints.
 611

The literature on genetic algorithms is composed of either applications in various domains or

analysis of its performance for optimization of simple unconstrained functions with continuous
variables. For these relatively simple test cases, GAs are relatively well understood and accepted
as a search technique. However, GM still needs to be handcrafted for successful deployment of
GPs in real world problems. One practical problem is how to incorporate constraints into the
model.
There are two approaches for incorporation of constraints in the GM. One is to apply a
penalty for the constraints that are violated. In this case, the chromosome corresponding to the
objective function doesn't change; however, a penalty is charged to the fitness value of this
structure. This may, and frequently does, cause the algorithm to slowdown due to resulting
infeasible solutions. If this method is used, initial population should be setup properly. Several
feasible solutions has to be inserted into the initial population to reduce the execution time. This
method is insensitive to how close an infeasible solution is to the feasible region. One advantage
ofthis approach is that it keeps the structure shorter. The shorter the defining length, the more
efficient is the information exchange. However, the population size would be larger than usual in
order to explore legal solutions. In this approach, the time to obtain a legal solution may be
excessive due to the number of constraints, the way of setting the initial population, the size of the
search space, and the penalty parameters.
The other approach that can be adapted is to form a penalty function and include all the
constraints in the structure (chromosome). The basic difference between the two approaches lies
in the fact that the latter tracks the deviation from the feasible region [6]. In this case, as the length
of a chromosome increases, the efficiency of the crossover operator decreases due to the reduction
in quality of the information exchanged. Therefore, one may need to apply multi-point crossover
in lieu of one-point crossover. The parameters that are recommended for each of the approaches
are discussed next.
The generic parameter setting is given as Pc = 0.6, Pm = 0.001 and N = 50. The current
literature on the effect ofpararneters on the search is limited to small size and/or benign problems.
For the problem given above, the following experiments were run:

N Pc Pm Policy # of trials

100 0.4/0.6/0.8 0.00110.03/0.1 Elitist 5000

 612

The following table shows the effect of mutation and crossover rate for Application 2. The
values listed are for the penalty function with hundred units of penalty for each unit deviation from

any constraint.
The values are negative due to the effect of penalty parameters. However, this shouldn't hide
the convergence properties of each-Jlarameter setting. It can be seen that highest mutation rate
produces the slowest convergence. The same behavior could be also seen when the average
characteristics are inspected. This result is consistent with the expected effect of the mutation

Parameters Pc= 0.4 Pc= 0.6 Pc= 0.8 Average

Pm = 0.001 -5.11 -5.14 -5.15 -5.13

Pm =0.03 -5.13 -5.12 -5.11 -5.12

Pm= 0.01 -5.11 -5.0 -4.90 -5.00

Average -5.11 -5.08 -5.05

operator. Principally, it is the crossover operator that makes the GAs different from the random
search techniques. As can be seen from the column labeled average, lower crossover rates result
in faster convergence. This behavior is contrary to the general wisdom. A formal question that
should be answered is why lower rate of information exchange is better? At the extreme, this can
be construed as random search being superior to the GA, for the problem tackled here. The
relation between crossover rate and convergence can be better attributed to the simple GA form
used. This GA was deficient in producing legal (feasible) solutions (unless some legal solutions
are introduced to the initial population). One possibility is to update the crossover operator so that
crossover produces only legal structures-feasible solutions (including initial population). For
constraint models with continuous decision variables, it is easy to apply a crossover method that
generates feasible-legal structures.

For the problem given above, a fitness function that adapts some sort ofiagrangian relaxation
was used. This fitness function is given below:

~=f(x}+~(p(b-g(x}}}2
 613

where p is a penalty parameter that balances the effect of the square power on numbers between
-1 and 1. "Il", unlike lagrange multiplier, is a binary parameter that takes the value of 1 when the
constraint is violated and 0 otherwise. Thus, a penalty proportional to the deviation from the
feasible region is applied to the fitness value of the corresponding chromosome. This approach
appears to be superior to the two simple approaches mentioned before. Using this approach, the
effect of control parameters on the efficiency of genetic algorithms for multi-constraint models
were analyzed. More widely-accepted control parameter settings for GAs were tested with this
design problem. Representative results are shown in Figure 4a, b. As can be seen in Figure 4a.,
the optimal population size is 50, although the sensitivity with respect to this parameter is not that
significant. On the other hand, as can be seen in Figure 4b. the higher crossover rates (e.g. 0.7)
beneficial for this problem. This conclusion conflicts with the earlier observations discussed during
the analysis of interrelation between the mutation operator and crossover rate. The cross
interaction of GM and control parameters is a fruitful area for future research.

Fitness (thousands) Fitness (thousands)

50.-----------------~ 50,-----------------~

40 •
• 40 •
popUlation -40
•
• •
Crossover=O.s
30 30
population ~ so
• •
Oossover-O.7

population - 60

••• •
Crossover-O.6
20 20
•
.. -
III
••
10
•• -:111. 10 •··'1
•••••• • ••
••• • •••
00
•••
2 4 6 8 10 12 °0 2 4 6 8 10 12
Generations Generations

Figure 4 a, b. Convergence Characteristics of Genetic A1goritluns

 614

Knowledge Acquisition and Machine Learning

Background

The development of knowledge based systems for planning and scheduling has been limited by the
knowledge acquisition bottleneck. Formalizing knowledge and implementing within knowledge
bases are major tasks in the construction of large hybrid systems that efficiently integrate
algorithmic and symbolic methods. The hundreds of rules and thousands of facts have to be
obtained through various means shown in Figure 5. Some of the modes of knowledge acquisition
involve human, i.e. "Domain Expert and Knowledge Engineer" while some of the knowledge
acquisition modes utilize machines. The interactive and incremental development of knowledge
based systems allow experimentation and validation of knowledge, structure, and methods at
various levels. The stages and type of information involved in acquisition of knowledge are
summarized in Figure 6.

Knowledge -------j.~ Expert System

Mechanism

Expert • Intelligent Editing

Program
Inference Engine
Data • Induction Program (Solution Knowledge)

Textbook • Text Understanding Knowledge Base

(Domain Knowledge)
Program

Expert • Knowledge Engineer

Figure 5. Mechanisms of Knowledge Acquisition

Advances for automation of the knowledge acquisition process would constitute a major
 615

Reformulations

Redesigns

Refinements

Design structure
+
Identify A Find Concepts B C Fonnullie Rul.. D Validate Rule.
Problem f---t-- to Represent -.. to Orpnize f----.- to Embody f---. that Orpnize
Cham:teristict KnowIedse KnowIedse Knowledse Knowledse

Identification Conceptualization Formalization ImplemenIation Testing

Legend
A.- Requirements B.- Concepts
C.- Structure D.- Rules

Figure 6. Stages of Knowledge Acquisition

contribution to AI technology and such techniques are within the domain of machine learning.
There is no generally accepted categorical definition of learning and maybe it is best to give a
prototype instead. A learning system can be modified by itself to improve its performance over
an extended period of time. A machine learning program will allow more efficient execution of
computer software, in comparison to ones with no machine learning component.
The literature on machine learning is vast and rapidly expanding. However, these techniques
are conveniently summarized within the two volume handbook on machine learning [21]. A
taxonomy of various symbolic machine learning topics is shown in Figure 7, where the paths of
the topics discussed here are highlighted. One of the earliest approaches to learning is called Rote
Leorning. In rote learning, computed solutions are stored along with the corresponding attributes
that contribute. It is quite effective whenever computation is more expensive than recall. Learning
element in computer programs that played games such as chess utilized rote learning. Today,
Exp1anotion Based Leorning systems seem to capture the spotlight. They attempts to learn from
a single example by explaining why it is an example of the target concept. Genetic Algorithms
discussed earlier would be useful in learning through observation and discovery.
 616

LEARNING STRATEGIES

Rote Learning by Learning by Learning by earning b'

Learning instructIon Analogy Induction deduction

~
progra- ~elI)o­
mrrung nzatIon

Learning from Learning from observat'

examples and discovery

p
c

examples examples
source type one trial iJ

teacher external
environment
~
system
itself
pOSItIve

Figure 7. Classification ofLeaming Mechanisms

Symbolic approaches to learning, connectionist learning approaches, and genetic programming

all offer significantly different points of view, each with strong and weak points. Genetic
Algorithms were discussed earlier, connectionist approaches and their comparison with statistical
methods are readily available in literature including the papers presented at this NATO ASI [28,
 617

36]. Our focus here will be to give a hybrid approach that combines the symbolic and connectionist
approaches to machine learning.

A Hybrid Approach to Machine Learning

The knowledge extraction and background knowledge encoding problems associated with
connectionist learning algorithms and inability of the symbolic learning algorithms to handle with
continuous variables effectively, and allow parallel knowledge representation have prompted the
researchers in AI to seek a hybrid approach. One development in this direction is Symbolic
Connectionist network (SC-net) [32]. The SC-net, designed specifically for constructing Expert
Systems, is based on a hybrid of Symbolic and Connectionist architectures. The system allows
extraction of knowledge in the form of rules and can handle both scalar and fuzzy variables. The
following features of SC-net make it a connectionist method
1) A highly parallel and uniform representation of knowledge
2) Fault tolerance and noise resistance.
3) A built-in ability to deal with non-crisp inputs and outputs.
and the following features make it a symbolic method
1) The ability to encode rules to support knowledge refinement.
2) Allow for rule extraction as a direct means to elicit learned knowledge and support the
implementation of Expert System standards such as consultation and explanation facilities.
3) Provides means to represent symbolic constructs such as variables, comparators and
quantifiers. This leads to a more powerful language for describtion of knowledge.
The network topology in SC-net is based on the training examples unlike the user specified
ones in connectionist approaches, promising an optimal representation. Fuzzy logic is used to deal
with uncertainty. The next section describes the different learning stages involved in SC-net with
enough emphasis on each learning component.

Learning Mechanism

Different learning steps involved in SC-net are shown in Figure 8. Each time SC-net is given a
training set that consists of input and corresponding output information, it invokes a call to the
Recruitment of Cells Algorithm (RCA). The main function of RCA is to map the training set into
 618

RCA I-----~ Redundancy / - - - . . , PSA

?

SimplifY Yes
Network
?

GAC

No
Generalize
?

Figure 8. Learning with SC·net

a network representation for further processing through the other algorithms.

The network generated consists of Input Cells (IC), Infonnation Collector Cells (ICC),
Negative Collector cells (NC), Positive Collector cells (PC), UnKnown (UK) cells, and Output
Cells (OC). The input cells represent the attributes ofthe training set, infonnation collector cells
transform the input infonnation into an intennediate form, and the NC and PC cells collect
negative and positive evidence present in the training set for a conclusion. The NC and PC cells
are connected to all output cells and IC cells. The UK cells act as threshold for the NC and PC
cells, enabling them to propagate. The activation is based on the type of evidence (positive,
negative, unknown). The output cells represent the output classes.
In the SC-net, the network structure consists of cells modelling the min and max operators

offuzzy logic. Every cell contains a bias value CB j, which indicates the type offuzzy function the

cell models and its value lies between -1 and +1. The min cells have a bias value of -1 and the max

cells have 1. The absolute value ofCB j represents an upper threshold on the cell activation, CA;.
These cells are connected by links which carry weights. The link weights represent the degree of
 619

contribution the lower cell makes for the upper cell's concept. If cell Ci ( with CA; ) and cell Cj (

with CA j ) are connected then the weight of the connecting link (CWi) is finite, otherwise (i.e.,
not connected) the corresponding weight is zero. The new cell activation CA'i is computed
according to the formula given below.

CB i * min { C~ * CWij } if Ci is a min cell.

CB i * max { C~* CWij } if Ci is a max cell.

CA;' = CA; positive + CA; negative - Y2 if Ci is a final output cell.

1- (C~* CWij) if Ci is a negate cell.

Here, a cell is a min cell ifthere is negative evidence, a max cell ifthere is a positive evidence,
and negate cell if there is no evidence.
The final output of cell Cj is always bounded by [0,1] range using the following relationship,
0i = min { 1, max { 0, CA;'} }
Every training instance is presented to the network as a single feed forward pass. After the
pass has been completed, the actual and the expected activation for each output are compared.
Three possible conditions may result from this comparison:
1) If the example was correctly identified (error is below some epsilon, € < 0.02), then no
modifications are made to the network.
2) If the example was similar to at least one previously seen and stored instance (error within
5€), the bias of these cells are adjusted to incorporate the new instance.
3) If the example was not identified by the network, then this results in the recruitment of a
new cell or information collector cell (ICC). Appropriate connections from the network
to the ICC are created. The ICC cell itself is connected to either positive collector (PC)
or negative collector (NC) cell. The PC is used to collect positive evidence, whereas the
NC accumulates the negative evidence. The unknown cell always propagates an activation
of 0.5, guaranteeing the PC cell only attains activation greater than or equal to 0.5,
whereas the NC cell propagates activations ofless than or equal to 0.5.
The network generated by the aforementioned procedure is next processed through the Pre-
Selection of Attributes (PSA) algorithm that introduces redundancy into the network. Since the
RCA provides a minimized representation of the input domain, certain input features may be
incorrectly identified as important. This is especially true when the number of training cases is
 620

insufficient. PSA algorithm addresses this problem by selecting all the input attributes that
represents a concept and was found to be quite effective in pattern-matching.
The network topology of SC-net is set based on the training examples, unlike the user
specified ones in other connectionist approaches. The cells are recruited at a single level and the
cell growth proceeds horizontally, suitable for paralelization. One of the major problems with the
RCA learning component is the linear growth of the network, under worst case conditions, each
example may result in recruitment of one to three new cells. This problem may be rectified through
injection of background knowledge (i.e. if then rules) in the training set. In doing so, the learning
system makes use of fuzzy logic language hedges (low, medium, high), quantifiers (never,
sometimes, always), and operators (max, min). As a consequence, one ends up with a natural
language like interface. The network growth is further controlled, independently, through the
global attribute covering (GAC) algorithm.
Given the RCA (with/without PSA) generated network as its input, GAC generates (except
for contradictions and inconsistencies) an equivalent network, minimized both in the number of
cells and links. As a side effect, extraction of highly general and simple rules are possible for
explanation purposes. At the first step of GAC, all links to information collector (IC) cells are
disconnected, CWij = O. This forces IC cells to propagate an activation of 1 (firing state),
regardless of the inputs presented to the network, during the feed forward pass. Since all IC cells
fire, they will also activate the corresponding output cells they are connected to. All IC cells that
are incorrectly activated are identified and form an conflict list. Furthermore, all links entering the
IC cell as inputs will be considered as potential inhibitors for network minimization. By
reconnecting specific links, IC cells can be prevented from firing for the same inputs. An
"undesirability" measure associated with each link is used to prune the network.
The GAC generated network can be further generalized through use of the GEneralization of
NETwork (GENET) learning component. In symbolic terms, generalization is accomplished
through replacing single conditions by a disjunction of similar conditions, allowing extraction of
rules that cover a larger domain. This can be hard coded into the network, through use of domain
specific knowledge.
The process of generalization involves grouping of the set oflinks entering an IC cell that
have a common parent cell. Each of these possible groups (containing more than one link) are first
represented in the form of a max-cell, and the links within the group form the input connections
to this cell. Here, there is a possibility of over generalization i.e., it is possible for output cells to
 621

fire even when they should not have fired. Once again, this is controlled by the "undesirability"
measure which is applied to the conflict vector. After each pass through this conflict vector, the
link with the highest undesirability index is selected and removed from the max-cell. This process
is continued until the conflict vector set is empty and the resulting network is more general. The
generalized network may be able to handle an input pattern that does not exactly match the one
presented for training. The SC-net system can handle crisp variables as well as fuzzy quantifiers.
Figure 9 shows a sample program that includes fuzzy quantifiers and operators. This
representation helps in dealing with uncertain/fuzzy domains and for generating rules which are
similar to that of a natural language.

Range 0.00- 0.05 # completelyJalse #

0.05- 0.20 # kind of false #
0.20- 0.40 # sometimes false #
0.40- 0.60 # unknown #
0.60- O.SO # kind_oCtrue #
O.SO- 0.95 # mostly-true #
0.95- 1.00 # completely-true #

Quantifiers a1mostall: 0.90-1.0

always: 1.00-1.0 1* same as always true *1
sometimes 0.70-0.9
never 0.00-0.0 1* same as not *1

Operators almost := they almostall are completely-true

incorrect := they always are completelyJalse
hardlyJalse := they are never completelyJalse

Rules if almost (ul,u2,u3,u4,u5) then u7 (I)

if incorrect (ul,u2,u3,u4,u5) then uS (I)
if never (greater (ul,O.S» then u9 (0)
ifhardlyJalse (ul,u2,u3,u4,u5) then ulO (I)
if and (sometimes (ul), never (u2), always (u3» then u7 (I)
Figure 9. Sample Program Containing Quantifiers and QuantifYing Operators

Another powerful feature of SC-net is its ability to incorporate background knowledge.

Background knowledge could be provided before, during and after the learning process. This
background knowledge could be domain specific and/or meta-knowledge. This feature is
particularly helpful in cases where there are infrequent examples that may never be covered by
the learning algorithms.
 622

Before the learning component of SC-net is invoked, it is possible to encode a set of rules into
the network. This can affect both the network growth, the training time, and the number of rules
generated. The background knowledge base can be constructed either through constructing a
prototype or via the Life-Cycle approach. In the first method, a set of rules may be used as
background knowledge along with training examples. In the second approach, the rules generated
from examples are modified by the domain expert(s), whenever needed and maintained in the
learning system The latter approach is illustrated in Figure 10.

Figure 10. Life Cycle Approach to Knowledge Refmement

During the learning phase, the background knowledge can be utilized through GAC algorithm
to select the best groups oflinks. This background knowledge can enter SC-net in three ways:
1) Unknown value heuristic: This allows the specification of attribute values, which are
considered to represent concepts such as doesn't apply, unknown, etc. This feature is
helpful when a domain exhibits a strong degree of ignorance i.e., where many of the
features are deemed not important or redundant.
2) Complementary value heuristic: In cases where the attributes contain complementary
values such as (high,low), (good,bad), (strong,weak) etc., this information can be used to
direct the GAC algorithm by temporarily discarding groups of links in that emanate from
cells representing complementary value heuristic.
3) Important attribute heuristic: Here, the background knowledge is entered in the form of
 623

description. They specify, for each output concept (class), only the attribute of importance.
This helps to reduce the search space learning algorithm operates on.
The GENET algorithm allow encoding background knowledge after learning as well. This
process can also be done through injecting the meta-knowledge after each learning cycle, resulting
in informed learning.
In the next section, an example that illustrates SC-net's learning ability is presented.

Example: Design of Complex Column Sequences

The design of complex distillation column sequences has been of interest to chemical engineers.
We used the method proposed by Tedder & Rudd [34] to test the capability of SC-net. The
optirna1/best column sequencing is based on separation and cost efficiency as given by Tedder and
Rudd. The sequences used are shown in Figure 11 and are abbreviated as
° 1 - Direct sequence O2 - Indirect sequence 03 - Side stream rectifier

04 - Side stream stripper Os - Complex design 06 - Side stream columns

07 - Side stream columns

For training purposes, the SC-net is presented with a set examples covering as many different
permutations of the selected attributes as possible. The training set used for this example is given
in Table 3. Only 15 training cases were used for learning. The input variables in the training vector
are very general and represented using fuzzy numbers. This fuzzy representation helps us to
represent a region (pi-shaped) as opposed to a single point. The following notation is used for
input variables:
ESI - Ease of Separation Index
MP - Middle Product
OR - Over Reads
BP - Bottom Product
OHBP -lOR-BPI indicates the how close overheads and bottoms are
Design - The type of Column Sequence.
These fuzzy variables are defined in the following range
It n - less than n, any number in [o,n-e]
ge_n - greater than or equal to n, any number in [n, 100]
 624

A B A

Dl c D2c
A B A

D3 c c
A
A.B

DS

A
c

D6 D7
c

Figure 11. Distillation Based Separation Schemes for Ternary Mixtures

b_nl_n2 - between nl and n2, any number in [nl,n2]

any - any number [0,100].
"n", "nl", and "n2" may have any value in the region R = [0,100] and e is a small real number
0(0.01). The output class has values Dl, D2, DJ, D5, D6, D7 which correspond to the type of
 625

Table 3. Training Set for the Complex Column Sequencing Example

ESI MP OH BP OHBP Class

It 1.6 b 40 90 any any It 10 D5

ge 1.6 j;(e 50 any b 5 20 any D5

It 1.6 j;(e 50 any It 5 any D6

It 1.6 It 50 any It 5 any Dl
j;(e 1.6 ge 50 any It 5 any D6
j;(e 1.6 It 50 any It 5 any D3
It 1.6 ge 50 It 5 any any D7
It 1.6 It 50 It 5 any any D2
j;(e 1.6 j;(e 50 It 5 any any D7
ge 1.6 It 50 It 5 any any D2
It 1.6 It 50 It 50 ge 50 any D2
ge 1.6 It 50 It 50 ge 50 any D2
It 1.6 It 50 ge 50 It 50 any Dl
It 1.6 It 50 any any It 10 D3
ge 1.6 any ge 50 It 50 any D3

column sequencing. Figure 12 shows the network generated by SC-net for this example, the input
attributes ESI, MP, OH, BP and OHBP are in the input cell layer and the interconnection network
consists of the network created for defining each of the attributes values. The output cells from
the interconnection network will be the input attributes with each of its values presented in the
example case, only few of them are shown in the figure. These cells are linked to the IC cells
depending on the example cases presented and the IC cells connect to the PCINC cells depending
on the type of evidence available (positivelNegative) in examples and the outer layer has all the
outputs presented in the examples.
The knowledge extracted from the SC-net is shown in Table 4. In terms of simplicity and
accuracy, these rules are very similar to the ones proposed by Tedder and Rudd[34] .
 01 02 03 05 06 07

Outout
Cell
Layer

ICC
Layer ())
I\)
())

8
INTERCONNECfION NETWORK

Input
Cell
Layer

Figure 12. The Final SC-net Generated for the Complex Colwnn System Selection
 627

Table 4. Rules Generated by SC-net

/f and (ES! [ItJ6), MP[lt_50}, BP[lt_5}) then Dl (1.00)

/fand (ES! [ItJ6), MP[lt_50}, OH [ge_50}, BP[It_50}) thenDl (1.00)
/fand (MP[It_50), OH [1t_5}) thenD2 (1.00)
Ijand(MP[It_50J, OH [1t_50J, BP [ge_50J) thenD2 (1.00)
Ijand (ES! [geJ6J, MP[It_50J, OH [ge_50J, BP[lt_5J) then D3 (1.00)
Ijand (ES! [ltJ6J, MP[II_50J, OH [ge_50J) thenD3 (1.00)
Ijand (ES! [geJ6J, OH [ge_50J) IhenD3 (1.00)
Ijand(ES! [ItJ6},MP{b_40jOJ, OHBP [ltjOJ) thenD5 (1.00)
Ij and (ES! [ge J 6J, MP[ge_50], BP [b _5_20J) then D5 (1.00)
Ijand(MP[ge_50}, BP [1t_5}) thenD6 (1.00)
Ij and (MP[ge _50J, OH [11_5J) then D7 (1.00)

After the training phase was over, the SC-net was given some new cases. The new cases used
are shown in Table 5. The classification accuracy is very encouraging as it is able to identify all
the classes correctly. Due to the fuzzy nature of the input attributes, in some cases, more than one
design was identified as optimal. This feature is particularly desirable in uncertain domains where
there are large number of attributes with interdependency and the attributes are best represented
in a qualitative fashion.

Conclusion

The objective of this paper was to provide an overvIew of various approaches to

planning/scheduling problems and knowledge acquisition/machine learning. The evaluation of
these techniques within the context of batch processing systems engineering will naturally take
several years. However, in each topical area, a more promising and new approach is discussed in
more depth with motivating examples. Thus, the conclusions of this chapter will be limited to these
cases.
 628

Table 5. Comparison of Sc·net Performance with Simulation

ESI MP OR BP ORBP . Expected SC-net

1.3 65 17 18 1 D5 D5
1.8 70 15 15 0 D5 D5
1.3 80 18 2 16 D6 D6
1.3 20 78 2 76 D1 Dl,D6
1.8 65 31 4 27 D6 D6
1.8 20 78 2 76 D3 D3,D6
1.3 65 4 31 27 D7 D7
1.3 20 2 78 76 D2 D2,D7
1.8 65 4 31 27 D7 D7
1.8 20 2 78 76 D2 D2,D7
1.3 25 20 55 35 D2 D2
1.8 25 20 55 35 D2 D2
1.3 25 55 20 35 D1 Dl
1.3 10 45 45 0 D3 D3
1.8 15 60 25 35 D3 D3

Genetic programming is a robust search technique that has been applied effectively in several
different solution spaces. This robustness of Genetic Programming could be exploited in
interfacing distinct decision modules of a batch production decision environment. The promising
performance ofGP is expected to increase its application in NP-hard combinatorial optimization
problems such as scheduling. The examples shown illustrate this potential as well as significance
of the modelling element. One of the strong elements of Genetic Programming is its robustness
in allowing the user great flexibility to bring insight both in modeling and search.
A hybrid learning mechanism which has the features of both symbolic and connectionist
approaches has been analyzed and its applicability for a chemical engineering design problem is
illustrated. The ability of learning system to extract knowledge in the form of expert/natural
language type rules makes it a very good choice for building expert systems. Its performance
accuracy, despite generalization of the network during the learning phase, makes it a valid choice
for overcoming the knowledge acquisition problems. Due to the background knowledge encoding
 629

feature, it can be used to generate rules in domains where there is a lack of expert knowledge and
systems which require some a priori knowledge.

References

I. Biegler,1.T., I.E. Grossmann and G.V. Reklaitis: Application of Operations Research Techniques in Chemical
Engineering. In: Engineering Design, Better Results through Operations Research Methods (R.Levary, ed.),
Elsevier Science Publishing Co., Inc. 1988
2. Booker,L.: Improving Search in Genetic Algorithms. In: Genetic Algorithms and Simulated Annealing ( 1. Davis,
ed.) Morgan Kaufinan Publishers, Inc., Los Altos, CA, 1987
3. Brooks, RA.: A Robust Layered Control System for a Mobile Robot. IEEE Journal of Robotics and Automation
RA-2 I (1986)
4. Cukierman, D., R Owans, and S. Sioseris: Interactivity Activity Scheduling with Object Oriented Constraint Logic
Programming. In: Application of Artificial Intelligence in Engineering VIII. Vol:II Computational Mechanics
PublicationslElsevier Applied Science. (G. Rzevski et al. ed.), 1993
5. Das, H., P.T. Cummings, and M.D. leVan: Scheduling of Serial Multiproduct Batch Processes via Simulated
Annealing, Computers in Chemical Engineering, 14 (12), 1351-1362, (1990)
6. Davis, L: Handbook of Genetic Algorithms. NY: Van Nostrand Reinhold, 1991
7. DeJong,K.A.: An Analysis of the Behavior of a Class of Genetic Adaptive Systems. Doctoral dissertation,
University of Michigan, 1975
8. Famili, A, D. Nau, and S. Kim: Artificial Intelligence Applications in Manufacturing, AAAI PresslMIT Press,
1992
9. Fuernsinn, M.,and G. Meyer: The Configuration of Automobile Manufacturing Plants Using FAKIR. In:
Application of Artificial Intelligence in Engineering VIII. Vol: I Computational Mechanics PublicationslElsevier
Applied Science. (G. Rzevski et al. ed.), 1993
10. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning. Reading, MA: Addison-
Wesley 1989
II. Grefenstette, J.J., LDavis, and D. Cerys: GENESIS & OOGA - Two Genetic Algorithm Systems, Melrose MA:
TSP 1991
12. Hammond, K., Chef: A Model of Case-Based Planning. In: Proceedings of AAAI-86 (1986)
13. Hillard, MR, G.E. Liepens, M Palmer, M Morrow, and J. Richardson: A Classifier Based System for Discovering
Scheduling Heuristics In: Genetic Algorithms and Their Applications. Proceedings of the Second International
Conference on Gnetic Algorithms, 231-235 (1987)
14. Holland,J.H., :Adaptation in Natural and Artificial Systems, The University ofMichigan:Ann Arbor 1975
IS. Kapanoglu, M., Genetic Intelligence in Scheduling of Modular Manufacturing Systems, PhD dissertation (in
preparation), U.of South Florida, 1995
16. Kempf, K. C. LePape, S. Smith, and B. Fox: Issues in The Design of AI-Based Schedulers: A Workshop Report.
AI Magazine II (5) (1991)
17. Lambrou, S.K., and AJ. Dentsoras: A Machine Based System for Valuation and Consultation on Machine
Assemblies during Design for Configuration. In: Application of Artificial Intelligence in Engineering VIII. Vol: I
Computational Mechanics PublicationslElsevier Applied Science. (G. Rzevski et al. ed.) 1993
18. Laszlo, H, M, Hoffineister, D.W.T. Rippin: Gantt: an Interactive tool, This volume, p. 706
19. Liepins, G.E., M.R Hilliard, J. Richardson, and M.Palmer: Genetic Algorithms Applications to Set Covering and
Traveling Salesman Problems. In: Operations Research and Artificial Intelligence: The Integration of Problem
Solving Strategies (1990)
20. Mah,R.S.: Chemical Process Structures and Information Flows, Boston: Butterworths 1990
21. Michalski, J., G. Carbonell, and T.M. Mitchell.: Machine Learning, Volumes I and II, Los Altos: Morgan Kaufinan
1986
22. Minton, S. Machine Learning Methods for Planning, Morgan Kaufinann, 1993
23. Mitchell, T.M.: Becoming Increasingly Reactive. In: Proceedings AAAI-90
24. Nilsson, N.J.: Principles of Artificial Intelligence Palo Alto: Morgan Kaufinann 1980
25. Noronha, S.J., and V.V.S. Sarma: Knowledge Based Approaches for Scheduling Problems: A Survey, IEEE
Transactions on Knowledge and Data Engineering, 3, 160-171 (1991)
26. Patel, AN., RS.H. Mah, and I.A. Karimi. Preliminary Design of Multi-product Noncontinuous Plants Using
 630

Simulated Annealing, Computers and Chemical Engineering, IS, 451-469 (1991)

27. Pearl, Judea: Hewistics. Addison-Wesley, 1985
28. Raich, A., X. Wu, and A. Cinar: Comparison of Neural Networks and Nonlinear Time Series Techniques for
Dynamic Modeling of Chemical Processes. This volume, p. 309
29. Realff, M,: Learning Localized Hewistics in Batch Scheduling. Internal report LISPE 88-053, MIT, 1989
30. Realff, M,: An Analysis of the Chemical Batch Production Problem and a Detailed Methodology for Scheduling.
Internal report LISPE 88-054, MIT, 1989
31. Rich, E and K Knight: Artificial Intelligence, McGraw-Hill, 1991
32. Romaniuk, S.G. Extracting Knowledge from a hybrid Symbolic, Connectionist network. Ph.D thesis, Univ. of South
Florida 1991
33. Smith, P., E. Fletcher, F. Gronskov,1. Malmgren-Hansen, KHo, B. Gaze: CIM.REFLEX- Using Expert System
Technology to Achieve Real Time Shop Floor Scheduling. In: Application of Artificial Intelligence in Engineering
VIII. Vol:I! Computational Mechanics PublicationsIElsevier Applied Science. (G. Rzevski et al. ed.) 1993
34. Tedder, D.W., Rudd, D.F. : Parametric Studies in Industrial Distillation. AICHE J. 24, No.2, (1978)
35. Tsukiyama, M, K Mori, and T. Fukuda: Strategic Level Interactive Scheduling and Operational Level Real-Time
Scheduling for Flexible Manufacturing Systems. In: Application of Artificial Intelligence in Engineering VIII.
Vol:II Computational Mechanics PublicationslElsevier Applied Science. (G. Rzevski et al. ed.) 1993
36. Venkatasubramanian, V.: Fault Diagnosis Through Neural Networks. This volume, p. 631
Integrating Unsupervised and Supervised Learning
in Neural Networks for Fault Diagnosis

Venkat Venkatasubramanian*, Surya N. Kavuri

Laboratory for Intelligent Process Systems, School of Chemical Engineering, Purdue University,

West Lafayette, IN 47907, USA

Abstract: Recently, there has been considerable interest in the use of neural networks for fault
diagnosis applications. To overcome the main limitations of the neural networks approach,
improvements are sought mainly in two respects: (a) a better understanding of the nature of
decision boundaries (b) determining the network structure without the usual arbitrary trial and
error schemes. In this perspective, we have compared different neural network paradigms and
developed an appropriate integrated approach. A feedforward network with ellipsoidal units
has been shown to be superior to other architectures. Two different types of learning strategies
are compared for training neural networks : unsupervised and supervised learning. Their
relative merits and demerits are discussed and a combination has been proposed to develop a
network that meets our diagnosis requirements. Unsupervised learning component serves to
identify the features and establish the network structure. Supervised learning serves to fine-
tune the resulting network. We present results from a reactor-distillation column case study to
demonstrate the structure of the measurement pattern distribution and the suitability of
ellipsoidal units approach. By considering the transient behavior in the diagnosis framework,
we point out that the problem of fault diagnosis can be treated on the same footing for both
batch and continuous processes.

Keywords: Fault Diagnosis, Neural Networks, Artificial Intelligence, Learning

1. Introduction

Managing process operations requires that the process situation be properly identified. This
requires the ability to answer [14] : What is the current state of the plant?

* Author to whom all correspondence should be addressed.

 632

Before we proceed further, it is useful to illustrate the problem of fault diagnosis in a

pictorial manner that brings out the essential characteristics of this problem. Figure 1 shows the
schematic of a typical distribution of the fault classes in the space of sensor measurements.

1.0

Nonnal behavior
0.8

~ O,S
Abnonnal behavior
i:!
& Faults known
E
~ 0.4

0.2 D Abnonnal behavior

Fault unknown

0.0
0 .0 0.2 0.4 O.S 0 .8 1.0
Concentration

Figure 1. Distribution of fault classes in the measurement space

Examples of such distributions in the space of two significant sensors may be found in [20,
18]. From the day-to-day operations of a process plant, sufficient patterns can be acquired for
the normal behavior which is therefore well defined. In Figure 1, the central region represents
the region where the measurements corresponding to the normal process operation fall. When
a pattern falls in the normal region, the fault classifier should be able to identify that the process
behavior is normal. Otherwise, the process behavior is abnormal. For the abnormal behavior,
only a few patterns may be available either from simulations or process history, covering only
portions of the abnormal region. These regions define the known faults (causes of abnormality)
as shown in Figure l. In the case of abnormal behavior, the classifier should identify whether
the behavior falls under one of the known fault classes and if so identify the fault class. If the
behavior has a novel symptom pattern, the classifier should accordingly identify the behavior as
abnormal while indicating that its cause is unknown. Fault classifier should be able to do this
using only the examples of known classes.
 633

This question is not easy to answer because of the following reasons:

• the measurements may be insufficient
• the measurements may be unreliable - sensor biases or failures
• process unit malfunctions - unnoticed leakages or blockages
• process unit degradation - catalyst deactivation, heat exchanger fouling
• large unknown external disturbances - Increase in the coolant temperature
Data reconciliation techniques traditionally serve to rectify data or measurement
inconsistencies by using the redundancy provided by the process model. If the process model
is in itself of doubt for its validity, this framework needs to be extended. In this perspective,
fault diagnosis is a complementary subdiscipline of data reconciliation where the goal is to
determine what is wrong without assuming either the process structure (model) or the
parameters are correctly known. In the general context of fault diagnosis, the definition of a
fault includes :
• gross parameter changes in a model (e.g.change in the feed concentration)
• structural changes (change in the model itself, e.g. failure of a controller, valve
stuck, pipe broken etc.)
• measurements are grossly corrupted (e.g. sensor biases, failure, saturation etc.)
Feedforward neural networks have been an attractive alternative for the design of classifiers
for fault diagnosis. The interest in them stems mainly from the fact that they provide a means to
parameterize the classifier without a specification of the functional form of the discriminant
function. This has an advantage over other schemes such as linear and quadratic classifiers [5],
used commonly in statistical pattern recognition, where a commitment is made a priori to a
particular functional form. Neural networks have been successfully demonstrated in their
ability to approximate arbitrary functions [3]. This generality also comes with a lack of control
and sometimes undesirable effects such as extrapolation problems. Application of neural
networks to· fault diagnosis has been a very active area of research in recent years [7, 19, 17,
15, 12]. Attempts were made to characterize their classification and generalization performance
on one hand, and the speed of their training on the other. Use of networks using ellipsoidal
units [9, 10] have been proposed as a viable alternative to the standard backpropagation
network on two accounts. First, the standard backpropagation network uses linear activation
function that classifies regions by generating intersecting hyperplanes. These hyperplanes of
infinite extent have been noticed to cause (a) erroneous extrapolation (b) place decision
boundaries close to the class boundaries (c) sometimes generate unintuitive classification of
some regions of the input space. Second, the training time of backpropagation networks can be
very high for large networks. This requires the use of decomposition techniques that help
speed up the training. Networks with ellipsoidal units permit the development of
decomposition techniques which result in the training of small networks with few training
patterns.
 634

2. Fault Diagnosis of Continuous and Batch Processes

Traditionally, fault diagnosis of continuous processes has been concerned with steady state
models and the deviation of the process from the nominal steady state. The situation with batch
processes is different in that the process is in unsteady state. Detection and diagnosis of faults
in batch processes has to be done using dynamic models instead of steady state models and
dynamic trends instead of steady state values. These distinctions would, however, blur when
one considers the transient dynamics resulting from a fault in a continuous process as well. In
such a general context, the problem of diagnosis becomes one of detecting and isolating a fault
from the time trajectories of measurements for the different faults. The problem of diagnosis is
to partition the state space into various fault classes and the normal region. Partitioning may be
performed by the generation of envelopes that enclose possible trajectories for each of the fault
classes. Envelopes so generated serve to isolate each trajectory to a possible fault class. Figure
2 shows the state-space trajectories of a process transient for two different faults.

Figure 2. Transient state-space trajectories of two different faults

 635

The classifier approximating the different fault distributions in the state-space should
develop envelopes around the region where the sample trajectories for that fault exist. When
trying to classify a trajectory to a particular fault class, two checks have to be made: (a) nearness
of the trajectory to envelopes; (b) similarity in the dominant directions in which the trajectory
and the envelopes spread in measurement space. Classifiers which use Euclidean distances can
identify whether the trajectory is close to a given envelope. However, modeling the direction in
which the envelope spreads is not possible with the use of radially symmetrical measure of
distance such as the Euclidean distance. Use of an asymmetric distance metric is warranted for
such situations. The required asymmetry is brought into the network by the ellipsoidal
activation as will be discussed next.

3. Linear Versus Ellipsoidal Activation

The standard backpropagation network uses a linear activation function, a sigmoid squashing
function and RMS error minimization. The ellipsoidal activation function [9] E(x) is the
following:
E(x) = - (x-m)T(DTD)-I(x-m) + I (I)
E(x) = 0 defines an n-dimensional ellipsoid oriented along the reference axes, where m is the
center of the ellipsoid and D is the diagonal matrix of the half-lengths of the principal axes.
Activation E(x) is incorporated in the squashing function to determine the bounded network
value. Since the squashing function, a sigmoid, is a monotonic function, the node value would
also have ellipsoidal contours of constant value. The node value for a sigmoid with the range (0
I) is given by:
(2)
l+exp(-E(x))
When x is on the ellipsoid, E(x) = 0, and the node value is 0.5.
When x is inside the ellipsoid, E(x) > 0 and the node value is in the range (0.5 1). As
x moves away from the ellipsoid towards the center, the node value tends closer to 1.
When x is outside the ellipsoid, E(x) < 0 and the node value is in the range (0 0.5).
As x moves away from the ellipsoid, the node value tends closer to zero.
Figure 3, schematically shows the difference between linear and ellipsoidal activation
networks. In networks using linear activation functions, decision regions are formed by the
intersection of hyperplanes generated at the hidden nodes. Decision regions so carved out may
include regions of input space which are not represented by training patterns. In the case of
linear activation, because of the infinite extension of these hyperplanes, the network classifies
the entire input space into the various classes. This causes these networks to display
unsatisfactory generalization characteristics, and make them not suitable for fault diagnosis.
 636

Ellipsoidal activation functions, on the other hand, generate closed decision regions. This
avoids arbitrary inclusion of input regions not represented by training patterns.

Figure 3. Schematic representation of activation functions in the network

Figure 4a shows a schematic of the classification by linear activation networks for the fault
diagnosis problem shown in Figure 1. Figure 4b shows the classification using ellipsoidal
activation networks. The following discussion compares these networks and explains why
ellipsoidal activation is more desirable for diagnostic applications.

Linear Activation Networks

Decision regions are fonned by hyperplanes.

Hyperplanes have an infinite extent, resulting in arbitrary classification of regions of
input space not covered by training patterns.
There is no region of indecision.
The network would classify all measurement patterns to one of the known classes. A
pattern that is far from the training patterns may belong to a known fault class or to
some unknown fault class. In Figure 4a, we see that all of the measurement space has
been apportioned to the known classes. This does not allow any scope for indicating
that the cause for an abnormal measurement pattern cannot be ascertained.
Decision regions for fault classes can be unbounded.
In Figure 4a, all the fault classes have unbounded decision regions though the fault
classes themselves are bounded.
Hyperplanes are not unique.
Hyperplanes formed by the network can be spatially translated to some extent without
affecting the classification of the training patterns. However, this translation changes
the decision regions formed. While such a translation of hyperplanes may not affect
the correct classification of the training patterns, it would change the class membership
 637

of some of the regions in the measurement space. As a result, generalization results

of the network are arbitrary and are sensitive to the network training. Different initial
weights or learning coefficients can result in different generalization results. This
arbitrariness results in a non-robust classification.
No notion of distance is used by the classifier.
All the patterns that fallon the positive-side of the hyperplane get high activation, close
to one. This activation provides no information on the distance of these patterns from
the training patterns. If a hyperplane is adjacent to the class boundary, patterns
immediately on the negative-side of the hyperplane would have low activation even
though they are very close to the training patterns.

1.0

08 Normal behavior

•
~ 0.6

R
C
Abnonnal behnvior
FllUllS known
~ 0.'

02
D Abnormal behnv;or
Fault unknown

0 .2 0.' 0 .6 0 .8 1.0
Concentration

Figure 4a. Fault space classification oflinear activation networks

Ellipsoidal Activation Networks

Decision regions are formed by ellipsoids.

Ellipsoids are bounded regions. Their size can be controlled so that the decision
regions do not include regions of input space far from the training patterns.
Identifies regions of indecision.
Ellipsoidal units have a decreasing output value as one moves away from their centers
radially. Since these units are placed on the training patterns, they have lower values
when they are far from these patterns. As a result, patterns wh ch are far from all the
training patterns are clearly identified as not belonging to any ( f the known classes
including the normal region. They are interpreted as an indical lon of abnormal
behavior of the process whose cause cannot be determined from the available
information.
 638

Decision regions for fault classes are always bounded.

This is because the ellipsoids are bounded regions. The intention is to generate
decision boundaries close to class boundaries, allowing scope for reasonable
generalization but preventing erroneous extrapolations.
Ellipsoids use a notion of distance.
Generalization results are thus based on this notion and hence robust. A pattern that is
closer is more likely to belong to a class than a pattern that is farther.

1,0

08 Normal behaVIor

~ 0.&

"R.
E
Abnormal behavior
Faults mown
~ 0._

02
D Abnorm31 behavior
Fault unknown

0,0
00 0 .2 0_ 0 ,6 0 .8 1.0
Coocentrabon

Figure 4b. Using ellipsoidal activation functions for fault diagnosis problem

From Figure 4b, we see that a single ellipsoid is sufficient to represent the normal region
and ' the fault classes 2 and 3. However, to represent fault classes 1 and 4, we need more than
one ellipsoid. From Figure 4b, it is clear that the network training algorithm has to address
three issues carefully in order to make such a representation possible: (a) it has to determine the
number of ellipsoids required to approximate each fault class. This is crucial since too few
ellipsoids would result in a poor representation of the fault class and too many would increase
the complexity of network training; (b) It has to locate these ellipsoids correctly and determine a
shape based on the local distribution of training set; and (c) it has to ensure that the size of the
ellipsoids generated is contained.
In the following sections, we address these issues related to network architecture and
training. We characterize the problems faced in the training of these networks and propose
solutions. The main issues of focus are (a) robust generalization characteristics and (b) compact
representation so that a class is represented by as few ellipsoids as possible.
 639

4. Unsupervised and Supervised Learning Strategies

A satisfactory solution to these problems require the consideration of two different approaches
to learning, narnely, unsupervised and supervised learning. We compare these two approaches
in this section and then go on to propose a strategy that integrates both.

4.1. Unsupervised Learning

Unsupervised learning is used when the class information is not available but the
patterns are to be grouped based on similarity.
Methods used for this purpose are familiarly known as self-organized learning or
competitive learning algorithms. The idea is to partition the training patterns into
disjoint groups so that each group has patterns similar in some respect.
The features extracted by these methods are such that they identify which patterns are
similar. The class information of these patterns is of no relevance and is not used in
the construction of these features.
Objective of these methods is to give credit for similar patterns to group together. No
penalty is given for grouping patterns, belonging to different classes, as being similar.
Since the objective uses a notion of distance (similarity), patterns far away from a
group can be accordingly judged in terms of the distance metric. Thus, in Figure I,
patterns far from fault classes 1 and 2 would be considered as belonging to the region
of indecision or to other classes.
In Figure I, fault classes 3 and 4 are very close and appear similar. A simple measure
of distance has no reason to view them as distinct. Without the class information, they
are treated as belonging to the sarne or similar group.
Competitive learning (or self-organized learning) is used as a means to grol!P the training
data by similarity. The grouping procedures give credit for patterns which are similar but do
not pay any attention to their membership to different classes. In Figure I for example, all
patterns of classes 1 and 2 are properly categorized just by grouping by similarity. However, a
simple measure of similarity will not solve the classification problem for classes 3 and 4. Even
a representation by many clusters may not solve the problem of misclassification.

4.2. Supervised Learning

Supervised learning is used for classification problems where the class information is
available.
 640

Typically, these methods use an error criteria such as RMS error minimization.
Objective of these methods is to penalize patterns which are wrongly classified. No
credit is, however, given for detecting similarity between a group of patterns.
The features extracted by these methods are such that they can be useful for
discriminating between the patterns of different classes. Notion of similarity plays no
direct role.
The objective of these methods does not penalize for including regions which are not
similar to the patterns in a class. This is one of the causes for the extrapolation
problems which arise in classification. Since the objective does not consider the effect
on patterns which are not a part of the training set, there is no built-in mechanism for
avoiding extrapolation completely. Thus, in Figure I, there is no reason why the
decision regions for fault classes 1 and 2 should not include the regions of indecision.
This results in poor generalization.
Performs credit assignment to various features so that a finer classification can be
brought about even when two classes are close and sharing nonlinear decision
boundaries.

4.3. Integrating Unsupervised and Supervised Learning Strategies

Supervised learning tries to determine features which serve to result in a maximum separation
between the classes. Features determined solely on this criteria do not consider the notion of
similarity. As a result classes may be arbitrarily assigned regions of measurement space where
no information (training patterns) is available. Unsupervised learning tries to determine features
by grouping patterns based on similarity. However, no information of class is considered
during the feature determination and the resulting classification can be poor. To avoid the
arbitrariness and get good classification, what is necessary is a method that uses features which
use a notion of similarity and then perform a credit assignment to these features to achieve good
classification. Unsupervised learning provides a way to group patterns by similarity thus
identifying required features. Supervised learning would assign relative credit to these features
to achieve finer classification. This integrated learning strategy brings together the good
characteristics of both methods while overcoming their undesirable characteristics. In Figure I,
clearly neither unsupervised or supervised learning will independently do the proper isolation of
classes 3 and 4. However, their combination will solve the problem. Unsupervised learning
can be used to identify the necessary features for each of these classes separately, without the
information of other classes. These features can be used to determine the structure of a network
[e.g. each feature forms a hidden node] and thus avoid the problem of arbitrarily determining
 641

the number of hidden nodes as practiced in the standard approach. Then, supervised learning
can be used to perform a credit assignment to these features, this time with the use of class
information, so that a proper classification is achieved.
Use of this integrated learning strategy also offers solutions to other problems [9, 10] :
solve the hidden node problem avoiding the unsatisfying and expensive trail-and-error
methods
overcome local minima traps which are severe when supervised learning alone is used
allow the development of decomposition strategies which make the fault diagnosis of
large-scale industrial processes tractable (discussed later in section 7)
avoid extrapolation problems arising through the introduction of bounding schemes

5. Clustering as Unsupervised Learning

Unsupervised learning is also known as associative learning, competitive learning or simply

clustering. The underlying purpose of these approaches is as follows: In the input space, there
may be regions where data samples are heavily concentrated. These regions are called clusters.
The centers of the clusters are the representative points of the training patterns and are much
fewer in number than the training patterns. The idea pursued here is that a cluster center along
with a local measure of distance should define the local density of the data distribution
sufficiently well. Clustering saves storage space since only a few cluster centers have to be
stored instead of a large number of training patterns. It saves computational time since the
computations (such as distance measures) have to be done only with cluster centers and not all
the training patterns. It gives good generalization because it captures the general characteristics
of the underlying distribution.

5.1. Why Mixture Density Representation?

If a density function has local regions where the densities are high, then it is important to know
the covariance of the density in each local region. The local covariance is used in the local
region for the purpose of measuring distance with a non-Euclidean distance measure. When the
density function of the class concerned has multiple modes, it should be represented by a
mixture density. The basic motivation behind this approach is that the modes of a continuous
density function are important in describing and analyzing the underlying data structure for
recognition purposes. The problem of isolating the modes is called feature extraction or
clustering. A mixture of density functions from a particular family, such as the Gaussian, can
 642

be used to approximate the density function of the underlying distribution. Upon detecting the
modes, one could approximate the local region by a Gaussian and the overall density by a
mixture of Gaussians.

5.2. Scope of Clustering

Clustering makes sense only in terms of a priori knowledge used. For example, the results of
clustering may not make sense when interpreted by a person who is using a field of knowledge
not incorporated in the clustering procedure. Clustering methods use some notion of a metric
and this can be validated only by assumptions made on the distribution of the input patterns.
Clusters could be very complex in the measurement space. For example, clusters for two
different classes could be interwoven as hyper-coil springs in a multi-dimensional space.
Without the knowledge that the respective classes have complex loci, the data points for
different classes will appear! close together and it will not be possible to separate points from
different classes. There may be a need to define locii such as hyper-spirals, hyper-oscillations
etc. in order to determine the appropriate distance metric a priori. Choice of a distance metric
used for clustering would thus presume something about the class distributions. Empirical
studies have shown that decision regions of many classification problems are ellipsoidal and
many class distributions are approximately Gaussian [1]. Most often, a multivariate normal
distribution (or a mixture of them) is assumed for each class, and although this assumption is
not often met, the resulting features are still considerably better than those that are randomly
derived.
A clustering procedure is practicable only when it forms a minimal number of cluster centers.
Otherwise, a tight upper bound should be available for the number of clusters a priori, but this is
generally not possible. What is desirable is a clustering technique which, once given a particular
choice of distance metric, determines a conservative number of clusters. Even if the preset upper
limit on the number of clusters is high, the algorithm should utilize only the required number of
clusters.

5.3 K-means Clustering

K-means clustering [4] is a simple and powerful clustering algorithm. K-means clustering
presupposes the number of clusters needed and would cluster the data accordingly. It utilizes
all the cluster centers so that each of the clusters is guaranteed at least one pattern. We present

!The look will be with a Euclidean distance measure.

 643

the k-means clustering algorithm in an adaptive fonn as a network [13]. This fonn makes it
convenient to compare with other methods to be discussed later. The network has a layer of
input nodes connected to a layer of output nodes. Each of the output nodes is a cluster. The
weights connecting the input nodes to an output node j give the cluster center coordinates.

Notation
c = number of cluster centers chosen
n = total number of inputs
Np = total number of patterns
Wji : weight connecting input node i to output node j
[Wjl , Wj2 , ... Wjn ] = cluster center j
dkj =Euclidean distance of jlh cluster center from the kth pattern

K·means Clustering Algorithm

1. Choose a priori the number of clusters c. Randomly choose c training patterns and assign
these as the initial cluster centers Wj.
2. For each training pattern k, find the distance dkj from all the cluster centers j. Let j* be the
cluster center that is closest to the pattern.
t\ Wj'i = (X ki -Wj'i)
Wj'i = Wj'i + Lrate*t\Wj"i
3. Go to 2 if there are changes in the cluster centers Wj-
Note that in step 2, a pattern is allowed to belong to only one cluster - the cluster that is
closest to. the pattern. Each pattern is grouped to a cluster center that is closest. The cluster
centers are then moved to the mean of the corresponding groups. This process is repeated until
the cluster centers stabilize. The main drawback of this method is that utilizes all the cluster
centers in the fonnation of clusters. Since the exact number of clusters in a distribution is in
general not known, one will have to overestimate the number of clusters in order to properly use
this method.
Figure 5 shows K-means clustering in the fonn of a network. For each training pattern,
distances are computed from each of the cluster centers and the pattern is included only in the
closest cluster. This is called a winner-take-all strategy where a cluster which is closest gets all
the credit for a pattern. This is a fonn of competitive learning which introduces a lateral
inhibition between the output nodes (or clusters). This is also sometimes referred to as hard
competitive learning, since it lets only one winner considering the rest to be losers.
 644

This node detennines the winning cluster

Winning cluster is the closest cluster

!
Output nodes or CI usters

W"....~cl~.=""
Input nodes

Winner-lake-all Competitive Learning

Figure 5. Network structure ofK-means clustering

A generalization to this, called soft competitive learning, can come in two forms:
• There are k winning clusters and the rest loose .
• All clusters have a fuzzy membership which specifies to what degree they
win or loose a pattern, instead of simply specifying whether they win or
loose a pattern.

5.4. Kohonen's Self-organizing Feature Maps

Kohonen's self organizing maps is related in principle to k-means clustering. Kohonen's

model also consists of two layers. The first layer contains the input nodes and the second layer
contains the output nodes. Adjustable weights completely connect all the input nodes to the
output nodes. Though the original Kohonen's algorithm [11] used output nodes in a 2-
dimensional array, others have used it in a linear array [8]. The neighborhood number is set
initially to ~ and gradually reduced to O. Topological neighbors are considered on either side of

the linear array_ of output nodes.

Kohonen's Algorithm

1. Randomly initialize weights Wji. Set the neighborhood size Nc to ~.

2. For each training pattern k, select the output node j* such that dk/ is minimum.
 645

Update Wji using the following rule:

l1Wji = (Xki - Wji)
Wji = Wji + Lrate * l1Wji
where j includes j* and each of its Nc neighboring nodes on either side.
3. Go to step IT if there are significant changes in weights.
4. If (Nc =0) stop
else Nc = Nc - 1 and Go to 2.
Notice that the above algorithm uses the notion of topological neighborhood. The topology
is in reference to the ordering of the output nodes (cluster centers) in the network. The algorithm
identifies the cluster center that is closest to a training pattern, and updates this cluster ~enter and
all its topological neighbors. Kohonen's algorithm (known familiarly as the self-organizing
map) is based on the idea that the neighboring cells in a neural network compete in their activities
by means of mutual lateral interactions, and develop adaptively into specific detectors of
patterns. The idea is that the locations of the responses tend to become ordered. Kohonen's
method differs from the k-means clustering in that there is no single winner anymore. Instead,
all the clusters in the neighborhood are winners. K-means clustering can be shown to be a
special case of Kohonen's clustering algorithm. If we set the neighborhood size Nc = 0 in step I
of the Kohonen's algorithm, we get the k-means algorithm.
After a neighborhood number is decided, the algorithm makes all the clusters in the
neighborhood to be the winners of the pattern. In fact, it gives the same weightage to all the
clusters, in the topological neighborhood, in Updating their centers for a pattern k. However, it
is better to have the weightage for updates dependent on the relative distances of their centers
from pattern k. This is essential because topological neighbors may not be neighbors by
distance and it makes sense to give a higher weightage to a closer cluster. A fuzzy-membership
scheme may be used such that a cluster center which is relatively far off from a pattern has a
very little movement towards the pattern. There are other advantages to the use of fuzzy
membership such as reducing the effect of gravity from dense regions which will be discussed
later.

6. Notion of a Fuzzy Membership

When a cluster belongs to the neighborhood, Kohonen's method updates the center, otherwise
it does not. This is a case of hard partitioning where a pattern has either a membership value of
oor 1 to a cluster. A fuzzy partitioning, on the other hand, allows us to give a soft partitioning
by allowing any value in the closed interval between 0 and 1. We use below a membership
function due to Bezdek[2] in his fuzzy-c-means algorithm. Huntsberger and Ajjimarangsee [8],
 646

used this membership function in a self-organizing model for unsupervised pattern recognition
based on Kohonen's self-organizing maps.
Ukj : membership value of kth pattern in the jth cluster
Ukj = 1 if dkj =0
=0 if dkl =0, (I * j, 0 $; I, j $; c-l)
dk .. 1
=L
c·1
(_l) otherwise
1=0 d kl
This fuzzy membership value is in the range [0 1]. Membership value is inversely
proportional to the distance of kth pattern from jth cluster,· dkj. Membership values are
normalized. Membership of any pattern in all the clusters add up to L This means that
membership considers only the relative distances of the pattern from the cluster centers.

6.1. Modified Kohonen's Algorithm with Fuzzy Membership

I. Randomly initialize weights Wj. Set the neighborhood size Nc to i. Set Lrate 2 value.

2. For each training pattern k, select the output node j* such that dk/ is minimum.
Update Wji using the following rule:
t:,.Wji=Ukj * (Xki - Wji)
Wji =Wji + Lrate * t:,.Wji
where j includes j* and each of its Nc neighbors on either side.
3. Go to step II if there are significant changes in weights.
4. If (Nc =0) stop
else Nc = Nc - 1 and Go to 2.
Figure 6 shows the clustering procedure in the form of a network. From a comparison of
Figures 5 & 6, it is clear that there is no single winning cluster for any pattern when
membership function is used. Instead, the membership specifies which clusters are close and
which are far. In the case of Kohonen's method, one could consider that there is a hard
membership function which gives a membership of 1 for the closest cluster and zero for the rest
of the clusters.
Use of fuzzy membership serves in four respects:
• The fuzzy membership moves a cluster center towards a pattern by an increment
inversely proportional to the distance of the cluster center from that pattern. As a result,
a cluster center which is relatively far off from a pattern gets a smaller move towards the
pattern than a cluster center that is closer.

2In all the case studies considered, Lrate was set at 0.001. It has been recommended in literature that the value of
Lrate be decreased with iterations. Larger Lrate means that cluster centers get larger movements.
 647

Nodes computing
fuzzy membership

I
Output nodes or Clusters

W. . . . .~"""'~_
Input nodes

No Single Winning Cluster

Figure 6. Network structure for Kohonen clustering with fuzzy membership

• Cluster centers may tend to move close when there are dense regions of patterns. There
is no mechanism in the clustering algorithm to prevent cluster centers from moving
towards each other until they coincide. Fuzzy membership provides such a facility
indirectly. Direct repulsion by penalty could be an alternative to avoid cluster centers
from moving closer to each other. However repulsion may inhibit the movements of
cluster centers thus resulting in local minima problems. When a pattern is relatively very
close to a particular cluster center, the fuzzy membership is close to one for that cluster
and is near zero for all other clusters. This particular feature discourages cluster centers
from moving close to each other. It prevents the clusters from getting attracted to
regions of high gravity.
• Fuzzy-membership also helps avoid excessive usage of cluster centers - i.e., cluster
centers are sparingly used. Redundant cluster centers do not get utilized because their
fuzzy-membership values drop to zero for all training patterns. This conservative
approach to the use of cluster centers results in a minimal set of features (cluster centers)
needed for the description of the distribution of the patterns. This is essential to
classification problems because the number of features determines the complexity of the
supervised learning to be performed for classification.
• Fuzzy membership provides a mechanism to allow overlapping clusters. When the
fuzzy membership of two clusters is high for a pattern, the pattern is a part of both the
clusters. The final clusters developed are thus allowed to overlap. The Kohonen's
basic algorithm like other clustering methods does not allow clusters to overlap.
 648

6.2. Avoiding Regions of Gravity

Gravity problem arises when there are clusters of disproportionate sizes. To illustrate the
gravity problem, we consider the distribution in Figure 7a. There are three clusters in the data
with two clusters having nine data points each while the third cluster has only two. Figure 7a
shows three nonoverlapping clusters. To illustrate gravity, we consider the initial cluster
centers within the large clusters. The starting cluster centers chosen are : (0.8, 0.9) (0.9, 0.9)
and (0.0, 1.0). When K-means clustering is applied to this situation, the algorithm identified
the following as the final cluster centers - (0.81, 0.81), (0.91,0.91) and (0.77, 0.11). Notice
that the first two cluster centers are trapped in a large cluster while the smaller cluster was left
unrepresented. This problem arises when the cluster centers get trapped in a region of higher
"gravity". This problem is attributable to the mutual exclusion of the clusters considered in the
K-means clustering.
Figure 7b shows the trajectory of the cluster centers for Kohonen's algorithm without the
fuzzy membership. Because of the notion of neighborhood, clusters are allowed to share
patterns and the cluster center is able to move out of the local cluster but is not able to steer
away from the gravity due to the larger clusters to move towards the smaller cluster. It
performs better than K-means clustering because the clusters are not localized with mutual
exclusion. Instead, the cluster sizes are started with large neighborhoods which is then
gradually reduced. However, the method still suffers from the influence of high gravity -
though there are clllster centers present in the gravity areas, other cluster centers are still
attracted to these denser regions.

1.2

1.0 + + +
+ + +
0.8 + + +

0.6

X2 + Data
0.4

0.2 + + +
+ + + +
0.0 + + + +

.().2+-_..,.._....._...,.___....,-_....--!
-0.2 ~O.O 0.2 0.4 0.6 0.8 1.0 1.2

Xl

Figure 7a. Clusters with disproportionate number of points

 649

1.0-r----------------:::-.

0.8

0.6
-0- Cluster center I

X2 -tr-- Cluster center II

0.4 --0-- Cluster center III

0.2

o.o ...._ - . , . . _ -_ _-..,..._ _..,..._ _ ~

0.0 0.2 0.4 0.6 0.8 1.0

Xl

Figure 7b. Trajectory of clusters centers - Standard algorithm

One could circumvent this problem if one has a scheme to give less importance to regions
(dense or otherwise) which are already represented by cluster centers. Figure 7c shows the
trajectory of the cluster centers when fuzzy membership is used. Since one cluster center
provides for the representation at each of the denser regions, the membership value for the third
cluster center is low at these regions of higher gravity. The third cluster center is thus able to
steer its way to the third smaller cluster. By using fuzzy membership, provision was made for
repulsion between the cluster centers. This repulsion helps the cluster center to steer against the
gravity of the larger centers towards the smaller cluster.

1.0..--------------:::--....,

0.8

0.6 --0- Cluster center I

X2 -ts- Cluster center II

0.4 --0- Cluster center III

0.2

0.2 0.4 0.6 0.8 1.0

Xl

Figure 7c. Trajectory of clusters centers - Algorithm with fuzzy membership

 650

7. Decomposition Strategies

For the diagnosis of large-scale industrial processes, network training task becomes
computationally intensive. It is imperative that decomposition techniques be used for achieving
practicable complexity. Speed up may be achieved by the following three decomposition
strategies which result in smaller networks with few weights and few training patterns:
(a) Network decomposition: The network is broken into smaller networks so that each of
the subnetworks can be independently trained. Since the network training complexity is at least
D(N3), where N is the number of weights, training the subnetworks independently is
computationally more efficient. IfNk is the number of weights in the kth subnetwork, then:
m
L Nk = N
k=1
m
and L (Nk)3 « N3
k=1
(b) Training set decomposition: When the training set is very large, it can dominate the
computational complexity as well. Training set decomposition exploits the fact that only a part
of the training set is relevant to the training of each of the subnetworks. Training patterns
which have negligible influence on the training of a subnetwork are identified and eliminated.
This results in small training set sizes.
(c) Input space decomposition: In a typical industrial process, the number of sensors is
quite large. In such a case, the input space dimension or the number of input nodes becomes a
dominating factor. Dimensionality reduction can be achieved by exploiting such factors as
sensor redundancy. Restricting to the use of local information (with respect to process
structure) is another possibility where only the close by relevant sensors are considered for each
fault class.

7.1. Network Decomposition

In networks with ellipsoidal units, hidden nodes are not shared between the output nodes, i.e.
each of the hidden nodes is connected to only one output node corresponding to a particular
class. This is because the ellipsoids formed by the hidden nodes are dedicated to a particular
class. Not sharing the hidden nodes amounts to reducing the number of connections
substantially thus decreasing the dimensionality (number of weights) of the training problem.
The network architecture is inherently decomposable since the hidden nodes are not shared
between the output classes in the network. Each of the subnetworks has one output node and
its own dedicated hidden nodes. Such a decomposition implies that there is no communication
 651

required between the subnetworks and hence their training can be done asynchronously on
different processors. Since the complexity of network training is at least O(N3), a linear
decomposition of the network would result in a faster network training i.e., the sum of the
training times for the subnetworks would be much smaller than that of the completely connected
network. Further, the decomposition allows for an asynchronous computation on the
subnetworks, resulting in a parallelism of 100% efficiency.

7.2. Training Set Decomposition

The training of feedforward neural networks exploits the parallelization that comes from
localized computation. The fact that computations can be localized at the processors allows for
a fast computation. When localized representation is used, as in the case of Gaussian units or
ellipsoidal activation functions, the network still retains this property of localized computation.
In addition, the network training can be modified so that it exploits the localized representation
of units in the input space. The value of a node with ellipsoidal activation function drops
radially to zero as one moves away from its center. For patterns far from its center, its value is
near zero. When the standard backpropagation is used in network training, the error
backpropagated to a hidden node by a pattern far from it, is near zero. As a result, the
contribution of this pattern is negligible to the modification of the weights connecting the input
nodes to the hidden nodes. If this near zero error is not backpropagated to the hidden node, its
weight updates are negligibly affected. This idea is being exploited in partitioning the training
patterns in the training of the network. All the patterns which give near zero value to a hidden
node are removed from the training set of that node. This considerably reduces the training
task.

7.3. Input Space Decomposition

So far, we had considered feedforward neural networks with ellipsoidal activation functions
and speedup of their training by network decomposition as well as training set decomposition.
However, when the number of input nodes is very high, the complexity of network training
will be dominated by the training of input-hidden nodes weights. One can reduce the
dimensionality by considering only the structural close sensor measurements. There is an
alternative method that utilizes only the distribution information of the training patterns. Recall
that each of the hidden nodes describes an ellipsoid in the input space and approximates a part
of the distribution of a particular class. Ideally, each of the principal axes of the ellipsoid
 652

should show how this data is distributed locally along those directions. If the distribution is
oriented to the axes of reference, approximating this with a set of unoriented ellipsoids is not
efficient. To use the power of representation that ellipsoids provide, we need to allow for the
orientation of the ellipsoids generated. For representing an oriented ellipsoid in an n-
dimensional space, the number of parameters N (center coordinates + axes lengths + axes
orientation vectors) grows as 0(n2). Since the network training is at least 0(N3), it is
imperative that we limit the growth of N with problem size. For large n, the problem of
network training (determining the parameters of oriented ellipsoids) would be very time-
consuming. To avoid this complexity, the number of parameters is traditionally kept at O(n) by
fixing the orientation of the ellipsoids to be along the reference axes· [6, 9]. Fixing the
orientation of ellipsoids a priori reduces the advantage of using ellipsoids instead of spherical
units. For example, in approximating an oriented Gaussian distribution, we would need many
fixed orientation (along reference axes) ellipsoids. We introduce a method of computing the
orientation, without increasing the complexity, by virtue of the fact that the network training is
performed in two phases: namely (1) clustering phase (2) ellipsoidal unit tuning. Clustering
phase determines the clusters present in a given distribution. Each of these clusters is then
approximated by an ellipsoid and tuned. If the cluster is oriented, a fixed orientation ellipsoid is
a poor approximation. If we can determine the cluster orientation, this orientation can be used
as the fixed orientation of the corresponding ellipsoidal unit. This is a more sensible choice
than the arbitrary choice of reference axes.
Let XI, X2, ... , Xm be the original variables defining the cluster patterns. Clusters would be
oriented, with respect to the reference axes in the x-space, if these variables are correlated in the
patterns of the cluster. Orientation of the cluster can be determined through principal component
analysis (PCA). It is essentially a mathematical technique involving transformation of the
original variables to a new set of uncorrelated variables called principal components. Principal
components are linear combinations of the original variables and can sometimes be identified as
being meaningful in their own right. Let YI, Y2, ... , Ym be the principal components. Then:
Yj =ajlXI + aj2X2 + ... + ajm Xm (j = 1,2, ... , m) or y =Ax
with the condition that A is an orthonormal matrix i.e., ATA =I.
Normalization of A ensures that distances are preserved between the patterns in the cluster
so that transformation by A only results in orientation of the cluster. Further, orthogonality
ensures that the new variables y are uncorrelated in the patterns describing the cluster. By
performing this transformation from x to y, we have determined a new frame of reference
where the variables are uncorrelated ie., the cluster which is oriented to the reference axes in the
x-space is now oriented along the reference axes in y-space. For example, an oriented
Gaussian distribution in x would be unoriented in y. Therefore, we can approximate the
distribution in y-space by an ellipsoid oriented along the reference axes. Characterization of
 653

the transformation can be done the following way: Let Lx be the covariance of the cluster in x-
space. Then:
LxcI> = cI>A
where:
cI> : matrix of the eigenvectors of Lx
A : diagonal matrix of the eigenvalues of Lx
Now, consider the transformation: y: <1>x.
mx = mean ofx
my: mean of y
Ly =A
Since Ly is diagonal, the y variables are uncorrelated. Thus, if we choose A : <1>, we achieve
the required transformation. This transformation to y-space chooses the directions of
eigenvectors of Lx as the new reference axes.
There is another advantage to this transformation besides determining the orientation of the
cluster. Because the measurements in a process are often redundant, ie. they are correlated, the
number of "independent pieces" of information is much fewer in number than the number of
sensors. Among these "independent pieces", only a few carry the significant information.
Each of the eigenvectors is an "independent piece" of information. Magnitude of the
corresponding eigenvalues gives the quantity of information (variance) associated with it.
These remarks can be quantified as follows: Let us assume, without a loss of generality, that the
eigenvectors are ordered in decreasing order of eigenvalues. Variance of the distribution along
the jth principal component Yj is given by the corresponding eigenvalue: A.j : var(Yj).
m m
Total variance in the data: L var(Yj) : L var(xj)
j=l j=l
The proportion of variance accounted for by the first p principal components is :
p
L var(Yj)
j=l
m
L var(Yj)
j=l
If this proportion is reasonably large (say 95%), then the first p principal components
provide an adequate description of the data. The rest of the components can be conjectured to
consist mostly'of noise. Not using the components with small variance serves to eliminate
noise. Further, if the original variables are redundant, Lx is singular, some of the Yj would
have zero variance ie., have no information along these directions. Considering only significant
components considerably reduces the dimensionality of the problem (from m to p, where p is
much smaller than m).
 654

8. Fault Diagnosis of a Reactor-Distillation System

We will consider the fault diagnosis of a reactor-distillation process system shown in Figure 8,
for illustrating the application of the techniques discussed so far. The CSTR processes reactant
A to result in a mixture of A and product B. Two P-controllers are used to control the reactor
temperature and holdup by manipulating the coolant flowrate and the product stream flow rate
respectively. The reactor product stream containing the binary mixture of A and B is fed into a
distillation column where it is separated in to a distillate stream with 98% B and a bottoms
stream containing 2% B. Two PI-controllers are used to control the overhead and bottom
product compositions by manipulating the reflux rate and the vapor-boilup rate, respectively. A
total often sensor measurements are available. Details of the process model and the listing of
faults considered may be found in Vaidyanathan [16]. The purpose of this case study is
twofold. One, to perform fault diagnosis on a system that is of interesting complexity. Second,
to examine the underlying structure of the fault space spanned by process measurements and to
bring out the asymmetric nature of fault classes that is often found in large scale problems. We
show how the redundancy in the data can be used to reduce the dimensionality of the data and
why ellipsoids are well suited for representing the distribution.
A total of eight faults were considered for training the network. Training patterns were
chosen from the simulation 40 minutes after the inception of the fault so that the fault signature
is significant in the measurements. Training data was next analyzed for its correlation structure.
For illustration purposes, we considered fault F4 (increase in the temperature of the coolant
feed) for the analysis. Data is analyzed for each fault by finding the principal components.
Principal components are the eigenvectors of the covariance matrix of the distribution. To
explain geometrically, they correspond to the direction of orientation of the distribution. For
example, if the distribution is an oriented Gaussian distribution, principal components
correspond to the directions of its principal axes. Corresponding eigenvalues give the variance
along each of these directions. To allow for oriented ellipsoids, the number of parameters grow
as O(n 2), where 'n' is the number of measurements. If the orientation is fixed, the number of
parameters grows only as O(n). The first advantage of principal components is that they
identify the data orientation so that by fixing the orientation of the ellipsoids along them, we not
only use the orientation information but also limit the number of parameters to O(n). The
second advantage of using principal components is that they are uncorrelated unlike the
measurements. This means that one can see how the data spreads along each of these directions
independently. The spread on the direction of a principal component is given by the variance
(corresponding eigenvalue of the covariance matrix). If the sensor measurements are
correlated, and they usually are because of their redundancy, then the variance of the
distribution is significant only along some of these directions. Along the rest, the variance is
 655

negligible or zero. This means that one can approximate the distribution very well by only
considering the few significant principal components. By mapping the distribution from
measurement space to the principal components, we can not only identify the orientation of the
distribution but also reduce the dimensionality by ignoring the insignificant components.
Principal components were identified for the distribution of fault F4. These components
were ordered in the data of decreasing variance along those directions. The first three
components are shown in Table 1. These three components account for 92% of the total
variance in the data, which can be interpreted as 92% of the information in the distribution of
the data. This means that the data in the distribution predominantly exists in the subspace
spanned by these three dimensions. The remaining space (7-dimensional) has nearly no
information on the distribution of the data. Distribution of the training patterns for fault class
F4 are shown in the space of these three significant components in Figure 9a.
In using a network with ellipsoidal units, a single ellipsoidal unit is used to approximate the
distribution for each fault class, in the space of components. That is, the principal axes of the
ellipsoid are now oriented along the principal components and not the axes of the measurement
space. Because the ellipsoidal unit can have different lengths along its different principal axes,
it is possible to limit the unit to the subspace spanned by the significant components with near
zero spread along the remaining. If one were to use hyper-spheres instead of hyper-ellipsoids,
we would be stuck with a large ten-dimensional sphere in this example as the hyper-sphere, due
to its spherical symmetry, would span the measurement space with equal variance (i.e. equal
radius) in all the directions. This leads to serious generalization problems. Figure 9b shows the
distribution in Figure 9a being approximated with a single ellipsoid. In this case, clearly a
single ellipsoidal unit approximated the class distribution. It is to be emphasized that the
ellipsoid 1S actually in the 10 dimensions spanned by the components with near zero or zero
variance along 7 of the dimensions. The largest axis of the ellipsoid is along the principal
component with the largest variance, the second largest axis along the second component with
the second largest variance and so on.

Table 1. Significant directions in data distribution for Fault 4

Significant Principal Components Variance

[0.091 -0.075 0.206 0.203 -0.127 0.932 0.134 0.019 O.013]T 0.981

[0.4140.352 -0.508 0.0460.154 0.0410.5 -0.196 0.354 -O.092]T 0.125

[ -0.213 0.422 -0.112 -0.3 -0.097 0.073 0.346 0.36 -0.42 0.483]T 0.084
 RJ OJ To

I - l _ - - ' _.....
_ FJ 1]
~_" 1 I
... I
I
.~ I
FJO T JO . I
I
a>
• I DX! U1
• I a>

8---~
G
e--4-
IQ-~
IV I! G I
I ,
I
I
I

B Xb

Figure 8. Reactor-distillation cohunn case study

 657

•.J .1.

Figure 9a. Distribution of training patterns in the space of principal components for fault F4

Figure 9b. Ellipsoidal approximation of the training patterns for fault F4

 658

The same procedure was perfonned for all the other fault classes and a single ellipsoid was
found sufficient to approximate each of the classes. Only 100 epochs of tuning were used for
100% correct classification. This demonstrates the efficiency of the representation of data by
ellipsoids.

9. Summary

Classifiers have to meet the minimum requirements of robust classification and reasonable
generalization for successful fault diagnosis applications. In this paper, we have shown that
linear activation neural networks have serious problems in meeting these requirements and the
use of ellipsoidal activation is more appropriate. After comparing unsupervised and supervised
learning strategies, we propose an integrated learning scheme which has the advantages of both
the approaches and avoids the problems of each. In this integrated approach, a fuzzy
unsupervised learning is used to determine the features and the structure of the network. Then,
supervised learning is used to perfonn credit assignment to various features by fine-tuning the
network parameters. Decomposition techniques have been introduced to result in smaller
networks with fewer training patterns. This makes the network suitable for large-scale process
plants. Since the hidden nodes are not shared between classes, the network decomposes into
smaller networks one per class. Since the ellipsoidal units are localized in the input space, only
the training patterns in its proximity need be considered for tuning its parameters. This results
in training set decomposition. Finally, the measurement space is decorrelated through principal
component analysis. This helps decrease the dimensionality of the input space. A reactor-
distillation column process system has been used to show the structure of the distributions in
the measurement space for various faults and the suitability of ellipsoidal units was addressed
in this context. We also point out that fault diagnosis of continuous and batch processes can be
treated uniformly if the transient behavior is considered, ie. using trends to perfonn diagnosis
instead of steady state values. Ellipsoidal units provide a means to develop envelopes which
enclose the trends for individual fault classes. Network output values give a measure of how
close or far we are from the envelopes of different fault classes.

Acknowledgments

The authors gratefully acknowledge the National Science Foundation (Grant ECS-9013349)
and the National Institute of Occupational Safety and Health (Grant OH02740-A1) for their
support of this work.
 659

References

I. Batchelor, B., Practical approaches to pattern recognition, Plenum Press, New York, 1974.
2. Bezdek, J. c., Pattern recognition with fuzzY objective function algorithms, Plenum Press, New York, 1981.
3. Cybenko, B., "Approximation by superposition of a sigmoidal function, Math. Control. Signals Systems, 2,
303-314, 1989.
4. Duda, R. O. and Hart, P. E., Pattern classijication and scene analysis, Wiley, New York. 1973.
5. Fukunaga, K., Introduction to pattern recognition, Acadentic Press, New York. 1972.
6. Holcomb, T. and Morari. M. "Local training for radial basis function networks: Towards solving the hidden
unit problem", 2331-2336, American Control Conference, 1991.
7. Hoskins, J. C. and Himmelblau, D. M., "Anificial neural network models of knowledge representation in
chemical engineering", Comput. Chem. Engng., 12,881-890, 1988.
8. Huntsberger, T. L. and Ajjimarangsee, P. "Parallel self-organizing feature maps for unsupervised pattern
recognition", Int. J. General Systems, 16, 357-372, 1990.
9. Kavuri, S. N. and Venkatasubramanian, V., "Using Fuzzy Clustering with Ellipsoidal Units in Neural
Networks for Robust Fault Classification", Comput. & Chem. Engng, 17, 8, pp. 765-784, 1993.
10. Kavuri , S. N. and Venkatasubramanian, V.,"Solving the hidden node problem in networks with ellipsoidal
units and related issues", International Joint Conference on Neural Networks, Baltimore; June 1992.
II. Kohonen, T., Self-organization and associative memory, Springer-Verlag, Berlin. 1984.
12. Leonard, J. A. and Kramer, M. A. "Limitations of the backpropagation approach to fault diagnosis and
improvement with radial basis functions", presented at the AICHE Annual Meeting, Chicago; November,
1990.
13. Moody, T. 1. and Darken, C. 1., "Fast learning in networks of locally tuned processing units", Neural
Computation, 1, 281-294, 1989.
14. Rippin, D. W., "The future of process operations", presented at the workshop Computer Aided Process
Engineering, World Congress of Chemical Engineering, Karlsruhe, Germany, June 1991.
IS. Ungar, L. H., Powell, B. A. and Kamens, S. N., "Adaptive networks for fault diagnosis and process control,
Comput. Chem. Engng., 14, 561-573, 1990.
16. Vaidyanathan R, Process fault detection and diagnosis using neural networks, Doctoral Thesis, Department of
Chemical Engineering, Purdue University, December 1991.
17. Vaidyanathan, R., and Venkatasubramanian, V., "Representing and Diagnosing Dynamic Process Data Using
Neural Networks", Engineering Applications of Anijicial Intelligence Journal. S, 1, pp. 11-21, 1992a.
18. Vaidyanathan, R., and Venkatasubramanian, V., "On the Nature of Fault Space Classification Structure
Developed by Neural Networks",Engineering Applications of Anijicial Intelligence Journal. 5,4, pp. 289-
297, 1992b.
19. Venkatasubramanian, V., Vaidyanathan, R. and Yamamoto, Y., "Process fault detection and diagnosis using
neural networks: I. Steady state processes", Comput. Chem. Engng., 14, 699-712, 1990.
20. Venkatasubramanian, V., "Recall and Generalization Performances of Neural Networks for Process Fault
Diagnosis ", in Proceedings of the Founh International Conference on Chemical Process Control. South
Padre Island. Texas, Feb 17-22. 1991.
 Overview of Scheduling and Planning
of Batch Process Operations·

G. v. Reklaitis

School of Chemical Engineering, Purdue University, West Lafayette, IN 47907, USA

Abstract: Scheduling of batch operations is an important area of batch process systems

engineering which has been receiving increasing attention in the last decade, especially in its role
within computer integrated process operations. In this paper, we review the basic issues which
scheduling methodology seeks to address and outline some of the reasons for the growth of
interest in this field. The components of the scheduling problem are described and the main threads
of the available recent solution methodology are reviewed.

Keywords: Batch size, campaign, cycle time, equipment, multiproduct plant, multipurpose plant,
network flowshop, operating strategy, planning, reactive scheduling, recipe, resource constraints,
storage, time discretization.

Introduction
The period of the last fifteen years has seen an increasing level of interest, research, and
publication in production planning and scheduling methodology for the chemical processing
industry. In this paper, we will examine this growth first by describing the basic issues which this
methodology seeks to address and then outlining some of the reasons for this growth of interest.
The components of the batch process scheduling problem will be described in detail and the main
threads of the relevant research directed at its solution will be summarized.

IThis paper draws extensively on a plenary lecture entitled "Perspectives on Scheduling

and Planning of Process Operations", presented by the author at the Fourth International
Symposium on Process Systems Engineering, Montebello, Quebec, Canada, August 5-9, 1991.
 661

Distinction between Planning & Scheduling

In the chemical processing context, production planning and scheduling collectively refer to the
procedures and processes of allocating resources and equipment over time to execute the chemical
and physical processing tasks required to manufacture chemical products. Usually, the production
planning component is directed at goal setting and aggregate allocation decisions over longer time
scales measured in months, quarters or a year, while scheduling focuses on the shorter time scale
allocation, timing, and sequencing decisions required to execute the plan on the plant floor. The
division between these two decision problems is to a large degree a matter of tradition, reflecting
the conventional hierarchical distinctions between corporate level planning and plant or production
line level decisions to execute those plans. The difficulty with that division has been the absence
of effective methods for aggregating/disaggregating information and decisions originating at one
level and exploiting them at the other level. The conventional solution to this problem of
coordination of the planning and scheduling levels in the presence of uncertain information has
been to employ the rolling horizon strategy under which planning is performed for several time
periods in the future, scheduling is performed for only the first period of the plan, reconciliation
of demands and output are made when the period is completed, and the process is repeated.
While it is often convenient to make reference to various decision levels m
planning/scheduling, the trend in computer integrated manufacturing is to fully integrate all levels
of decision making, thus focusing on the interactions rather than the distinctions between levels.
Therefore, for purposes of the present discussion, we will not differentiate between planning and
scheduling, refening to the whole as the scheduling problem. We simply recognize that scheduling
can be performed over different times scales from planning a campaign of runs, to developing a
master schedule for a given campaign, to modifying the master schedule to respond to changes
and unexpected perturbations which inevitably arise as the schedule is executed.
Scheduling is required whenever there is competition among activities for limited resources
which are available over a finite time period. It involves three key elements: assignment of
resources, sequencing of activities, and determination of the timing of the utilization of resources
by those activities. The assignment component involves the selection of the appropriate set of
resources for a given activity. The sequencing component concerns the ordering of the execution
of activities assigned to resources, while the timing component involves the determination of
specific start and stop times for each ofthe activities undergoing scheduling. Consider a plant with
two reactors and six different products each requiring a batch to be produced. The assignment step
 662

may involve selecting the first reactor for the first three batches and the second reactor for the
other three. The sequencing step would involves electing the order in which the batches are to be
produced on each line and the timing component would, of course, involve the determination of
the exact start and stop times for the production of each batch. These three components of
scheduling lend themselves to mathematical description and computer based solution approaches.
Indeed, scheduling techniques have been extensively investigated as a core area of the operations
research domain since the 1950's, as documented in the classical books by Baker [2], Coffinan et
al [11], and the text by French [19]. Because the need for scheduling arises in so may different
commercial and manufacturing contexts, various forms of the methodology have been widely
investigated in many fields of engineering, computer science, management science, and business.
The literature is thus quite large, diverse, and diffuse. In this paper, we will primarily focus on that
portion of the literature most relevant to batch chemical processing applications.

Stimulus for Scheduling in the CPI

In contrast to the long term concern with scheduling in the discrete parts manufacturing industries,
the attention devoted to computer aided scheduling methodology in the chemical processing
industry and the chemical engineering profession is much more recent, beginning in the middle
1970's (see reviews [56] and [35]). The principal exceptions are refinery scheduling and
distribution logistics LP applications which were already being investigated in the 1950's. The
accelerating interest in scheduling methodology in the chemical processing context, particularly
in the last decade, has been stimulated in part by major changes in the business environment faced
by the industry as well as changes in the information and computer technologies which can be
deployed to solve these highly complex decision problems. We briefly review these factors since
they provide the environment within which advances in the field of chemical process scheduling
must be pursued.
First, within the last decade, the CPI of the industrialized nations has changed dramatically
and, perhaps irreversibly, as enterprises have recognized that they must operate on a world-wide
basis to be competitive. Manufacturing operations must be globally coordinated, production must
be highly cost-efficient with tighter inventory control and high capital facilities utilization. The
quality movement in the industry has focused on responsiveness to customer needs, predictable
product delivery, and consistency in meeting product specifications. Furthermore, the strategic
plans of the leaders of the industry have identified more sophisticated chemical and biochemical
 663

products, with high value-added and short product life-cycles as the future of the CPl. The
increased tailoring of products to specific customer needs has led to more product types and
grades and lower production requirements for individual products. This has led to increased
sharing and coordination of production resources to meet the needs of multiple products and less
reliance on dedicated facilities. As a result, multi product batch and semi continuous operations
have become, if not more prevalent, (they have been historically widely used throughout the

industry) more central to corporate long term business plans. Shared production facilities are no
longer a temporary expedient to be replaced by a continuous dedicated plant when the market

expands, instead, they have become a permanent reality.

Secondly, rapid changes in information technology have made essential enterprise information
quickly and widely available. Automated, on-line plant information systems, no longer mere
vendor novelties, are available to provide instantaneous equipment status reports. Computerized
data base systems for handling customer orders, inventory status, and work in progress have made
timely and informed order processing decisions possible. The ClM movement in the discrete parts
manufacturing industries has drummed home the message that the electronically available
enterprise information can and should be used to do something more than generate management
reports and archive hourly, daily, or monthly production averages. Since unit level control systems
have already been widely implemented in the chemical industry, it has become abundantly clear
that the next levels of application lie in supervisory control and scheduling.
The advances in information technology have, if anything, been outstripped by those in
computer systems. Continuously improving, low cost computing power and memory are beginning
to make established but previously too costly solution approaches for scheduling problems such
as minced integer linear programming or branch and bound methods, feasible for selected
applications. Developments in graphical user interfaces offer promise in making comprehensible
the extensive information required to capture the details of scheduling problems and make
software easier to use for non-programmers. The advent of affordable high performance desk top
computing make it possible to deliver effective interactive scheduling tools to plant personnel even
in smaller production facilities. Finally, the realities and prospects of advanced architecture
machines make feasible the rigorous solution of schedule optimization problems heretofore
deemed intractable.

Clearly, the stage has been set for significant and rapid developments in the exciting field of
process scheduling in the next five years. In the following sections, we will seek to define more
 664

precisely the directions of the most promising avenues. We begin with a more detailed examination
of the scheduling problem and its various components and forms.

The Scheduling Problem

In its most general fonn, the scheduling problem consists of an operating strategy, a set of plant
equipment, a set of resources such as operators, utilities, and materials, a set of product recipes
and product precedence relations, specifications of resource availabilities and product
requirements, and one or more criteria which must be optimized. In the following, we characterize
each of these problem components:

Operating Strategy

A key decision which underlies any scheduling application is the strategy which has been selected
for organizing the manufacturing process. As shown in Figure 1, [55], the choices of operating
strategies for batch processing can be represented in terms of a two dimensional space which is
characterized by degree of product recipe similarity and by the relative length of time of individual
production runs or campaigns. Plants which must deal with products which have low recipe
similarity and short runs (one or two batches) require different modes of organization than those
which have high recipe similarity and long runs or campaigns (dozens of batches). It is reasonable
to assume that the issue of selection of operating strategy will have been resolved at plant design
or at the time of the most recent retrofit. Thus, for scheduling purposes, it will be assumed that
the strategy has already been determined and therefore is a problem input. Additional more
detailed aspects of operating policy which are implied by specific choices of operating parameters
such as inventory targets and tardiness penalties also have to be treated as input parameters.

Equipment

The equipment items of the plant consist of processing units, storage tanks, transfer units, and
connecting networks.
1) Processing equipment: These are the individual units available to the plant, often grouped
into types which have equivalent processing characteristics. Equipment items are characterized
by a nominal processing capacity and, possibly, a range of capacities over which operation is
feasible. Batch equipment is characterized by volume and semicontinuous equipment by rates.
 665

Degree of Recipe Similarity

high medium low

short
Flowshop Jobshop

Relative
Campaign medium
Length

long
Multiproduct Multiplant Multipurpose

Figure 1. Operating strategy space

2) Storage: Intermediate, raw materials, and product storage tanks may be of different type
and are characterized by volume.
3) Connections: the individual plant equipment items are linked through a network oflines
which may be fixed or flexible. The network represents all feasible connections between
equipment. Equipment is often organized into groups consisting of units used simultaneously
(in-phase). Groups perfonning the same function are organized into stages, where groups in a
stage are used sequentially (out-of-phase) as shown in Figure 2.
4) Transfer units: Transfer lines are devices for moving materials between process equipment
and are characterized by a rate or possibly by an allowable range of rates. They may include
conveyors, blowers, manned transfer vehicles or even autonomous vehicles. Transfer lines are
associated with a specific set of connections, may be used on a shared basis, and thus may be
entities subject to scheduling.
All of the equipment items may have release times and lists of block-out times associated with
them. The release time for an item sets the time at which that item first becomes available for
scheduling. Block out times are time periods during which an item is unavailable for processing
because of maintenance or other requirements.
 666

Group 1
Batch r--------.
:00
1.4.7 •...
Batch
1
L ________ .J
Group 2
r--------.

F----.-.,: 0 D :. .---~
I
I
D
L ________ .J
I
I

0 D
Group 3
Batch r - - - - - - - - •
3.6.9 •.. ~ I
3.6,9, ...
I I
L ________ .J

Figure 2. Generalized structure of a stage

Resources

Secondary resources are globally shared production inputs which may be either renewable or
nonrenewable. Renewable resources are those inputs whose availability levels are restored to
original levels immediately after usage. For example, upon completion of a task, an operator
become available for a subsequent task. The most common renewable resources are manpower,
electricity, heating and cooling utilities. Nonrenewable resources are those that are consumed and
depleted in the process and thus must be replenished after usage. For instance, when a raw
material charge is made to a unit, the inventory of that material is reduced and can only be restored
through specific action, e.g., delivery of new raw material.

Recipe

A product recipe can be viewed as a directed network of processing tasks which must be
performed to manufacture a given product. The nodes are the tasks while the directed arcs denote
the precedence order among the tasks. While usually the network structure is fixed and
 667

independent of the system state, the possibility exists of conditional tasks, that is, tasks whose
execution depends upon a state dependent decision. For instance, a reprocessing task may be
initiated if a composition is outside of desired ranges or a devolatilization task may be executed
if the need arises for temporary storage of the output of a given task.

Task

A task is a collection of elementary chemical and physical processing operations which must be
completed in a serial fashion within a certain type or set of process equipment. Each task has
associated with it a required processing time, a set of material balance or size factors, a final state,
a stability condition, a storage condition, change over times, resource requirements, and resource
substitution relations. Each these components may depend on the equipment and resources which
are selected for that task.
1) Processing time: The task processing time (for batch operations) or rate (for
semi continuous operations) in general may be a function of the specific type of equipment
employed, the amount of material processed, the level of assigned resources, as well as of selected
state variables. The overall processing time of a batch task may contain transfer time components
to reflect the time associated with material transfer steps which may be batch size dependent.
2) Feasible equipment list: a set of equipment types which can be used to execute the given
task. In general, the list may be ordered by priority.
3) Size factor: the mass or volume which must be processed in a given task in order to
produce a unit amount of final product. In general, there may be a size factor associated with each
member of the feasible equipment list for a task.
4) State: The final state of a task is characterized by composition, phase, temperature, and
other appropriate system variable values.
5) Stability condition: The stability condition indicates whether the material is unstable and
thus must be immediately processed (Zero Wait, ZW), can be held for a specified period of time
(Finite Wait, FW), can be held indefinitely (Unlimited Wait, UW), or can be made UW upon the
execution of a further processing task (Conditionally Stable, CS).
6) Storage condition: The storage condition reflects the material holding options available
upon task completion: storage only within the processing unit (No Intermediate Storage, NIS),
storage only within an external storage vessel or specified set of vessels (Finite Intermediate
Storage, FIS), both of the above options (Conditional Finite Intermediate Storage, CFIS), or
 668

unlimited external storage (VIS).

7) Change-over times: At the completion of a task there is a clean-out time which is incurred
before the equipment assigned to that task can be employed to process another task. In general,
for a given task, the change-over time will depend on the successor task and the specific
equipment item employed.
8) Resource requirements: The resource requirements are the amounts or rates at which the
various resources which are necessary for the execution of the task are consumed. Although often
assumed to be constant, these requirements may vary over the duration of the task in a specified
fashion.
9) Resource substitution relations: The requirements for specific resources can sometimes be
met through several alternate means according to specified priorities. For instance, a task may
require one of a certain class of operators but, if none is available, a more highly skilled, swing
operator may be employed. Substitution relations can be expressed in the form of a set of ordered
lists where the first member of the list is the normal choice of resource type while the subsequent
members are the allowed ordered substitutes.
In general, each task will itself be composed of a series of individual chemical and physical
steps, such as filling, mixing, heating, addition of a reagent, reaction. cooling, decanting, addition
of a solvent, and emptying. Each of these individual steps may be defined in terms of a set of
specific characteristic values of the above eight components. Thus, the recipe description can
extend beyond the network of tasks to a detailed description of each task with multiple input and
output streams associated with each task.

Product Precedence Relations

Product precedence relations are directed graphs consisting of nodes representing the products
and directed arcs denoting the required (partial) order in which products need to be produced. For
instance, a given product may be an intermediate which is required in the production of one or
more other products and thus may need to be produced prior to those products.

Resource Specifications

.
The availability level of each resource type is a function of time, which may be constant over the
entire time period under consideration or could be a periodically repeating or a general piecewise
nonlinear function.
 669

Demand Specifications

Product requirements are given as a set of orders, where each order has associated with it the
identity of the product, the required amount, an earliest start time for the production of the order,
a due date, and a priority in case of tardiness.

Performance Measures

The most general measure of the effectiveness of a schedule is an economic one which includes
components such as sales revenue, task specific operating costs based on the consumption of
resources, labor, and materials, inventory carrying charges, change-over cost, and order tardiness
penalties. Often, direct plant performance measures such as time to complete all orders or the sum
of the tardiness of all orders are used for simplicity.

Scheduling Problem Solution

The solution to the batchlsemicontinuous plant scheduling problem will in general consist of three
components:
1) A production plan which will indicate the sequence in which the orders will be processed
and the time frames within which orders or campaigns will be completed.
2) The specific assignment of equipment items and secondary resource levels to individual
production tasks.
3) The detailed schedule of operations expressed as starting and completion times of each
task, the predicted distribution of inventory levels over time, and the detailed profile of resource
utilization levels over the scheduling period.
The specific details of these three components will differ considerably depending upon the
particular strategy under which the plant is operated.

Special Problem Forms

To facilitate the subsequent review of methodology, it is expedient at this time to outline some of
the more important special cases and forms of the resource constrained scheduling problem
(RCSP). The classification of these special cases can be made according to the presence or
 670

absence of active constraints on resources other than equipment and the mechanism employed for
dealing with uncertainty in the scheduling data. Within each of these subdivisions, the cases can
be further divided on the basis of the operating strategy which is employed. Finally, one can also
consider an extension of the general RCSP, the multiplant problem, in which scheduling
encompasses the coordination of a number of geographically distributed and interacting
production facilities.

Equipment vs Resource Dominance: While the resource constrained problem is clearly the more
general, there does exist a universe of applications in which the principal limiting resource is the
process equipment. This so called equipment dominant case has received the bulk of the attention
in the scheduling literature, including that devoted to chemical processing applications. The next
level of complexity is the case involving shared intermediate storage. The presence of shared
storage, for instance, a set of storage vessels which can be employed for material from several
different products and tasks, introduces the need to consider the impact of the availability of
sufficient storage resource on the start and stop times of tasks which will require storage. Since
the time over which storage is required is not fixed as a property of that resource or of the product
recipe but is driven by the needs of processing units, shared storage introduces a coupling between
processing units and tasks that is not present in the equipment dominant case. Conceptually, this
coupling requires that the usage of storage be checked at each point in time at which there is a
change in the assignment of an equipment resource. The coupling induced by multiple shared
resources makes such checking even more burdensome and contributes significantly to the
complexity of the general resource constrained problem.

Rolling Horizon vs Reactive: The classical scheduling problem form is that in which all ofthe
production requirements, equipment status, and resource availability profiles are assumed known
at the time the schedule is generated. In many manufacturing environments, however, the
equipment and resource availabilities change frequently and unpredictably, task processing times
are subject to variations, order information is amended and new orders accepted continuously.
Thus, while a master schedule generated based on a priori information is an important planning
too~ the schedule must be almost continuously updated and revised. In principle, these situations
can be addressed using two limiting approaches: a moving horizon rescheduling strategy and a
reactive scheduling strategy.
In the first approach, an initial schedule is generated and then a form of moving horizon
 671

approach is used in which the schedule is periodicaIly regenerated whenever a sufficiently large
number of new events or inputs are encountered. Rescheduling can be done using an a priori
scheduling procedure which only considers tasks which have not yet been completed. Each new
schedule generated in this fashion will generally result in some resequencing and reassignment of
equipment and resources and, thus, will lead to some degree of schedule "chatter" or
"nervousness" as orders with intermediate due dates are repeatedly juggled in successive
schedules. Such repeated schedule changes can be undesirable because, rather than mitigating the
effects of process upsets and order changes, they, in fact, induce further changes. For instance,
materials requirements plans, work assignments, shipping schedules, and other associated
supporting functions, which have lead times of their own, may have been established based on an
earlier schedule and could be significantly impacted upon rescheduling. In many chemical industry
applications, it is highly desirable or even essential to minimize undue changes in the plans for
preparatory tasks and supporting activities to avoid the associated costs and delays.
The second approach seeks to reduce system nervousness by using a master schedule as the
basis for planning followed by a reactive scheduling strategy which responds to each processing
variation, equipment or batch failure, or order change by appropriately readjusting the master
schedule in a least cost or least change way. A range of adjustment modes can be employed:
readjustment of only task start and stop times, reassignment of intermediate storage capacity,
reassignment of equipment items or resource types, and resequencing of tasks or orders. In
specific instances, one mode will be preferred to another. Thus, the selection of modes might be
made so as to minimize the overall cost of the change or to minimize the difference in the
completion times of orders between the master schedule and the revised schedule.

Operating Strategy Alternatives: The three limiting cases of operating strategies which have
received the most attention are the multiproduct plant, the network flowshop, and the
multipurpose plant.
As shown in Figure 3, the multiproduct strategy is employed for products with highly similar
recipes and long product campaigns. Since the equipment network structure is fixed and all
products essentially visit the same equipment in the same order, the scheduling solution consists
of only the first and third components. The production plan consists of the determination of the
order in which campaigns of individual products are made and their length (lot size), allowing
multiple campaigns ofthe same product in order to reduce inventory charges. At the lowest level,
 672

STORAGE

STRIPPER a:
I "'
a.
a.
SURGE ii:
f-
Ul

' ';". J" i

SURGE ~
Ul

Figure 3. Multiproduct plant

the detailed schedule simply reduces to a set of single product line schedules. Each line schedule
consists of the start and stop times for a sequence of batches of the same product sweeping
through the fixed equipment network.
The network flowshop is appropriate for the case with similar product recipes and short
campaigns. The schedule for this case is similar to the multiproduct strategy, except that there is
no campaign structure, rather each product batch is treated as a separately scheduled entity. Thus,
the schedule consists of the order in which the individual batches are started in the plant together
with start and stop times for each batch on each of the units that it visits.
The multipurpose plant strategy is used for products with dissimilar recipes and longer
campaign lengths. Schedules under this strategy will involve all three solution components: the
production plan, the production line configurations used in each campaign, and the detailed
production line schedules. The production plan identifies the campaign structure, that is, the sets
of products which are to be produced at the same time, the campaign sequence, and the campaign
durations. The production line configuration or campaign formation portion involves the
assignment of specific equipment items to specific product tasks. These assignments serve to
define the individual production lines which will be used in executing the campaigns. The line
schedules themselves again reduce to single product schedules. This organization is illustrated in
Figure 4. As shown, the first campaign involves products A and B, while the second processes A
and C. In each campaign the plant equipment is reorganized into two different production lines.
 673

I
I
-I
I
I
I
Product A I
Product A
1
I
I
I
1
I
Product B I Product C

CAMPAIGN 1
• CAMPAIGN 2

Figure 4. Multipurpose plant

Note that in both the network flowshop and the multiproduct cases it is not essential that all
products employ ail equipment. Specific paths may be assigned to specific products or product
families. For instance in the batch resin process studies by Ishikawa et al [25] and shown in Figure
5, products may follow five different paths which use units I through 7; 8 and 2 through 7; 9
through 16; 8, 10 through 16; and 8, 10, 11, 12, 17, and 16.
In practice, it is not uncommon to find mixed strategies in which the above limiting strategies
are combined. For instance, a portion of the plant may operate using dedicated lines, another
portion may be operated in the multipurpose mode, while the last portion may operate like a
multiproduct plant. Alternatively, the particular strategy used at one point in time may change as
demands and product mix change.

The Multiplant Extension: In many instances, products and major intermediates are made using
multiple plant sites which may be geographically widely distributed and may exchange
intermediates. However, each plant is operated under its own strategy and schedule. In this case,
the scheduling problem must include an extra decision level which makes the assignment of the
 674

Storage

batch unit K

Figure S. Example batch resin process

different products and intermediates to different plants, tracks inventories, and accounts for the
time losses due to shipping of materials between the plants. However, this extra decision level
must be closely coupled to the specific plant and production line schedules, since the production
rates achieved in the individual lines will affect the overall assignment of products to specific lines
in specific plants and vice versa.
Before we begin with a review of the approaches to solving these various forms of the general
problem, we briefly summarize some of the difficulties attendant to the scheduling problem and
its solution from both a computational and a practical point of view.

Impediments to Scheduling Applications

Scheduling applications have very large information demands and lead to optimization problems
of considerable computational complexity. Engineers attempting scheduling studies are often
stymied because of the great diversity of reported problem forms, specialized solution approaches,
and nonstandardized terminology, generally are handicapped by the lack of convenient and reliable
off-the-shelf solution tools, and can be hamstrung by organizational impediments.

Information Load: Large amounts of information are required for the specification of the
 675

problem: commercial data, equipment data, recipe data, and plant status information. This
information load directly contributes to the complexity of the problem and has materially impeded
the adoption of scheduling methodology in practice. Unquestionably, the key first step towards
routine computer assisted scheduling is having the appropriate database of current and reliable
plant and order status information.

High complexity and dimensionality: The resource constrained scheduling problem (RCSP) is
inherently a high dimensionality, mixed discrete/continuous domain optimization problem
(MINLP) which, in its general and even in most of its simpler forms, remains among the most
severe challenges for optimization methodology. Mathematical programming formulations for
resource constrained scheduling problems have been under study since the middle sixties but
solution approaches to problems of practical scope have in the past been stymied because of the
formidable combinatorial complexity of the problem.

Many forms and special cases: In practice, the RCSP exists in many different forms, with very
particular structural constraints and site specific operating restrictions and rules. As noted above,
the research community in turn has proposed for investigation a number of idealized problem
forms as a means of identifYing and exploring elements of solution procedures for more realistic
forms of the RCSP. Unfortunately, these two sets of problem forms have not always enjoyed
significant commonality. In addition, a very diverse, nonstandardized terminology has arisen for
describing problem features. These factors have led to undue confusion and entropy in the field
and have served to discourage practitioners from exploring and exploiting the substantial
knowledge base that does exist.

Paucity of quality software: Although in recent years there has been much research on solution
approaches, the development of high quality process scheduling software has lagged considerably.
Available commercial quality scheduling software typically only offer LP, at best MILP, or
heuristic dispatching/assignment rules as solution vehicles. MILP and even heuristic methods are
quite restricted in the size of applications which can be successfully attacked. Several good
graphical display tools to support scheduling have been announced recently but these either leave
the main scheduling decisions to the user or offer only very simple heuristic schedule construction
mechanisms.
 676

Organizational barriers: A team undertaking a scheduling study often encounters considerable

impediments arising from the need to cross various traditional jurisdictional boundaries between
the corporate planning and logistics functions, the sales and marketing arm, plant management,
and direct plant supervision. Management support can be difficult to garner because the benefits
from better scheduling are difficult to predict a priori. Operating supervisors are often too busy
fighting fires to invest the time on another plant wide application and may be reluctant to allow
that the operation, in fact, requires improvement. Operating staff are suspicious of "help" to
improve their productivity, viewing such efforts as interference in their well-established routines.
Although much remains to be done, these difficulties can all be mitigated sufficiently so that
at the present time, using existing methodology, practical scheduling applications can be
expeditiously implemented and made operationally successful. However, the implementation of
such applications does require strong management support and a fair level of expertise in both
scheduling problem formulation and its solution.

Status of Methodology

In this section, the main approaches proposed for the solution of various forms of the general
scheduling problem will be reviewed. No claims are made of providing an exhaustive catalogue
of all references on the subject. Instead the focus is on contributions which are representative of
the available approaches to solving an interesting class of problems and are relevant to chemical
processing applications. We first consider deterministic scheduling, beginning with the equipment
dominant case and continuing with the resource constrained case. This is followed by a discussion
of reactive scheduling approaches.

Equipment Dominant Scheduling

The literature on deterministic scheduling of equipment dominated plants can be separated into
approaches dealing with the network tlowshop, the multiproduct plant, and the mUltipurpose plant.
By far the greatest portion ofthe literature has focussed on vanous forms of the flowshop. The
three limiting forms of this problem, the single unit problem, the simple serial tlowshop and the
singlc stage parallel system have been studied quite extensively as descnbed in the survey papers
by Gupta and Kypansis 124], Graham et a1. [21], [Graves, [231, and Lawler, et a1. [40]. Results
from complexity theory [20] indicate that most instances of these limiting problems belong to the
class of NP complete optimization problems. Consequently, it is conventionally argued that
 677

effective solutions to these problems and the more general network flowshop can only be expected
via heuristic methods. This generally has been the case, although recent work by Miller and Pekny
[44] with the classical traveling salesman problem offers clear evidence that rigorous solution of
NP complete problems is possible for problem instances of practical size by careful exploitation
of problem structure, effective bounding procedures, and the use of parallelization. As shown by
Pekny et a1. [52], the single machine shop and the serial flowshop under zero wait restrictions,
both with sequence dependent change-over costs, transform to asymmetric traveling salesman
problems and, thus, large instances of both problems (100 plus batches/jobs) could be routinely
solved. The ATSP solution approaches developed by Pekny and Miller are now routinely used by
DuPont to sequence processing lines to considerable economic advantage.

Network Flowshop: The general network tlowshop problem has to date received relatively little
attention, however, various special cases have been investigated, especially in recent years. Recall
that under this operating policy it is assumed that each product is only produced in a few batches
and, thus, scheduling involves consideration of each batch as a separate entity. Multiple batches
of the same product are thus effectively treated no differently than those of other products.
Furthermore, since under this strategy all products are assumed to follow the same recipe
structure, then under the assumption of no reuse of any equipment item by different tasks of the
same product, tasks of a recipe and the stages of the network become equivalent structures.
Consequently, the task stability and storage conditions are merged and described as properties of
the stages of the equipment network Thus, for example, the illS, NIS, ZW and FIS conditions are
all treated as stage and/or network properties.
The available literature on network flowshop scheduling can be divided into efforts focussed
on networks operating under illS or ZW conditions and those operating with finite storage.

UISIZW Networks: Most of the work in this area has been confined to networks in which the
parallel units in a stage are identical and are employed out-of-phase. The first effort in this
category is that due to Salvador [62], whose investigation appears to have been motivated by a
nylon manufacturing application. In this work, an exact algorithm for minimizing makespan for
ZW networks with identical processors within each stage was proposed. A branch and bound
algorithm was used to determine the optimal sequence, while a dynamic programming subproblem
was solved at each node to determine the completion time. No computational experience was
reported, but clearly the computation time can be expected to be severe for problems of practical
significance.
 678

Kuriyan and Reklaitis [37,38] have investigated approximate algorithms for the network
flowshop operating under UIS or ZW conditions, with the objective of minimizing makespan. A
two level approach was proposed which decomposes the problem into a simplified sequencing
subproblem followed by a sequence evaluation subproblem. The batch processing sequence is
determined using heuristic sequencing algorithms drawn from those available for simple serial
flowshop and parallel network problems, namely, dispatching rules, bottleneck sequencing, local
search sequencing, and best fit sequencing. The sequence evaluation subproblem is solved using
two different strategies: one product at a time (PAT) and one stage at a time (SAT). In the former
case, a product (batch) is assigned to a unit at every processing stage before another product
(batch) is considered. In the SAT scheme, every product is assigned a unit at a stage before
assignments for subsequent stages are determined. Computational tests with a broad series of
numerical test problems in which 20 products were scheduled on 2, 3, and 4 stage networks,
"hourglass" networks, and parallel networks with randomly selected processing times, support the
following conclusions. First, the best fit and local search procedures are the most effective
sequencing procedures. Second, the quality of the initial sequence can significantly affect the
performance. Since the difference between the PAT and SAT methods are not significant, the
former is recommended because it can be used for both UIS and ZW networks.
Citing the results of a survey of 20 companies, Musier and Evans [45] proposed the
investigation of multistage network flowshops with intermediate storage between each stage of
parallel units, operating under the objective of minimizing total tardiness of orders. As first step
towards the solution of this form of the network flowshop problem, an approximate procedure,
called the Heuristic Improvement Method, was developed for dealing with the single stage parallel
system. The procedure considered all single and pairwise interchanges of products (batches) to
improve an initial sequence. Results were reported for problems with up to 100 batches and 12
parallel units with a maximum deviation of 1396 from optimality. The performance of the method
declined as the number of batches and parallel units increased, presumably because of the
significant growth oflocal optima (with respect to pairwise exchanges).
The extension of the method to multistage processes with UIS between stages and the total
tardiness criterion has been reported by Musier and Evans [46]. The extension involves the
repeated application of the single stage heuristics in a reverse flow direction for a specified number
passes through the network. As is true with virtually all heuristic approaches to parallel and
 679

network flowshops, the method can readily accommodate nonidentical parallel units and restriction
of units to specific products. Computational results with randomized problems involving four
stages with up to seven units per stage, up to 280 batches, and integer problem data indicated that
good solutions can be obtained, although clearly optimality can not be claimed. Overall the quality
ofthe solutions is difficult to assess since no solution lower bounds are reported.

Flowshops with Finite Intermediate Storage: Most batch chemical plants operate under finite
storage (FIS) capacity constraints and may require different storage conditions at each stage of
the process. Storage represents the most direct example of a shared resource and thus
investigations of even simplified forms of scheduling problems with storage restrictions are
important for insights that they offer to the general RCSP. To date most of the finite storage work
has focussed on the simple serial case.
Early work on FIS scheduling addresses the two unit serial system, where storage capacity
is measured in number of batches. Dutta and Cunningham [15] proposed an exact algorithm using
dynamic programming. However, because of heavy computational demands, two approximate
methods were also developed. Papadimitriou and Kanelakis [51] showed that the two unit FIS
problem is NP-complete and proposed a two step solution strategy in which the product sequence
is determined by solving the ZW problem and the FIS schedule is determined using an unspecified
completion time algorithm. Wiede et aI [77], developed an optimal polynomial time algorithm for
the completion time calculation of the two-stage FIS case. When combined with approximate
methods for determining product sequencing using the related UIS or ZW problem, a two step
solution approach resulted which yielded makespans on average within 2.4% of the optimum
makespan.
Wiede and Reklaitis [75] further investigated the multi-unit FIS problem where the available
storage capacity is shared by all stages of the production line. An approximate solution method
which finds the storage assignment resulting in the earliest completion time for each product was
proposed for the makespan minimization problem. Products were scheduled one at a time but in
the worst case the storage feasibility check required multiple passes thus raising the computational
complexity of the completion time procedure to O(M2N2) where M denotes the number of stages
and N the number of products. Wiede and Reklaitis [76] extended the basic logic of this approach
to accommodate the Mixed Intermediate Storage (MIS) case in which different stages operate
under different ZW, NIS, UIS, or FIS strategies. Ku and Karimi [32] reported a completion time
 680

a1goritlun which reduces the computational complexity to O(MN) by only making feasible storage
assignments and thus eliminating the storage feasibility check. This approach was extended to
accommodate the mixed intermediate storage case. Ku and Karimi [34] also examined the less
general FIS problem in which finite but dedicated storage units occur between process stages as
well as the CFIS case. An exact solution approach using recurrence relations for the completion
time calculation and an MILP for sequence determination were proposed. Because of excessive
computational demands, an approximate approach was also developed using local search
procedures. These ideas have been extended to the MIS case with transfer and set-up times by
Rajagopalan and Karimi [54].
The only study to date to deal with the scheduling of multi-unit flowshops under a tardiness
based performance function is that ofKu and Karimi [33]. In this work, a branch and bound
solution method is proposed for the UIS serial problem with the goal of minimizing weighted
tardiness. Computational costs for this approach were found to increase exponentially so that
solution of problems with more than 15 products was found to be impractical.
Finally, a study of a specific network flowshop consisting oftwo stages of parallel processors
with nonidentical units and finite intermediate storage was reported by Kuriyan et al [39]. In this
work, the storage contents was updated in terms of actual quantities rather than as integer number
of batches. The problem was treated by using a simplified model for sequence determination and
a discrete/continuous plant simulation to predict the completion time of the schedule. It was
shown that the best fit heuristic using as initial list the products ranked by decreasing order of
differences between first and second stage processing times was most effective among a suite of
sequencing heuristics. The study demonstrated that a simulation can be effectively used as a
practical completion time calculation mechanism which can readily accommodate plant specific
constraints and monitoring of in-process inventory levels.

Multiproduct Plant: The key distinction of the multiproduct plant from the network flowshop is
that many batches of each product are assumed to be required, thus operation is organized into
longer campaigns. The principal advantage oflonger campaign operation is that it can be assumed
that, within a given campaign, production will occur in regularly repeating patterns. Such
repeating patterns obviate the need for scheduling of individual batches. Instead, the structure and
parameters of one instance of the repeating pattern can be used to characterize the production rate
for the entire campaign. Thus, the scheduling problem is reduced to determining the characteristic
 681

parameters of the repeating batch processing pattern for each campaign and the duration of the
individual campaigns. The available literature can be divided into those approaches in which the
repeating pattern involves batches of the same product (single product campaign) and those in
which the repeating pattern involves batches of more than one product (mixed campaigns). Within
each category, the work can be differentiated based on whether or not the determination of
campaign length is explicitly considered as part of the scheduling procedure.

Single Product Campaigns: The classical form of the multiproduct plant involves operation in
long single product campaigns where within each campaign batches are processed in a periodic
fashion described by a characteristic cycle time and batch size. The repeating pattern is thus one
batch of the given product. The ratio of batch size to cycle time for each product constitutes the
production rate for that product and therefore campaign lengths can be determined from product
demands by simple calculation. If sequence dependent change-over losses must be considered,
then the order of the campaigns can be determined by using single facility sequencing methods
[24] such as solving a mixed integer linear programming formulation. These models can also
readily accommodate inventory considerations and, thus, lot sizing decisions, that is, breaking of
long single product campaigns into several shorter campaigns so as to optimally balance inventory
holding costs against change-over costs.
The key parameters of batch size and cycle time for each product are readily computed for
the simple case of constant task/stage processing times, no reuse of equipment items by multiple
tasks of the same product, and identical parallel groups for each stage. However, in the presence
of batch size dependent processing times and transfer rates and unequal groups, as first noted by
Mauderli [43], the batch size and cycle time ofa batch can become dependent upon the particular
set of groups that are assigned to that batch. If the groups in each stage are used in a repeated
cycle, with each group used once in each cycle in fixed order, then batches will be produced
sequentially along NP paths of the production line, where NP is the least common multiple ofthe
number of groups in each stage. The repeating pattern in this case becomes the sequence ofNP
paths called a simple path sequence. In general, depending upon the specific order in which groups
on the different stages are assigned to batches, different simple path sequences can be generated,
each with a different overall average processing rate. For instance, in the four stage example
shown in Figure 6, the second and third stages each contain two groups, the Band C units
constituting equipment used in-phase. Since any simple path sequence will contain two paths,
 682

Path 1
Path 2
Path 3
Path 4

Figure 6. Path sequencing

paths 1 and 4 and paths 2 and 3 constitute two simple path sequences.
Wellons and Reklaitis [69] developed a mixed integer nonlinear programming formulation for
the problem of determining the simple path sequence for the general single product production line
which maximizes the average production rate ofthe line. Since combinatorial analysis showed that
the formulation is subject to a large number of degenerate solutions because of rotational and
operational equivalences, (Wellons and Reklaitis [70]), an explicit enumeration procedure was
developed for identifying only unique simple path sequences. Furthermore, a reformulation of the
MINLP was reported which restricts the feasible region of the model to distinct path sequences
and is much more efficient for larger applications, (Wellons and Reklaitis [73]). Interesting general
observations derived from this study are first that the production rate of lines with unequal groups
can be significantly increased through the use of simple path sequences as the basic repeating unit
for the campaign. Second, in the presence of variable processing times, the optimum batch size
for a line, whether using a single batch size or path dependent batch sizes, can in general be less
than the maximum batch size allowed by the capacity of the units in the line. Thus, significant
production rate enhancements can be obtained by proper choice of repeating batch pattern and
optimized batch sizes.

Mixed Product Campaigns: All noted by Wittrock [78] in the case of flexible manufacturing
systems, the productivity of a flowshop can be increased by replacing a series of long single
product campaigns with combinations of batches of several products which are repeated in a
 683

periodic fashion. Specifically, consider three products, A, B, and C, which are to be produced in
a three stage network in campaigns of three batches each (Birewar and Grossmann [4]). If the
time-limiting stage is different for each product and ifthere are no time losses due to changeovers,
then as, shown in Figure 7, production in a cycle of one batch each of products C, B, and A,
results in a reduction in the overall makespan to 42 hours in the case of ZW operation and to 38
hours in the case of VIS operation. Of course, if time losses due to change-overslclean-ups are
large enough, then the benefits of this so-called mixed product campaign (MPC) mode can be lost.
The implications of this strategy on the design and operation of multiproduct plants were
investigated by Birewar and Grossmann [4], who considered the plant with a single unit per stage
and Cerda, et al [8], who report heuristic procedures to guide the introduction of parallel units.
In the former work, it is shown that for sufficiently large numbers of batches, makespan
minimization can be replaced by overall cycle time minimization. MILP formulations suitable for
determining the optimal cycle which included one batch of each product were reported. This work
C A • B
Staoe I

SlagO 2

SIag83

44Hr

SIag8 I

Slago 2

SlaI/e 3

I b) 42 Hr

CA3 CA3

~
Slagel

SlaI/e 2
411
SIag8 J
I_I
11 Hrs

iel J8 Hr

= = = ......
PrcaUCl A Proaucl 9 ?!COUCt C Inllrmediate Storage

Figure 7. Scheduling with alternate cycle structures

 684

was extended in [3] to allow consideration of more general repeating patterns involving multiple
instances of batches of the same product by using a representation in terms of pairs of consecutive
batches of products. This led to a compact LP formulation for determining the number of pairs
of each type present in the optimal cycle and to graph constructions for devising specific cyclic
schedules. While the work reported to date is effectively confined to single unit process trains and
cases where the number of products is small compared to the total number of batches to be
produced, the MPC concept merits further investigation as it may well be an effective production
strategy for new batch plant concepts such as plants using autonomous transfer vehicles for which
change-over costs are less significant.

Scheduling with Lot Sizing: In the classical multiproduct case, campaign lengths are governed by
aggregate demands and, thus, only the order of campaigns is significant. The combined slack times
and clean-out times between each potential pair of successive campaigns can readily be
precomputed given the cycle schedules for each campaign. Thus, the optimal campaign sequence
can be obtained as the solution of a modestly sized integer program. In the more general form of
the multiproduct plant problem, inventory costs are considered and processing occurs in response
to orders which have associated due dates. Thus, consideration of mUltiple campaigns of the same
product and choices of campaign length (often referred to as lot sizing) must be added to choices
of campaign sequences. Only a limited amount of work has been reported to date on this more
general form of the problem.
Specifically, Birewar and Grossmann [5] considered the case of one or more nonidentical
parallel production lines, consisting of a single unit per stage and processing multiple products.
A multi-time period mixed integer programming formulation was presented in which the integer
variables are the number of batches of a given product which are to be produced on a given line
in a given time period. Average inventory levels over a time period were considered but
change-over and set-up costs and times were neglected. The formulation accommodated both
single product campaigns and mixed product campaigns by embedding the MPC formulation of
[3]. If the integrality condition is relaxed, then the problem reduces to an LP. A two phase solution
strategy is, thus, proposed in which the relaxed problem is solved first, the variables violating
integrality conditions are rounded down to the nearest integer and the LP is resolved with the
integer, variables maintained at fixed values.
Musier and Evans [47], investigated a single stage process consisting of nonidentical parallel
 685

units taking into account order quantities and due dates, inventories of finished products, and
sequence dependent clean-out requirements between products. The scheduling goal was to
determine the assignment of orders to units and the sequencing of orders on units so as to
minimize the sum of the times during which the net inventory of a product is negative. The
solution approach involved several heuristic components. First, an estimate was made of the
maximum number of batches of each product which need to be produced. Next, an initial feasible
schedule was determined using a composite best fit heuristic. Finally, the resulting initial schedule
was subjected to a local improvement procedure, consisting of single batch repositioning and
pairwise interchanges of batch positions. Each heuristic improvement candidate was tested for
inventory feasibility and the process was terminated when a local optimum is reached. Multiple
randomly selected starting points were used to improve the likelihood of finding the global
optimum to the problem. Computational results showed that the number oflocal optima increased
with the number of batches being processed and thus the likelihood of attaining global optima
decreased dramatically with problem size. The main limitations of the approach are the restrictive
network structure and the total stockout time criterion which does not differentiate between
products/orders or the amount of the stockout.
Recently, Sahinidis and Grossmann [60] addressed the closely related problem of cyclic
multiproduct scheduling of semi-continuous parallel lines for constant product demands, It is
assumed that production occurs on nonidentical parallel lines, each operated using a characteristic
fixed cyclic schedule which involves a subset of the products. A cost based objective function is
used which accounts for production, sequence dependent transition, and inventory holding costs
as well as a credit for freeing-up capacity. The result is an MrNLP which, after reformulation to
eliminate some inconvenient nonlinearities, is solved rigorously using a Benders Decomposition
approach. The key assumptions of the approach are that the production lines share no resources,
hence operate independently, that demand rates are constant over time, and that inventory costs
can be computed independently for each production line. The key assumption which avoids the
need to discretize time is that of cyclic operation. In that sense, the representation of time is
analogous to that used in the mixed product campaign case.

Finally, Kudva et al. [36] consider a more general form of the network flowshop, shown in
Figure 8, with multiple nonidentical units in each stage in which not all products can be processed
on all units. The formulation accommodates intermediate product draw-offs, raw materials feeds
to any stage, blending of intermediates to form specific final products, finite intermediate storage,
 686

feed

Unit
C

Unit
o

Figure 8. General network flowshop

and order deadlines. The solution strategy uses a fixed discretization of the scheduling horizon and
a two step approach to schedule generation. In the first step, orders are scheduled sequentially
according to priorities as close to due dates as possible whiie ensuring that inventories are
maintained above safety levels. In the second step, schedule heuristics are employed to improve
the schedule by aggregating orders and coalescing runs of the same task. Products are deblended
via a backwards component explosion and the required intermediates are scheduled as separate
orders, if not available from inventories. Different schedules are generated by changing order
priorities and evaluated using a cost based objective function which includes inventory costs,
changeover costs and deadline violations. The algorithm has been successfully tested on historical
data from an existing multiproduct plant, was found to give significantly better schedules than
those manually generated by plant staff, and is now implemented on the plant site for routine use.

Multipurpose Plant: The multipurpose plant problem is the most general form of the scheduling
problem. Production occurs in campaigns where each campaign involves one or more parallel
production lines and each production line itself will contain out-of-phase groups in a stage,
multiple in-phase units in a group, and, possibly, intermediate storage. Thus, planning of
campaigns, formation of campaigns, and scheduling of production lines must all be considered.
We will divide the review of the literature into those efforts which address only the production
 687

planning aspects of the problem and those which deal with all three levels of decision making. The
latter category is especially interesting because it represents an area of applications where
integration of all three levels is essential because of the strong interaction among the decisions
made at each level.

Production Planning: In contrast to the other operating strategies, in the multipurpose plant case
the production planning problem involves not only the selection of campaign sequence and length
so as to balance set-up and inventory charges but also involves the decision of selecting among
alternative campaigns so as to meet production needs in the most efficient way. The selection of
campaigns in effect is equivalent to assigning the production rates of the various products, where
these assignment are mutually constrained by the availability of a limited set of equipment. In
principle, to fully reflect the inherent flexibility of this operating mode, any production planning
approach must, thus, address all of these decision aspects. Unfortunately, many of the reported
formulations do not.
The production planning models proposed by Suharni and Mah, [65] and Rich and Prokopakis
[58] assume that a fixed set of equipment is used for each product and that the product batch size
and processing times are fixed. Suhami and Mah use a preprocessing procedure to assign due
dates and lot sizes for final and intermediate products given the due dates for the final products.
Based on the results of the preprocessing procedure, alternative production sequences are
generated and the tardiness of each schedule is determined by solving a linear program. The
schedule with least tardiness is retained.
Rich and Prokopakis [58] use an MILP formulation to sequence and schedule production
runs. A variety of scheduling objectives, such as minimize tardiness and minimize makespan are
investigated. A key assumption of their formulation is that every order due date must have
associated with it a corresponding production run, so that the campaign lengths are implicitly fixed
by the due dates. In effect, the due dates become the means of fixing the discretization of the time
domain. The main result of the solution of the MILP model is thus the sequencing of the orders
on the available equipment in the plant. Rich and Prokopakis [57] extend this approach to allow
production of a product over several different predefined sets of processors.
Both of these approaches concentrate on the precedence relationships among the products to
ensure the timely production of intermediates. This is indeed very important because in many
multipurpose plant applications, product recipes are long and thus it is expedient to break long
 688

recipes into several subrecipes associated with intermediates so as to allow the use of the same
equipment for several of the intermediates. This is, in effect, a means of reuse of the equipment
for several tasks ofthe overall product recipe. However, a major deficiency of both approaches
is that only one production line, and, thus, one production rate, is considered for the production
of each product While this does simplify the planning problem, it does neglect the inherent
flexibility of a multipurpose plant to accommodate a variety of production rates through a large
number of possible production line configurations.

Combined Planning and Scheduling: Only Mauderli and Rippin [42], Lazaro and Puigjaner [41],
Wellons and Reklaitis [71,72] have addressed all three components of the multipurpose plant
scheduling problem, especially the generation and evaluation of alternative campaigns from an
existing set of equipment items. In this section, we outline the essential elements of the three
approaches and note their limitations.
MauderIi and Rippin [41] employed evolutionary enumerative techniques to generate
alternative single product production lines. They first enumerated all possible equipment
arrangements with only one equipment group on each stage. After eliminating inefficient
candidates, these set-ups were then combined into all possible single product lines. Heuristic rules
were used to control the potentially explosive number of alternatives that were generated. The
effective production rates of these lines were determined assuming arbitrary sequencing of the
groups on each stage, ZW transfer between stages, and all batch sizes set at maximum path
capacity. The single product lines were then combined in an enumerative fashion to form a
combinatoriallY large number of alternative single and multiple product campaigns. An LP
screening procedure was then used to extract a much smaller subset of dominant campaigns. The
LP procedure obtains these dominant campaigns by identifying the campaigns forming the convex
hull of the set of feasible campaigns. A campaign is dominant if its vector of rates can not be
expressed as a linear combination of the rates of other campaigns. This concept is illustrated in
Figure 9 for two products A and B, whose production rates are shown on the axes. Campaign 2
is a dominant campaign since its production rates can not be obtained as any linear combination
of the rates of two other campaigns. Campaign 3 on the other hand is inferior as it could be
exceeded by campaign 2, for example. Given the set of dominant campaigns, a multi-time period
MILP production planning model is used to allocate the available production time in each period
to specific dominant campaigns. Constant change over times are assumed between campaigns and
 689

35
o d0a11nant. oa.m.pa1cn
30 - noniD.terior rat ••
o inrerior caD1paicn

2
:25
3::

:20
< 40
0.:::
:5
15

10

0
0 5 10 15 :20 :25 30 35

R
B
Figure 9. Illustration of dominant campaigns

the time losses due to start-up of the production lines are neglected.
Although this is the pioneering work on the subject, the work does have some notable
limitations. Specifically, since the campaign generation procedure is a heuristically controlled
partial enumeration, there are no assurances that the dominant campaigns are indeed truly
dominant over the set of all possible campaigns. Furthermore, the production rates determined for
the individual candidate lines are not optimized and, hence, may lead to poor campaign selection
decisions. Finally, the campaign screening procedure simplifies lost production time due to
changeover and start-up times by using average values, thus potentially leading to infeasible
production plans because the effective production rates may be either under or overestimated.
Nonetheless, the work is noteworthy for its recognition of the importance of all three decision
levels on the efficient operation of a multipurpose plant.
Lazaro and Puigjaner [41] used an exhaustive enumeration procedure to generate alternative
single-product production lines, allowing only in-phase operation of parallel equipment items on
a stage. A set of good production lines was chosen based on a heuristic selection index which
incorporates the production rate, equipment utilization fraction, and production cost per unit of
final product. Production planning was performed in a hierarchical fashion. A single product
planning problem was solved for each product to determine the number of batches to be produced
in each dominant production line in each planning period. The task sequencing and scheduling for
 690

each period were accomplished with a heuristic job shop scheduling algorithm [2]. Since the plant
is treated as a classical job shop, the solution procedure assumed VIS operation for all stable
intermediates and relaxes the requirement that products be produced in campaigns. However, the
job shop scheduling algorithm is able to take into account constraints on manpower and utilities.
Since the production plan for each product is determined independently, the planning problem is
applied recursively until a feasible production plan is achieved.
The main limitation of this work is that the scheduling procedure treats the plant like a job
shop and assumes unlimited intermediate storage. The resulting schedules can be inefficient
because out-of-phase operation of units is not considered or difficult to implement because
unfinished batches of several different products effectively have to be moved in and out of storage
as the plant produces several different products at the same time on shared equipment. The
campaign structure of multipurpose plant operation is thus obscured if not lost. Furthermore, the
production lines for each product are chosen on a greedy principle based on the value of the
selection index. Consequently, the lines selected for different products may require many of the
same units, making simultaneous production of these products difficult or inefficient.
Wellons and RekJaitis [71,72] revisited the multipurpose plant scheduling problem, developing
rigorous MINLP formulations for all three decisions levels and solving them using mathematical
programing techniques. The solution to the problem is approached via three components: the
single product campaign formation problem, the multiple product campaign formation problem,
and the overall production planning problem. The selection ofthe equipment items to form single
or multiple production lines and the determination ofthe most efficient schedule for each line are
tightly coupled decisions which do not lend themselves to simple decomposition. Thus, the
MINLP formulations for these problems include not only the 0-1 variables, which assign specific
items to specific equipment groups in the stages of production lines, but also incorporate the single
product scheduling formulation of Wellons and RekJaitis [69,70,73]. The production planning
problem on the other hand, relies only on the gross parameters of the dominant campaigns and
thus can be solved decoupled from the campaign formation and line scheduling problems. We
briefly summarize the salient features of each of these three component problems.
The single product campaign formation problem incorporates the following assumptions and
conditions:
1) The production stages may contain nonidentical groups
2) Each equipment item is used only once in a production line
 691

3) Batch aggregation is not allowed and batch splitting is only allowed within an equipment
group
4) ZW and NIS operating policies are considered but not intermediate storage
5) Transfer times fromlto batch units to/from semicontinuous units are accommodated
6) Processing and transfer times may be batch size dependent.
The problem is posed as an MINLP in which the equipment and path assignment decisions are
represented by three key 0-1 variables and the objective is to maximize the production rate. The
first variable type assigns an equipment of a given type to a specific position in a specific
equipment group on a specific stage. The second determines the number of equipment groups on
each stage and the third denotes the existence of specific paths of a simple path sequence. The
resulting problem is nonconvex and potentially has large dimensionality. A Benders
Decomposition approach is developed in which the complicating variables are the number of
groups in each stage. The two resulting subproblems are both MINLPs but one is significantly
smaller and the other has substantially fewer degeneracies. Comparisons with campaigns generated
in Mauderli's work shows that the proposed formulation yields better or equivalent results in all
cases tested, with improvements of up to 12.5% in the production rate.
The multiple product campaign formation problem involves the creation of multiple parallel
lines which have the highest production rate for a given product ratio. It is thus a multicriterion
problem in which one seeks campaigns such that the production rate of each line is as large as
possible. The desired set of solutions to this problem will be the dominant campaigns in the sense
of [41]. However, rather than generating a large number of possible campaigns and then
extracting from these those that are dominant, Wellons introduced the use of the Noninferior Set
Estimation method [12]. The NISE method allows the set of dominant campaigns to be generated
in a sequential fashion starting with the set of optimized single product campaigns by repeated
solution ofa suitably modified form of the single product campaign formation problem. As in that
case, the solution of the resulting MINLP is best obtained using a Benders Decomposition strategy
analogous to that developed for the single product case. Computational results using the Mauderli
and Rippin [41] problems show that the proposed approach yielded campaigns that dominate
Mauderli's results in 22 of38 cases, yielding as much as 20% improvement in the production rates,
and are equivalent to MauderJi's campaigns in the remaining cases.
The multi-time period production planning problem formulated by Wellons is an MILP which
selects from among the dominant campaigns those which maximize net profit subject to upper and
 692

lower bounds on sales. The formulation takes into account both the changeover time between
campaigns and the start-up time for each campaign. As shown in Figure 10, the change over time
is denoted by COT while the start-up time is denoted by su.
Clearly if the actual production time T ofa campaign in a given period is relatively short, then
change over and start-up times can be quite significant. The key 0-1 variable in the formulation
selects a given campaign for a given period. The number of these variables clearly grows with the
number of campaigns and, especially, with the number of multiple product campaigns.
There are three significant limitations of all three of these comprehensive approaches to the
multipurpose plant problem. First, the formulations do not explicitly include the decision of how
to subdivide the production of a product into several separately scheduled intermediates. Instead,
that decision must be made manually. Second, the structure ofthe individual lines does not permit
the inclusion of intermediate storage vessels which would allow changes in batch size in the train.
This can be done in principle using the approach of Yeh and Reklaitis [79] at the cost of some
additional complexity. The third and key limitation, however, lies in the need to construct the set
of dominant campaigns. In essence, the campaign formation component seeks to identifY the most
efficient campaigns for all of the different combinations of products and for all possible different
production scenarios. Given P products to be produced and allowing for the possibility of
campaigns involving from 1 up to P products, the number of distinct product combinations is 2P.

c:=:J c==J c:=:J

o 0 c:
.L

CO'J1Ac

c:=::J
CJ
c==..
/+--~-~'If--------- ----)
COT Be SUBe I TBep

CLop
Figure 10. Model of a multiple product campaign showing separate changeover and startup times for
each production line in the campaign
 693

In principle, a separate multiple product campaign fonnation problem would have to be solved to
identify the dominant campaigns from among these combinations. If the problem fonnulation is
extended to incorporate consideration of resource constraints, then the structure of the dominant
campaigns will also depend on the levels of resource availability. Thus, the campaign fonnation
subproblems would in principle have to be repeatedly solved for different resource level scenarios.
It is, thus, apparent that, in order to accommodate the consideration of limited resources, an
approach is needed which only fonns campaigns as and when they are required for specific
resource and production needs.

Resource Constrained Scheduling

The subject of resource constrained scheduling has seen limited, although increasing, attention by
workers in the chemical engineering research community but has been more extensively
investigated in the operations research and artificial intelligence communities. As noted earlier,
the resource constrained problem is inherently more difficult than the equipment dominant case
because in contrast to process equipment, resources such as materials and utilities are divisible.
Thus, in addition to sequencing of tasks in a temporal dimension, it is also necessary to consider
the feasible grouping of simultaneously executed tasks so as to jointly utilize resources in a feasible
fashion.

Chemical Process Oriented Contributions

Since the work of Lazaro and Puigjaner [41] who reported on the adaptation of a job-shop
algorithm for scheduling multipurpose plants with resource constraints, a growing series of papers
has emerged to address various forms of the RCSP. The work has largely resulted from the efforts
of research groups at ETH, Imperial College, and Purdue University.
Egli and Rippin [16] considered the problem of determining a short tenn schedule for a
multiproduct batch plant subject to resource constraints which meets a specified demand pattern
while minimizing costs associated with storage, changeovers, and utility consumption. The
scheduling procedure accounts for constraints on working patterns, limited availability of shared
resources, inventory requirements for intennediate products, and raw material deliveries. The
procedure is enumerative in nature: all possible production sequences are generated and then
progressively eliminated by imposing the problem constraints so that only the favorable sequences
 694

are fully evaluated. Feasible schedules are generated by shifting each batch schedule forward or
backward in time until all resource constraints are satisfied. Results are reported for a relatively
small example with four products and eleven equipment items. The work is notable for its
compilation of all relevant problem elements but offers a solution methodology of very limited
scope.
Tsirukis and Reklaitis [67,68] considered the resource constrained form of the multipurpose
plant scheduling problem and proposed a strategy which uses global approximation methods to
reduce the solution domain sufficiently so that math programming approaches can be more
effective. In this solution approach, the problem is decomposed into two main decision levels. The
first decision level is concerned with the efficient sequencing of actions that may not overlap in
time and, therefore, occupy disjoint intervals on the time axis. The second decision level is
involved with actions that must be executed within the same time interval. The first level
corresponds to the assignment of orders to campaigns and the allocation of a time frame for each
campaign. This decision problem is represented as a relaxed MINLP formulation which is solved
as a generalized Hopfield network. The second level involves the allocation of equipment and
production resources to the specific orders within a given campaign. These allocation decisions
are highly interdependent and constrained and are approached using a novel feature extraction
search scheme which serves to narrow the search domain, followed by detailed optimization over
the reduced decision domain. Figure 11 shows the hierarchical organization of the MBP
scheduling system. The MBP system has been used to successfully solve scheduling problems with
nine products, forty equipment items, and four resource types whose direct formulation would
involve up to 1120 binary variables [68].

CAMPAIGN
GENERALIZED

l
FORMATION
HOP FIELD
SUBPROBLEM
NETWORKS
(MINLP)

I
~
RESOURCE
FEATURE
AND
EXTRACTION I-
EQUIPMENT
ALGORITHM
ASSIGNMENT

Figure 11. MBP scheduling system

 695

Each task is allowed to have associated with it a constant processing time, a constant
resource requirement, and a set offeasible equipment. To model the use of limited plant resources
over time, the scheduling horizon is discretized into a number of uniform time intervals. The basic
time quantum for the discretization is selected to be sufficiently small so that the processing and
other characteristic times can be sufficiently accurately represented. Under this representation, the
key set of decision variables are those which control the allocation of a unit to a task in a specific
time period, that is, binary variable Wijt takes on the value 1 if unit j is assigned to processing task
i in time quantum t and zero otherwise. By assuming constant processing times and resource
utilization rates during the occurrence of a task, the material balance, resource utilization, capacity
bounds, and unit allocation constraints can be expressed in linear form. The result is a mixed
integer-linear programming problem whose solution can be sought using general purpose MILP
solvers or specialized branch and bound strategies which exploit the particular problem structure.
The authors have reported variations of the formulation to handle sequence dependent set-
up/clean-out constraints, several types of storage restrictions and both short term and campaign
type operating modes. Continuous processing step are treated by discretizing runs into discrete
increments.
The time discretization approach employed in this work is the classical means of dealing with
resource allocation problems in the presence of globally shared resources whose utilization is
distributed over time. The well-known difficulty with uniform time discretization is that the
number of integer variables required to represent practical problems with realistic processing
times, numbers of tasks, resources, and equipment can be quite large. The classical dilemma,
which then arises in practice, is the following. If the problem data is rounded so that the time
quantum can be increased ( and thus the number of decision variables decreased), then the solution
which will be obtained will be approximate and either too conservative or infeasible. If the problem
data are accurately treated, then the MILP can become too large for routine solution. In addition,
if sequence dependent factors are taken into account, the number of constraints grows quite large,
making the LP relaxations which must be solved as the branch and bound solution progresses large
and often resulting in large integrality gaps. On the positive side, the MILP framework is quite
flexible and offers considerable expressive power in accommodating application specific features,
such as equipment connectivity restrictions, finite shared storage, and conditional post processing
steps. To date investigations of various forms of single product, multiproduct, and multipurpose
 696

Prodllctl

10% 21"

IntBC

211r IlIr

F.,dC

Figure 12. State-task network example (after Kondili et al)

operations have been reported with successful solution of applications which are of modest size
in terms of the operational problem but large in terms of variable dimensionality (up to 4000
variables).
From a conceptual point of view, the large number of uniformly spaced time discretization
points is a modeling artifact which is introduced to simplifY the verification that resource
constraints remain satisfied throughout the scheduling horizon as task to equipment assignments
are made. Yet, assuming constant processing times and constant resource utilization rates during
task execution, the actual times at which checking of resource utilization is required is limited to
the points in time at which the resource availability changes and the start and stop times of the
individual tasks. These times correspond to the original input data of the scheduling problem and,
thus, ideally should be the only event times which ought to be explicitly considered. As shown in
Figure 13, the number of the data-related resource utilization event times and the model-related
event times can differ significantly. Effectively, the data-driven event representation leads to a
sparser, nonuniform discretization which can potentially be a much more efficient approach to the
RCSP.
An initial attempt to exploit this concept was given in [80,81] where the development of an
 697

Data-related: I I I IIII I II I
Model-related: 1111111111111 11111111111111111 11 1111 1111 111111111I11
Figure 13. Data and model related resource time events

enumerative search scheme for handling a general form of the short term scheduling problem is
reported. The key component of that approach is an interval processing framework which employs
tree structures to generate a nonuniform set of time intervals, one for each recipe task during
which the task may float freely and not violate resource constraints. Given this set of intervals and
the appropriate set of precedence constraints, the actual assignment of exact start times for tasks
can be accomplished through a linear program. The approach allows a very flexible treatment of
various forms of material stability conditions, intermediate storage, renewable and nonrenewable
resources, as well as detailed step descriptions for each task. However, the implementation using
the production system language OPS83 proved too inefficient for large scale application. In [82],
an MILP formulation of the interval processing framework is reported which overcomes most of
the limitations of the direct enumerative search approach. A preliminary comparison of the
nonuniform formulation with the conventional uniform discretization representation suggests that
for a significant range of problems substantial reductions in the number of 0-1 decision variables
can be achieved. Considerable work remains to be undertaken to develop efficient solution
algorithms for both formulations since with MILPIMINLP problems it is only through rigorous
computational experiments that questions concerning relative solution effectiveness can be
answered.

Operations Research Contributions: The operations research literature on the subject of resource
constrained scheduling is quite extensive. A comprehensive summary of the main avenues of attack
can be found in [7] and an updated review in [6]. As in the equipment dominated case, most of
the approaches reported address only severely simplified resource constrained problems and
employ either very large scale multi-time period MILP formulations or, in a few specific instances,
 698

polynomial time algorithms. The opportunities for direct transfer of solution approaches applicable
to the problem structures discussed earlier are, thus, limited.

Artificial Intelligence Contributions: Alternatively, the schedule for a resource constrained batch
process can be viewed as a sequence of actions in time that transform the plant from an initial
state to a final state in some optimal fashion. Under this perspective, the RCSP can be considered
as an instance of the planning problem which has been extensively studied by the artificial
intelligence research community. Planning research investigates the automatic generation of a
course of action that transform the world from an initial state to a desired goal state. The research
is focussed on two major problems: (i) the action representation in which a convenient medium
is sought for describing the interactions among problem variables and (ii) the action sequencing
problem where efficient solution methods are sought that can sequence the actions and produce
the desired result.
The most important action representation vehicle has employed first order logic. The world
is described by a set of logical variable whose interactions are represented by appropriate sets of
propositions. The actions are represented by operators that transfer the state of the world by
manipulating the truth value of the logical variables and the propositions. This action
representation was introduced by Fikes and Nilsson [17] and has been used with variations in most
subsequent efforts. The solution methods were initially based on intuitive rules that minimized the
differences of the actual world state from the goal state. However, the inability of these
approaches to guarantee feasibility let to more sophistical heuristic methodologies, such as
hierarchical decompositions, least commitment strategies [59] and temporal reasoning models [1].
More recently the planning problem was shown to belong to the class of NP complete
problems [9] and, thus, research has reoriented from seeking general systems for solving planning
problems to approaches tailored to more narrowly defined domains, such as discrete
manufacturing scheduling systems [18] and [50]. These efforts effectively are compilations of
problem dependent heuristic rules and intuitive decomposition strategies which are not transported
readily to other domains. Overall, the representational ability of planning methodology, linear or
nonlinear, is inadequate for representing chemical process scheduling problems which are
described by continuously varying quantities, discrete variables, and nonlinear relations among
variables. In addition, the RCSP requires solutions that are not only feasible but also, if not
optimal, then at least very good as measured by a performance function. It thus appears that the
 699

most effective approach to RCSP is the mathematical programming approach supplemented by

heuristics, approximation methods, and search strategies, as the practicalities of obtaining working
solutions dictate.

Reactive Scheduling

Although the need for frequent schedule readjustment has been actively discussed in the process
systems engineering literature [48], [35], and [74], concrete developments have been sparse.
However, the issue of rescheduling in a dynamic environment has been an active area of research
in the discrete parts manufacturing arena for some years and, thus, an extensive literature does
exist in that domain. We first consider contributions in the chemical processing domain.

Chemical Process Oriented Contributions: The first contribution to focus on the issue of reactive
scheduling of batch chemical plants is that ofCott and Macchietto [13 & 14]. This work reports
on the development of a system for real-time rescheduling of a batch plant subject to processing
time variations. The underlying methodology is a completion time calculation algorithm which
considers storage, processing units, and secondary resource limitations, and responds to
perturbations by readjusting the start and stop times of operations so that feasibility of the
schedule is restored. The start and stop time redetermination amounts to an earliest start heuristic
for the completion time calculation for a f1owshop under the restrictions that no task to resource
reassignments are allowed. Tongo and Reklaitis [66] presents a heuristic completion time
calculation approach for the resource constrained multiproduct plant. The formulation
accommodates splitting and merging of batches, multiple storage options, including CFIS,
product-unit combination dependent processing, transfer and changeover times, as well as order
release and due dates. The approach selects earliest completion of all orders while satisfying
resource limitations. Effectively, both developments can be viewed as simplified forms of discrete
event simulation. Discrete/continuous simulators, such as BATCHES [10], which are designed
to monitor resource requirements and proceed in the execution of tasks based on dispatching rules
such as earliest start, could in principle accomplish the same task, although with some additional
overhead. Unfortunately, any reassignment decisions which can be carried out within a simulation
are inherently myopic in nature and thus also suffer from the limitations of the earliest start
strategy.
 700

In addition to this fundamental limitation, these rescheduling strategies also do not consider
the full range of schedule modification alternatives, which are available in practice, from simple
time shifts to unit reassignments; to resource substitutions, and eventual complete rescheduling.
By way of example, Figure 14a shows two parallel two stage production lines, each processing
two product batches. Suppose that, as shown in Figure 14b, a processing time deviation takes
place on unit ul which leads to a time conflict in the schedule for unit u2. The simple time shift
solution to this conflict would be to delay the start of batch b2 on ul, thus disrupting preparations
for that batch which may have already taken place and propagating further delays to any
subsequent processing or product shipping steps. An alternate resolution of the conflict, which
causes no such disruption, is to assign batch bl to unit u3 which is idle during that time interval.
Of course, this may require an extra clean-out of unit u3, with its attendant costs.
A framework for balancing time shifts against unit reassignments which uses a beam search

c:::::::::
bl b1
ul ,I ,-,1
--I (a)

u2
,I ,,
,2· ts!

.,1
bl b2
113 ,1_
,I ,I
52
.4 .2

bl b2
sll
,
ul WU/4 ~.I (b)

.1_,
sl· it 1$2

b1 b2

,
113 • •1
I_ _ _ _ _
I _ _ 51
u4 51

b1 bl
ul .11 m C=:::sl
,II
(0)

,I
151
b1 b2
113 sl_511 • •1

,I_ _ _ _ _ _
,I _ _ _ .2
u4 51

C pIOductl • pIOduct1
m! pracessiDg lime dcYiaIion .. COIIflia

Figure 14. Conflict resolution (a) Original schedule (b) Simple shifting (c) Replacement
 701

with heuristic pruning was reported by in [30]. The approach employs the conflict
detection/conflict resolution paradigm under which a two level decision tree is constructed by
tracking each batch and tracing each task for each batch undergoing conflict until all conflicts are
resolved. The tree is pruned using heuristic pruning criteria as well as Zentner's multiple time
interval propagation technique for eliminating infeasible choices. The decision tree is subjected to
a beam search which seeks to minimize the weighted deviation from the master schedule. The
choice between time shifting or unit replacement is made based on the earliest completion of the
task which is subject to conflicting demands. Results reported using a three product, 51 batch, 26
equipment item test case show significant improvement over the use of a myopic criterion for
eliminating conflicts. The approach, nonetheless, suffers from the limitations of rapid growth in
computation time with increases in the beam search parameter and the suboptimal solutions which
can be introduced by virtue of the sequential nature of the batch conflict resolution mechanism.
Clearly, additional work is required to develop a reactive scheduling strategy which accommodates
the full range of rescheduling decision alternatives.

Discrete Parts Oriented Contributions: Nof et aI [49] provide a survey of literature on dynamic,
real time scheduling and rescheduling for job shops and f}owshops. Most of the systems developed
so far are directed at issues connected with integration of real time monitoring, scheduling, and
schedule modification in an interactive scheduling framework.
Ow et al [50] developed a system, OPIS, that performs reactive schedule revisions by
focussing on a short "conflict horizon". Based on the nature of the scheduling conflict, the conflict
is classified into a conflict class. Using a heuristic rule base, the system chooses from a set of
predefined remedial actions and applies orie such action to the schedule in the "conflict horizon".
These alternative remedial actions are investigated using a beam search. No attempt is made to
predict or account for the impact of a given decision on the remainder of the schedule. Rather, any
additional conflicts are merely imposed on the rest of the schedule. Kanai et aI [29] developed a
system which does reactive modification of a job shop schedule. The schedule is represented as
a constrained network of individual steps and a simple constraint propagation scheme is used for
propagating the effects of specific deviations through the rest of the schedule. Rerouting of jobs
in the face of machine breakdowns is not considered: all corrective actions are confined to time
delays.
Prosser [53] describes a scheduling system which has a combination of predictive and reactive
 702

components. The rescheduling problem is treated hierarchically in three layers. At the lowest level,
conflicts are resolved by time-shifting. If the system finds a conflict that cannot be resolved by
time shifting, a higher level tactical agent is engaged to attempt load balancing. If this fails to
resolve the conflict, control is passed to the highest level which employs job resequencing. A key
limitation of this system is that it chooses only one solution among the possible alternatives at a
given level and backtracks to the next level if that alternative fails. There are no provisions made
for interlevel backtracking. Grant and Nof [22] developed an automatic adaptive scheduling
system for a jobshop which similarly involves the simple components of time shifting and
resequeneing unscheduled orders. As deviations occur in the schedule over time, these are
accommodated through time shifting. As these shifts occur, the difference between the expected
and the actual completion times of tasks are monitored. When the difference exceeds some
tolerance limit, resequencing is invoked. With the exception of OPIS, none of these developments
perform a search over the decision space to find the best decision, rather, they proceed with the
first feasible alterative that is identified.
The concept of reactive plan modification has also been investigated by researchers in the AI
planning domain. For instance, [26,27,28] presents a formal theory of reactive plan modification
within the framework of hierarchical nonlinear planning. During the rescheduling process, an
attempt is made to minimize the disturbance caused to the applicable portions of the original plan.
At each node of the task network, the number of plausible choices is controlled using a heuristic
strategy which orders the choices with the aim of minimizing interactions. As in the basic
deterministic scheduling case, the main limitation of planning theory approaches derives from the
fact that the focus is entirely on obtaining feasible schedules rather than conducting a search for
the best (or at least very good) solution. Furthermore, an underlying premise is that enough expert
knowledge is available about the domain of interest for the development of a comprehensive body
of rules. In the case of highly combinatorial scheduling problems, the effectiveness of expert
knowledge can be severely limited.

Conclusions

In this paper, we have sought to define the essential elements of scheduling problems
characteristic of chemical manufacturing. The available methodology, especially the developments
of the last five years were reviewed. Much has truly been accomplished and a number of research
groups in the academic chemical engineering world now are active in the field. However, further
 703

work is clearly required. Specifically, further investigation is needed of alternative formulations

and solution methods for resource constrained scheduling problems, systematic treatment of
reactive scheduling within an overall framework for handling uncertainty in scheduling, and
effective approaches to the coordinated scheduling of multiple plant sites.

References
1. Allen, 1.F.: Maintaining Knowledge about Temporal Intervals. Communications of the ACM, 26, II, (1983)
2. Baker, K. R: Introduction to Sequencing and Scheduling. Wiley, New York, (1974)
3. Birewar, D. B., and I. E. Grossmann: Efficient Optimization Algoritluns for Zero Wait Scheduling of Multiproduct
Batch Plants. Ind. Eng. Chem. Res., 28,1333-1345 (1989)
4. Birewar, D. B., and I. E. Grossmann: Incorporating Scheduling in the Optimal Design of Multiproduct Batch
Plants. Comput, & Chem. Engng., 13 (112), 141-161 (1989)
5. Birewar, D. B., and I. E. Grossmann: Simultaneous Production Planning and Scheduling in Multiproduct Batch
Plants. -Ind. Eng. Chem. Res., 29, 570-580 (1990)
6. Blazewicz, 1., G. Finke, R Haupt, and G. Sclunidt: New Trends in Machine Scheduling. Europ. 1. of Operat. Res.,
37,303-317 (1988)
7. Blazewicz, J., W. Cellary, R. Slowinski, and J. Weglarz: Scheduling under Resource Constraints - Determinate
Models. Baltzer, Basel, (1987)
8. Cerda, 1., M. Vincente, 1. Gutierrez, S. Esplugas, and 1. Mata: Optimal Production Strategy and Design of
Multiproduct Batch Plants. Indust. & Eng. Chern., 29 (4), 590- 600 (1990)
9. Chapman, D.: Planning for Conjunctive Goals. Artificial Intelligence, 32, 333-377 (1987)
10. Clark, S. M, and Kuriyan, K: BATCHES-Simulation Software for Managing Semicontinuous and Batch Process
Paper 32f, AICHE National Mtg, Houston, TX (April, 1989).
II. Coffmann, E. G., Jr., (ed): Computer and Job/Shop Scheduling. Wiley, New York, (1976)
12. Cohon, 1. 1., R 1. Church, D. P. Sheer: Generating Multiobjective Trade -Offs: An Algorithm for Bicriterion
Problems. Water Resour. Res., 15,1001-1010, (1979)
13. Cott, B. 1., and S. Macchietto: A General Completion-Time Determination Algorithm for Batch Processes.
Presented at Annual AlChE Meeting, San Francisco, (Nov. 1989)
14. Cott, B. J., and S. Macchietto: Minimizing the Effects of Batch Process Variability Using On-Line Schedule
Modification. Compul. Chem. Engng., 13 (1/2), 105-113 (1989)
IS. Dutta S. K.,and A. Cunningham: Sequencing Two Machine Flowshops with Finite Intermediate Storage. Manag.
Sci., 21, 989-996 (1975)
16. EgIi, U.M, and D.w. T. Rippin: Short-tenn Scheduling for Multiproduct Batch Chemical Plants. Compul. Chem.
Engng., 10(4),303-325 (1986)
17. Fikes, R. E. and N. 1. Nillson: STRIPS: A New Approach to the Application of Theorem Proving to Problem
Solving. Artificial Intelligence, 2 (3/4), 189208 (1971)
18. Fox, M. S., and Smith, S. F.: ISIS - a Knowledge-Based System for Factory Scheduling. Expert Systems Journal
I, (I), 25-49, (1984)
19. French, S.: Sequencing and Scheduling: an Introduction to the Mathematics of the Job-Shop. Horwood, Chichester,
(1982)
20. Garey, M. R and D.S. Johnson: Computers and Intractability: A Guide to the Theory ofNP-Completeness.
Freeman and Co. (1979).
21. Graham, RL., E 1. Lawler, 1. K. Lenstra and A. H. G. Rinnooy Kan: Optimization and Approximation in
Deterministic Sequencing and Scheduling: A Survey. Ann. Discrete Math., 5, 287-326 (1979)
22. Grant, F. H., and S.Y. Nof: Automatic Adaptive Scheduling of Multiprocessor Cells. Presented at ICPR,
Nottingham, U.K, (August 1989)
23. Graves, S.C.: A Review of Production Scheduling. Operations Res., 29(4), 646 675 (1981)
24. Gupta, S. K., andJ. Kyparisis: Single Machine Scheduling Research. Omega Inl. 1. Of Mgmt. Sci., 15 (3),207-227
(1987)
25. Ishikawa A., H. Shinji and I. Hashimoto: Module-Based Scheduling Algorithm for a Batch Resin Process.
Proceedings of the ISN90 International Conference on Advances in Instrumentation and Control New Orleans,
LA, (Oct. 1990)
 704

26. Kambhampati, S. and lA. Hendler: Control of Refitting During Plan Reuse. Proceedings of International Joint
Conference on AI, 943-948 (1989)
27. Kambhatnpati, S.: A Theory of Plan Modification. Proceedings of AAAI Conference, 176-182 (1990)
28. Kambhampati, S.: Mapping and Retrieval During Plan Reuse: A Validation Structure Based Approach. Proceedings
of AAAI Conference, 170-175 (1990)
29. Kanai, N., S. Yakai, K Fukunaga, and Y. Tozawa An Expert Systern to Assist Production Planning. Proceedings
of International Workshop on Al for Industrial Applications, 219-224 (1988)
30. Kanakamedala, K B., V. Venkatasubramanian, and G.V. Reldaitis: Reactive Schedule Modification in Batch
Chemical Plants, Ind. Eng. Chern. Res. 33 (1),77-80 (1994)
31. Kondili, E., Pantelides, C. C., and Sargent, R.W.H.: A General Algorit1un for Scheduling Batch Operations.
Proceedings for the Third International Symposiwn an Process Systems Engineering, Sydney, Australia, 62-75,
(1988). See also, Comput. & Chern. Engng., 17 (2), 211-229 (1993)
32. Ku, H-M., and I. Karimi. Completion Time Algorithms for Serial Multiproduct Batch Processes with Shared
Storage. Comput. & Chern. Engng., 14 (I), 49-69 (1990)
33. Ku, H-M. and I. Karimi: Scheduling in Serial Multiproduct Batch Processes with Due-Date Penalties. Ind. Eng.
Chern. Res., 29, 580-590 (1990)
34. Ku, H-M. and I. Karimi: Scheduling in Serial Multiproduct Batch Processes with Finite Interstage Storage: A
Mixed Integer Linear Program Formulation. Ind. Eng. Chern. Res., 27, 1840-1848 (1988)
35. Ku, H-M., D. Rajagopalan, and I. Karimi: Scheduling in Batch Processes. Chern. Engng. Prog., 83 (8),35-34
(1987).
36. Kudva, G., A Elkamel, J. F. Pekny, and G. V. Reldaitis: A Heuristic Algorit1un for Scheduling Multi-Product
Plants with Production Deadlines, Intermediate Storage Limitations, and Equipment Changeover Costs. Fourth
International Conference on Process Systems Engineering, Montebello, Canada, (Aug. 1991). See also, Comput.
& Chern. Engng. 18 (9), 859-876 (1994)
37. Kuriyan, K, and G. V. Reldaitis: Approximate Scheduling Algorithms for Network Flowshops. PSE '85: The Use
of Computers in Chemical Engineering, IChE Symposiwn Series 92, 79-90, Pergamon, Oxford, U. K, (1985)
38. Kuriyan K, and G. V. Reldaitis: Scheduling Network Flowshops so as to Minimize Makespan. Comput. Chern.
Engng., 13, 187-200 (1989)
39. Kuriyan, K, G. Joglekar, and G.V. Reldaitis: Multiproduct Plant Scheduling Studies using BOSS Ind. Eng.
Chern. Res., 26,1551-1558 (1987)
40. Lawler, E. L.., Lenstra. 1 K, and Rinnooy Kan, A H G.: Recent Developments in Deterministic Sequencing and
Scheduling: a Survey. In M A H Dernster, 1. K Lenstral and A H G. Rinnooy Kan (cds.) Deterministic and
Stochastic Scheduling, Reidel Dordrecht, 35-73 (1982)
41. Lazaro, M., and 1. Puigjaner Simulation and Optimization of Multi-Product Plants for Batch and Semi-Batch
Operations. I. Chern. Symp. Series, 92, 209-222 (1985)
42. Mauderli, AM, and D. W.T. Rippin Production Planning and Scheduling for Multi-Purpose Batch Chemical Plants
Comput. Chern. Engng., 3,199-206 (1979)
43. Mauderli, AM : Computer-Aided Process Scheduling and Production Planning for Multi-Purpose Batch Chemical
Plants. Ph.D. thesis; E T H Zurich Nr. 6451 (1979)
44. Miller, D 1., and J F. Pekny: Exact Solutions of Large Asymmetric Traveling Salesman Problems. Science, Vol
251,pp.754-761,(1991)
45. Musier, R F. R, and L B. Evans: An Approximate Method for the Production Scheduling of Industrial Batch
Processes with Parallel Units. Comput. & Chern. Engng., 13 (Y2), 229-238 (1989)
46. Musier, R F. H, and 1.B. Evans: Schedule Optimization for Multi-Stage Chemical Processes with MUltiple Units
at Each Stage. AIChE Annual Meeting, Chicago, (Novernber, 1990)
47. Musier, RF.., and L. B. Evans Schedule Optimization with Simultaneous Lot-Sizing in Chemical Process Plants.
AIChE 1.,37, pp. 886-896 (1991)
48. Musier, R.F.H, and 1.B. Evans: Batch Process Managernent. Chern. Eng. Prog., 86 (6), 66-77 (I 990b)
49. Nof, S. Y., V. N. Rajan, and s.w. Frederick: Knowledge-based, Dynamic, Real-time Scheduling and Rescheduling:
A Review and Some Annotated References. Research Mernorandwn No. 89-16, School of Indus. Engineering,
Purdue University, West Lafayette, N (1990)
50. Ow, P.S., SF. Smith and A Thiriez: Reactive Plan Revision. Proceedings of AAAI Conference, 77-82 (1988)
51. Papadimitriou, C. H, and P.C. Kanellakis: Flowshop Scheduling with Limited Ternporary Storage. J. Assoc.
Comput. Mach., 27,533-549 (1980)
52. Pekny, J.F., D.L. Miller and G. 1. McRae: An Exact Parallel Algorit1un for Scheduling When Production Costs
Depend on Consecutive Systern States. Comput. & Chern. Engng., 14 (9), 1009-1023 (1990)
53. Prosser, P: A Reactive Scheduling Agent. Proceedings of AAAI Conference, 1004-1009 (1988)
54. Rajagopalan, D. and I.A. Karimi: Completion Times in Serial Mixed-Storage Multiproduct Processing with
 705

Transfer and Set-up Times. Comput. & Chern. Engng., 13 (Yz), 175-186 (1989)
55. Reklaitis, G. V.: Progress and Issues in Computer Aided Batch Process Design. Proceedings of the Third Int.
Conference on Foundations of Computer-Aided Process Design, CACHE-Elsevier, New York:, pp. 241-276
(1990)
56. Reklaitis, G.V.: Review of Scheduling of Process Operations. A1ChE Symposium Series, Vol. 78, No. 214,
119-133 (1982)
57. Rich, S.H., and G. J. Prokapakis: Multiple Routings and Reaction Paths in Project Scheduling. Ind. Eng. Chern.
Res., 26(9), 1940-1943 (1987)
58. Rich, S.H., and G. J. Prokapakis: Scheduling and Sequencing of Batch Operations in a Multipurpose Plant. Ind.
Eng. Chern. Process Des. Dev., 25(4), 979-988 (1986)
59. Sacerdoti, E.D.: The Nonlinear Nature of Plans. Advance Papers of the Fourth Jntemational Joint Conference on
Artificial Intelligence, Morgan Kaufinann, Los Altos, CA, 206-214 (1975)
60. Sahinidis, N. V., and I.E. Grossmann: MlNLP Model for Cyclic Multiproduct Scheduling on Continuous Parallel
Lines. Comput. & Chern. Engng, 5 (2), 85-103 (1991)
61. Sahinidis, N.V., and I. E. Grossmann: Reformulation of Multi period MILP Models for Planning and Scheduling
of Chemical Processes. Compul. & Chern. Engng., 15 (4), 255-272 (1991)
62. Salvador, M.S., Ph.D. Thesis: Case Western Reserve University, Cleveland, OH. (1978)
63. Shah, N., C.C. Pantelides and R.W.H.. Sargent: A General Algorithm for Short-term Scheduling of Batch
Operations - II. Computational Issues. Compul. & Chern. Engng., 17 (2), 224-244 (1993)
64. Rapacoulias, c., N. Shah, and C. C. Pantelides: Optimal Scheduling of Order-driven Batch Chemical Plants. In
L. Puigjaner and A. Espuna (eds), Computer Oriented Process Eng., Elsevier, Amsterdam, pp. 145-160 (1991)
65. Suharni, I., and R.HS. Mah: Scheduling of Multipurpose Batch Plants with Product Precedence Constraints.
Proc. of the Second FOCAPD Conference, A. W. Westerberg and H. H Chien, Eds; Amer Institute of Chemical
Engineers: New York, (1984)
66. Tongo, G.O. and G.V. Reklaitis: Completion Time Calculation of a General Multipurpose Batch Plant with
Resource Constraints. Paper 102a, A1ChE National Meeting, Orlando, (March, 1990)
67. Tsirukis, T., and G. V. Reklaitis: A Comprehensive Framework for the Scheduling of Resource Constrained
Multipurpose Batch Plants. Proc. of the Fourth International Symposium of Process Systems Engineering,
Montebello, Canada, (August, 1991)
68. Tsikuris, A., and G.V. Reklaitis: Feature Extraction Algorithms for Constrained and Global Optimization - I
Mathematical Foundations and II. Batch Process Scheduling Applications, Ann. Opns. Res.42, 275-312 (1993)
69. Wellons, M. C., and G.V. Reklaitis: Optimal Schedule Generation for a Single-Product Production Line - I.
Problem Formulation. Comput. & Chem. Engng. 13 (1/3),201-212 (1989)
70. Wellons, M.C., and G.V. Reklaitis. Optimal Schedule Generation for a Single-Product Production Line - II.
Identification of Dominant Unique Path Sequences. Comput. & Chern. Engng., 13 (Yz), 213-227 (1989)
71. Wellons, M. c., and G. V. Reklaitis: Scheduling of Multipurpose Batch Chemical Plants, I. Formation of Single
-Product Campaigns. Ind. Eng. Chem. Res. 30,671-688 (1991)
72. Wellons, M C., and G.V. Reklaitis: Scheduling of Multipurpose Batch Chemical Plants. 2. Multiple-Product
Campaign Formation and Production Planning. Ind. Eng. Chern. Res. 30,688-705 (1991)
73. Wellons, M. c.: Scheduling of Multipurpose Batch Chemical Plants, PhD thesis, Purdue U ni versi ty, West Lafayette,
IN, December, 1989.
74. White, c.H.: Productivity Analysis of a Large Multiproduct Batch Processing Facility. Comput. Chern. Engng,
13 (112),239-245 (1989)
75. Wiede, W., and G. V. Reklaitis: Determination of Completion Times for Serial Multi-product Processes - 2. A
Multiunit Finite Intermediate Storage System. Comput. Chern. Engng., II (4),345-356 (1987)
76. Wiede, w., and G. Y. Reklaitis: Determination of Completion times for Serial Multiproduct Processes - 3. Mixed
Intermediate Storage Systems. Comput. Chern. Engng., II (4), 357 -368 (1987)
77. Wiede, W., K. Kuriyan and G. Y. Reklaitis: Determination of Completion Times for Serial Multiproduct Processes -
I. A Two Unit Finite Intermediate Storage System. Comput. Chern. Engng., 11(4),337-344 (1987)
78. Wittrock, RJ.: Scheduling Algorithms for Flexible Flowlines. IBM J. Res. Develop., 29, 401-412 (1985)
79. Yeh, N.C., and G.V. Reklaitis: Synthesis and Sizing of Batch Sernicontinuous Process: Single Product Plants.
Comput. Chern. Engng., 11,639-654 (1987)
80. Zentner, M., and G. V. Reklaitis: An Interval Based Approach for Resource Constrained Batch Processing
Scheduling, Part I: Interval Processing Framework:. COPE-91, Barcelona, Spain, (Oct. 1991)
81. Zentner, M., and G. V. Reklaitis: An Interval Based Approach for Resource Constrained Batch Process Scheduling,
Part II: Assignment and Adaptive Storage Retrofitting. Paper 140d, AIChE Annual Mt., Los Angeles, (Nov. 1991)
82. Zentner, M., and G.Y. Reklaitis: An Interval-based Mathematical Model for the Scheduling of Resource
Constrained Batch Chemical Processes. This volume p. 779
GanttKit - An Interactive Scheduling Tool
L. Halasz, M. Hofineister, D.W.T. Rippin
TCL, Eidgenossische Technische Hochschule, CH-8092 Zurich, Switzerland

Abstract: A prototype interactive scheduling capability has been developed as part of the
BatchKit knowledge based system for planning and operation of batch chemical processes.
The main goal of the work was the verification of the potential of knowledge-integration
techniques in developing a flexible toolkit for a wide variety of tasks:
• Entity-relationship schema for the representation of pertinent complex objects and their
relationship
Active graphics support for man-machine cooperation
• Version management techniques to support case-based reasonmg and learning from
problem solving experience
• Generalization of branch-and-bound search algorithm
For each product one or more types of batches are defined, each of which represents the
allocation of the necessary processing tasks for that product to appropriate equipment items.
The schedule is constructed to satisfy the demand for a number of products for which due
dates within the planning period and corresponding demands are specified. Various constraints
can be imposed between successive tasks and successive batches of the same or different
products.
A simple branch-and-bound procedure is available for rapidly generating plausible
schedules and the merits of a schedule can be assessed by a performance criterion based on
how closely the cumulative demand curves for each of the products can be matched by the
scheduled production. Gantt charts of the schedules are displayed on the screen where they
can be interactively manipulated.

Keywords: Batch processing, scheduling, production planning

1. Introduction
Many problems in the design and operation of batch processes have been solved in recent
years. Increasingly sophisticated methods of mathematical programming are being applied to
the solution of larger and more general problems.
However, although special-purpose programs may be available to solve different problems
of batch processing such as design. medium term capacity planning or short term scheduling
considerable effort is still needed to transfer data from one to the other. Furthermore, however
 707

general the formulation, practical use of the programs almost always calls for some additional
features to be incorporated, which even if feasible can commonly only be realised with
substantial programming effort.
In addition there are still wide differences of practice in batch processing. Because of
different historical experience or the special circumstances of different batch environments
companies have different criteria, different practices and different calculation methods.
This situation favours the provision of a general batch processing environment in which a
variety of batch processing calculations can be carried out under the control of the user who
is kept informed and is free to intervene or interact with the programs.
Rather than a set of programs, we prefer to consider that the user has at his disposal a
toolkit which he can learn to use effectively and creatively. The BatchKit project has been
launched to investigate the potential of new technologies such as knowledge engineering and
artificial intelligence to provide the type of flexible environment or toolkit needed to aid
batch processing.
A related objective of the BatchKit project has been to gain experience in integrating
a relatively large set of problem formulations and solutions as opposed to dealing with
completely new problems. Thus some of the first capabilities to be realised in BatchKit
are similar to those previously available in earlier special purpose programs ([22], [16]' [14],
[6]) but in simplified form for demonstration purposes and with interactive facilities. The
individual modules (of which GanttKit is one example) will share conceptual knowledge of
the BatchKit system, as designed in the BatchKB knowledge base).
Research experience has shown that the main problem of such an integration, indepen-
dently of the domain, is the design of a common representation based on a unified data model.
For data modelling in BatchKit, first a relational data modelling system (AXI) has been
developed on top of IntelliCorp's KEE[lO]. AXI has proved to be useful for the formalisation
of the general structure which currently consists of two subsystems. The first subsystem has
been built around the SimKit discrete event simulation package in KEE. It made use of the
extensive graphical capabilities of this package and makes provision for the solution of some
problems of equipment allocation, batch sizing and capacity planning as described in[9] .
The second subsystem, GanttKit, a short term interactive scheduling system is described
here. The complete integration of the two systems is possible, however, they are separated
so that workstations which have KEE but not SimKit can still make use of GanttKit. The
GanttKit system was demonstrated during the NATO ASI workshop, June 1992, in Antalya,
Turkey.

1.1 Related Work

Current work in the field of production planning can benefit from the foundations laid
down in other fields :
Determination of the optimal sequence of operations (jobs) has been of interest in
operations research. Recent advances in the development of optimization algorithms
allow e.g. the formulation of some types of planning problems as MINLPs - see e.g.
[18] for review.
 708

Systems for production scheduling have been created for medium-term industrial needs (to
our knowledge, e.g. Sandoz, Ciba-Geigy, Rhone-Poulenc and ICI companies and certainly
many more are using such systems, see [17]), mostly based on simplified formulations
adapted to particular needs, many of them from the machine or aerospace industries.
Planning (i.e. finding an optimal sequence of operations (such as robot movements)
to reach a given goal) has been a research topic in artificial intelligence research. AI
research also established methods of temporal reasoning and the formulation and methods
for solution of the Constraint Satisfaction Problem [24].
In Computer Science, the establishment of constraint programming as a generalisation
of logic programming has opened up a new perspective in knowledge integration in
engineering domains[ll].

An example of a system for the study of batch processes is the BATCHES system [4]
- mainly used for simulation. A planning system ISIS based on constraints formulation has
been designed at Carnegie-Mellon University [8] and later extended to OPIS (Opportunistic
Intelligent Scheduler) [21].Other systems have been reported in [25] where e.g. the concept
of "plan critics" and local plan repair has been introduced to arrive at a final feasible and
acceptable, though suboptimal, plan.
An overview of the work in the area of scheduling and planning of batch process
operations can be found in [18].

1.2 Structure of the Paper

The general characteristics of a toolkit are described in Section 2, with some reference
to the BatchKit environment for the treatment of general problems in batch processing.
Attention is then directed to the GanttKit aid to scheduling which operates within the
BatchKit environment.
In Section 3 distinctive features of the problem are identified and an objective function
representing deviations from an ideal schedule is defined.
In Section 4 a partly heuristic solution to the scheduling problem is presented. This
generates favourable sequences of production which are then checked for feasibility.
Section 5 describes the structure of the knowledge base and the graphical representation
and in Sections 6 and 7 various operations on the graphical representations of the schedules
are presented.
Section 8 is about the DATAIMAGE system which is a powerful tool for interactive
data editing.
Section 9 provides an illustrative example showing the most important features of GanttKit
such as optimization, reconciliation of inconsistencies, manual modifications, sensitivity
analysis and version management.
Some conclusions are drawn in Section 10.
 709

2 Tool Kit
A tool kit is a collection of programs which operate in a common environment while
maintaining consistency throughout the operations. In contrast to a single program which
solves a well defined problem, a tool kit contains a set of problem solvers of varying capa-
bilities which may provide partial or complete solutions to the original problem formulation
or enhance the information available to the user in other ways.
GanttKit consists of the following main components:
1. KEE system implemented in the LISP language system and the graphical environment
of the X-Windows system.
2. AXI system - the relational tool.
3. BatchKB - The conceptual model of common concepts of batch processing in terms of
object classes and relations.
4. The DimageKB - The knowledge base of the general data editing subsystem.
5. GanttKB - The knowledge base providing the GanttKit-specific objects, relations and
functions.
6. The individual problem knowledge base storing the problem-specific data such as plant
structure, recipes, orders, products etc.
7. The problem solving algorithms:
a. General-purpose branch-and-bound and its parametrisation for the short-cut method
described below.
b. The schedule reconciliation function implementing the shifting algorithm described
below.

The most important issues of a toolkit design will be treated in the following sections.

2.1 Hierarchy of Problem Solving Levels

Typical complex problems such as production planning can be formulated and solved
(both by man and by the computer) at various levels of abstraction - ranging from detailed
calculations of single production steps to complete production plan optimization.
Maintaining problem state consistency while supporting the possibility of operating at
different levels of detail poses a difficult problem for which no general solution currently
exists. E.g., the low-level manipulation of the production plan - such as the shifting of
individual production steps - may violate assumptions made about the problem state by
higher-level problem solvers or by the user.

2.2 Problem Solution by Man-Machine Cooperation

The necessity of cooperation between man and the computer seems to be characteristic
of the very complex problems arising e.g. in planning. The first reason is the impossibility
of completely capturing in models the reality of a dynamically changing environment. The
second reason is the difficulty arising in solving problems formulated using more complete.
complex models. As a result, the modelling detail must be traded against the cost or time
 710

needed to formulate the problem and to find the "best" solution under constraints on available
resources (including time). Therefore, the domain expert who may be in possession of
information not represented in the model (or models) should be included in the solution
process. His knowledge may allow important decisions to be made when partial solutions
have been acquired, or complete solutions are to be compared.

2.3 The Role of Representation

The cooperative solution of a problem by combination of user and program operations at
varying levels of abstraction necessitates the maintenance of equivalence of representations
of the current problem state in all cooperating problem solving components (including the
human expert or agent).
This equivalence is most easily maintained if there is one common representation based
on a sufficiently general and powerful data model. In BatchKit, this task is fulfilled by the
AXI system - a relational extension of the KEE object-based representation which supports
the development of an Entity-Relationship schema [23]. The conversational language and
the retrieval system provided with AXI allows flexible manipulation of the problem state
representation using addition, deletion and modification operations. The overhead which the
KEE-AXI representation carries with it may call for transformation into a more efficient form
when the problem formulation is ready for solution.

2.4 Advanced User Interface

The cooperative problem solution requires intensive communication between man and
machine. The user interface design must provide representations which facilitate this com-
munication within the limitations of human perception and comprehension. The information
should be presented in a form which is easy to interpret for a human expert. Graphical rep-
resentation offers a suitable and powerful tool for manipulation by a human expert because
human comprehension is in many cases better with a graphical than with a textual repre-
sentation (a textual representation of a medium size graph is difficult to understand and to
manipulate when compared to its graphical equivalent).
The KEE software development environment uses a graphical interface for user interac-
tion. It also provides a sophisticated set of tools for the development of user interfaces.
E.g., the active graphics possibility provided by the KEE system allows to immediately
propagate changes to internal data towards the graphical interface and vice versa. This facility
can be used to maintain the equivalence between the graphical and the internal problem
representation by bidirectional propagation of changes also in the course of problem solution.

2.5 Version Maintenance

If the toolkit efficiently supports the maintenance of multiple problem versions, the user
can learn from problem solving experience and postpone the selection of final solution.
The KEE-AXI system uses three different representation techniques to maintain multiple
problem versions:
 711

Knowledge bases - suitable when the different problem versions are intended to be
independent of each other.
Object class instances allow simultaneous work with several versions and easy transition
between them.
KEE worlds: The reachability relations between versions (i.e., the possibility to reach a
version by incremental changes to another version) are expressed explicitly in the form of the
KEE world graph which can also be represented graphically using the KEE worlds browser.

3 Problem Analysis
As in other problem domains, the development of methods and algorithms for scheduling
of chemical batch plants pursues two conflicting objectives:
Generalisation - Successively more general problem formulations are being proposed,
resulting in increased efficiency problems.
Efficiency - The efficiency problems are countered by proposals of new, more efficient
search methods.
The efficiency problems are addressed by the following methods :
General efficient methods - which meet the original problem requirements but are more
efficient.
Relaxation - The requirements on the solution are relaxed in some sense (constraints or
optimality) and heuristic methods are proposed.
Specialisation - Within the general formulation, knowledge of special problem properties
(implicit in the class of formulations) is utilised to select or to construct a specialised
problem solver.
Problems of real life scheduling in both the chemical and machine industries are par-
ticularly difficult to represent completely and solve rigorously. A recent meeting of experts
(AMI Special Interest Group on Manufacturing (SIGMAN) [13]) made among others the
following statements :
Optimality is hard, if not impossible, to define in a realistic scheduling environment.
The computation of optimal schedules is a futile enterprise, given the unpredictability of
factory operations.
It is preferable to produce and maintain satisfactory schedules.
In view of these difficulties the scheduling feature of BatchKit does not attempt to provide
a rigorous optimization but explores a different approach. A partly heuristic optimization
procedure rapidly provides feasible schedules with some favourable characteristics. If desired,
this optimization procedure can be applied repeatedly to partially completed schedules, in
arbitrary alternation with manual adjustments to the Gantt chart displayed on the screen,
where an indication of the deviation from a hypothetical ideal schedule by cumulative product
balance curves is also always available.
The scheduling problems of multipurpose batch chemical plants have distinctive features
which call for a generalisation of earlier formalisations of the scheduling problem.
 712

An early formalisation - a flow-shop problem for 2 machines - is due to Johnson [12].

Gradually, as more powerful computers and algorithms became available, the problem was
extended and generalised. - [20] present a job-shop scheduling problem and propose the
representation in a directed disjunctive graph.
A definition of the job-shop problem according to Bellman [2], p48 is: When n jobs or n
items are processed by m machines, under a given ordering that is, given the machine order
to process each job, and given processing times of each job on each machine, the problem is
to determine the sequence of n jobs on each machine in order to minimize a given objective
function (performance measure).
In the present paper, this formulation is further generalised in that processes are not
characterised by a fixed, linear sequence of operations but by a set of precedence constraints
which may be represented by an acyclic digraph.
Further, the plant is assumed to be a set of equipment items with possibly restricted
interconnections. Therefore, the assignment problem also has to be solved before scheduling
can be performed. In GanttKit, the process-to-plant assignment is solved separately and a set
of selected solutions is given, represented by batch patterns.

3.1 Characterization of Batch Production

In this section, some characteristic differences between the chemical and machine industry
relevant to the scheduling problem are pointed out.
The assumption of treating the material as individual items can be made more frequently in
the machine industry than in the chemical industry where the materials treated are a continuum
and batches can be split and merged to any degree and thus lose their identity.
The nature of the physical and chemical processes taking place in chemical batch plants
allows some simplifications which are not generally applicable in the machine industry :
Chemical processes normally cannot be interrupted - no preemption and no inserted idle
time can be stipulated.
Each batch processing equipment item can process only one batch at a time -no overlap
is allowed.
A batch chemical product is produced in a series of process steps. The allocation of
each step to an equipment item (or machine) for a specified time defines a batch production
pattern. Different batch patterns may be defined by the assignment of the process steps to
different equipment items.
The definition of a batch pattern also includes the transfer times between equipment
items. These may significantly influence the timing of the schedule, particularly if large
volumes of material have to be transferred or there are long connections between successive
equipment items.
Constraints on the permitted delays between operations may be imposed when the
materials at the end of a step are unstable. The extreme case where no delays are permitted
between operations is designated as zero wait (ZW). The opposite extreme in which a batch
on which an operation has been completed may be allowed to remain in the equipment item
 713

indefinitely until the succeeding item becomes available is designated as no intermediate

storage (NIS). For intermediate cases minimum and maximum permitted delays between
operations may be specified.
The availability and role of intermediate storage will also influence the batch pattern
and may permit sections of a pattern to be decoupled and treated relatively or completely
independently of one another. The treatment of finite intermediate storage (FIS) is not
considered in the current prototype GanttKit. However unlimited intermediate storage (UIS)
can be specified.
Batches can be merged and split if some equipment items are used in parallel. Different
raw materials might be required and also different products of the same batch might be
produced at different points in time.
Jlroduct demand may be specified in different ways [19]. It is subsequently assumed
that orders are identified as triplets of product, quantity and due-date. A common feature
of multipurpose batch production is that the feeds for some products may themselves be
intermediate products produced in the same plant leading to a product precedence structure.
The planning horizon is a time span given by start and end times. The planning horizon
can be closed, or open at one end or at both ends. Some intervals in the planning horizon
can be disabled to represent holidays 01: special working conditions.
Constraints on the availability of resources, such as equipment (machines), utilities,
labour and raw materials are common. These are not currently modelled in GanttKit.
The changeover time needed between products on the same equipment item is likely to
be sequence dependent, i.e., determined by both the preceding and the succeeding products
(a vessel which has just produced black dyestuff needs a thorough cleaning before it can
be used to produce white dyestuff, if the production of the two dyestuffs is carried out in
reverse order, the cleaning time required may be much shorter) .. Changeover times but not
their associated costs are considered in GanttKit.
The list of constraints could be continued further which suggests that the constraints need
a general treatment which will be formalized in the following sections.

3.2 The Objective Function

Many alternative objective functions have been defined to measure the performance of a
schedule e.g. completion time, lateness, tardiness and machine idle time ([2], p64;[3], p8).
For the prototype scheduling system in GanttKit an objective function was devised which
measures how well different schedules are able to match a given set of product demands.
It is assumed that with an ideal schedule the correct amount of each product will be
produced at exactly the right time to satisfy each order of the demand. The total profit earned
by this ideal schedule is the sum over all products of the amount of each product supplied,
multiplied by its profit per unit. In practice the production pattern will not exactly match
the demand pattern. In the objective function, penalties are imposed for deviations from the
schedule. When products are produced too early they have to be kept in stock until required
by the order, thus incurring a storage cost. When products are produced too late the delivery
of the order has to be delayed, thus incurring a lateness charge. Both of these penalties are
 714

charged per unit amount and per unit time. The charges are not generally equal, the lateness
charge is likely to be higher.
A convenient way of recording deviations from the ideal schedule is by the cumulative
difference between production and demand over the planning period, also used in FMS/FAS
systems [1] and in other reported work on batch chemical production [26].
The production functions for all the products can be obtained from the set of currently
scheduled batches. It is assumed that all feeds are available before the start and all products
produced appear after the completion of the batch (see section 5.1). Thus if a terminal (initial)
step of a batch ending (starting) at time t produces (consumes) x mass units of product (feed)
i, a positive (negative) product increment of magnitude x becomes a set member of the
definition of Pi(i), the production function of product (feed) i at time t. (Increments which
occur simultaneously are of course added.)
The production function for product i is defined over the set of time points T; as Pi(tj),
Pi(t~), ... , Pi(t~.l, with tj+l > tj and to::; tj ::; i f Vj where to is the start time, t f is the end
time of the planning horizon. Pi(tj) is the amount of product i which becomes available at
time t~ and is of positive sign, any amount of the same product consumed as feed for any
batch appears in the production function with a negative sign. A surplus or shortage may
also be recorded at the starting time to.
The demand function of product i is similarly defined as amount increments over the set
of time points T; as qi(tD, qi(t~), ... , qi(t~.J. where tj+l > t~ and to ::; tj ::; t f Vj.
The cumulative balance of product i :

The cumulative production and demand can be analogously defined at all points including
the end point of the planning horizon.

Pi(t) = ~Pi(tj)
1'<1
}-

Qi(t) = ~ qj(tj)
1'<1
}-

Differentiated penalties ct, ci are imposed for positive or negative deviations from the
ideal production schedule reflected in positive or negative values of the cumulative balance.
The differential penalty at time t is a function of the cumulative balance at that time, Gj (Rj( t))
where the (asymmetric) function Gi( T) ofreal value T (quantity of product missing or in excess
of demand in our interpretation) may be defined e.g. as :
 715

The cumulative penalty until time t is then

J
t

Li(t) = Ci(R;(t'J))· dt'J

o

The symbols ct,ci are the penalties per unit amount of product i per unit time for
surplus and deficiency respectlvely.
The maximum attainable profit for product i is II;nax = Ii x Qi(i e ), where f; is the profit
per unit mass of product i supplied. Since it may not be possible to produce the whole of
the demand, the actual gross profit is

IIi = Ii x (min (Qi(ie),Pi(i e )))

The net profit for product i is

IIi = IIi - Li(te)

and the total net profit is

4. The Scheduling Problem

In this section, the formulation of the scheduling problem and the principles of its solution
are presented in general terms in the introductory subsection. The principles of the solution
as they are implemented in the GanttKit system are presented in the following subsections.

4.1 Problem Formulation

For the scheduling problem the following are given:

1. Partial plan comprising batches already allocated to equipment items, their timing and
other resource utilization.
2. Batch production patterns given by precedence graphs including constraints on timing of
successive steps.
3. Demands given by product - quantity - due-date triplets.
4. Change-over constraints between different batches.
5. Holiday constraints.
6. Equipment item break-down constraints.

The problem is to complete the given partial plan so as to fulfil all unfulfilled orders and
maximize the total net profit.
 716

4.2 Principles of Solution

In the first approximation, the solution of the scheduling optimization problem can be
seen as a two stage, generate-and-test procedure: the first stage generates feasible solutions
and the second stage tests them for improvement with respect to the current best solution. In
logic programming notation (which is very expressive for the formulation of chronological
backtracking, see [15]) :
(best-plan ?x) (pOSSible-plan ?x) ;; Generate
~
(improved-plan ?x) ;; Test
(asserta (best-sa-far ?x)) ;; Store
(fail). ;; Repeat loop
(best-plan ?x)~(best-so-far ?x) ;;Collect result
(improved-plan ?x) <- (not (best-sa-far ?y) ). ;; The first
(improved-plan ?x) <-- (best-sa-far ?y)
(better-plan ?x ?y)
(retract (best-sa-far ?y) .
In the above program, the predicate possible-plan generates and binds to the variable
?x the possible solutions, i.e., feasible plans. (An attentive reader will have noticed the
simplification made above: The set of feasible plans will be infinite e.g. in presence of
continuous variables e.g. when the time is modelled by a continuous variable. In such a
case, the assumption has to be made that ?x is bound to a suitably selected representation of
subsets of feasible plans - e.g., all solutions with a given sequencing of steps but without
the actual absolute position on the time axis.)
The identification of the optimal solution by the above program necessitates that the
union of the plans generated by the predicate possible-plan and bound to the variable
?x covers the complete set of solutions, i.e., all feasible plans. Due to the infeasibility of
the full enumeration of all feasible schedules, incompleteness is introduced in two ways into
the above procedure :
1. Only a subset of all possible feasible schedules is generated
2. An allowance is made in the optimality test in favour of efficiency.
This approach would work well if the set of solutions were representative, i.e., if we were
able to produce a sufficient number of essentially different solutions (not just in the vicinity
of a single local optimum).
The above approach is rather inefficient for scheduling problems of realistic size for two
reasons: firstly, the solution space is very large, and secondly strong bounding functions are
difficult to design.

4.3 Principles of Solution in GanttKit

The pmblem formulation in GanttKit is restricted in the following ways:
Batch splitting and merging can be introduced by explicitly defining batch patterns
(recipes) for the same product but with the allocation of more than one equipment item
to a process task. The relative timing of the steps of the batch pattern which process the
parts of a split batch must reflect the nature of production (e.g. in-phase or out-of-phase
parallel production at one stage).
 717

General resource constraints are not considered.

In GanttKit, the solution of the short-term scheduling problem has been decomposed into
two subproblems:
1. Optimization of a relaxed problem (short-cut optimization): Determine the number and
types of the batches required to satisfy the given orders and the order in which the
individual batches will be allocated to the plan.
2. Allocate sequentially the batches generated by the short-cut optimization to the current
state of the plan. This problem is solved using the batch allocation procedure which is
capable of allocating a single batch to a partial plan.
An advantage of this decomposition is that the batch allocation procedure can be used
also for other purposes. E.g. the user can add a batch "manually" to an existing partial plan
using the batch allocation procedure which determines a feasible timing for the batch; or
an infeasible plan can easily be reconciled, allowing fast generation of feasible plans under
different constraints; a development of this idea is the sensitivity calculation module which
is also based on the batch allocation procedure, where the change of the objective function
value is recorded under changing constraints.
In the"short-cut" optimization procedure, the maximum total load on an equipment item
is used as the objective function, which can be strongly bounded and is independent of most
of the constraints of the original problem.
The suggested procedure can be given in logic programming notation as follows:
(good-plan ?x) - (best-plan ?y) (feasible-plan ?y ?x) .
The first subgoal of the procedure generates an optimal solution ?y of the relaxed problem.
The second subgoal generates a feasible solution for the original problem ?x from ?y.
It is assumed that an optimal solution of the reduced problem results in a "good" solution
for the original problem.
On the other hand, the solution to the relaxed problem may result in a batch sequencing
for which no feasible schedule can be found by the second stage, although a feasible solution
exists. This may be the case in strongly constrained problems where batch sequencing
becomes critical.

4.4 "Short-cut" Optimization

The "short-cut" optimization is a branch and bound procedure which determines a series
of batches (with their number, type and desired latest starting time) to be added to the current
schedule which are required to fulfil the remaining outstanding orders.
It uses two heuristics :

1. Batches producing the most urgently required products are added first.
2. It seems to be appropriate to assume that only those schedule candidates can be easily
timed which evenly distribute the load between different equipment items, i.e., if the
maximal equipment item load is minimized.
 718

4.4.1 Product urgency The cumulative difference curves for each products yield the in-
formation needed to determine which product is needed most urgently. By considering the
missing quantities of products and the corresponding times, the latest time at which the pro-
duction must start to deliver the required quantity can be calculated for each type of batch
which produces the product. The product for which this time is the earliest is considered to
be the most urgent (or, the corresponding orders are the hardest to fulfill).

4.4.2 Equipment occupancy The maximum of the workload over the equipment items is
independent of the timing of batches. It can be calculated from the specification of the batch
patterns to be implemented. Thus a lower bound can be calculated for each partial solution
by estimating the smallest total equipment item load (for each equipment item) resulting from
fulfilling the remaining orders.
For each equipment item additional loads for unfulfilled orders are added to loads for
batches already assigned to give a lower bound for the workload on that item. The overall
bound is the largest of these workloads over all equipment items.
It should be noted that in estimating the additional workload needed to satisfy the
outstanding orders, additional loads are placed only on equipment items whose use is
unavoidable for at least one product. If a product can be manufactured by several batch
patterns the workload will be increased only for equipment items required in all the batch
patterns. Nothing is added to the load of any other item.
For minimization, the lower bound for a partial solution is an estimate of the objective
function value which is lower than the objective function value of any complete solution
derived from this partial solution. If the lower bound is greater than the objective function
value of an already established complete solution then the corresponding partial solution (and
thus all complete solutions which could be derived from it) will be eliminated. This may lead
to a substantial reduction of the search space.
The procedure for calculating the lower bound for any partial solution is as follows.
Let E = {el' e2, ... , en,} be the set of equipment items and let Ai be the current load of
equipment item ei. Let P = {PI, P2 • ... , PnJ be the set of products, let D = {d I, d2 • ••• , dnJ
be the set of total unfulfilled demands, so that di is the total unfulfilled demand of product
Pi. Let B = {b l , b2 , •.• , bnb } be the set of batch patterns.
Let L = L( ei, dj, bk ) be a function defining the load on equipment item eifE if the
demanded amount djED of product pjEP is fulfilled with batch pattern bkfB.
If some bk does not produce product pj then L( ei, dj, bk ) is taken as infinity. If h
produces product pj then L(ei,dj,bk) is taken to be zero for all equipment items which are
not used by bk otherwise it can be calculated from the mass balance and the equipment item
occupancy of the batch pattern.
The minimum load on ei to satisfy the unfulfilled demand dj of product Pi can be given by
li(dj ) = rrun
B
L(ei, dj, bkl; i = 1,2, ... , ne;j = 1,2, ... ne
The minimum load on ei to satisfy the unfulfilled demand of all products can be given by
lr in = L l;(dj); i = 1,2, ... , ne
D
 719

The lower bound value is

max(,\'
E I
+ [min)
I

4.4.3 The Algorithm The short-cut optimizer works as follows:

1) For all products cumulative difference curves are calculated resulting from the present
existing (consistent) schedule and the partial solution, i.e., the set of batches already added.
2) The remaining orders for which the quantity in stock plus the production predicted
for the current partial solution is insufficient to satisfy all orders are ranked by their urgency
value described in the preceding subsection. The most urgent product is selected.
3) All batch patterns for the most urgent product are considered for adding to the
production plan to produce a new short-cut partial solution.
4) Partial solutions are evaluated in terms of a bounding function estimating the minimal
necessary total load on equipment (described in the preceding subsection). Partial solutions for
which the bound on the load is higher than the value for a complete solution already identified
are pruned. i.e., eliminated from further consideration. The remaining partial solutions are
ranked according to their bound value and the batch of the partial solution with lowest bound
is added to the partial schedule. The partial schedule is not checked for consistency. The
choice between partial solutions with the' same bounding value is arbitrary.
5) The first four steps are repeated until all orders are fulfilled. i.e., the cumulative
difference curves for all products are non-negative at the end of the planning horizon.
At this point a solution is obtained for the sequence of batches to be manufactured to
meet the total product demand. This solution is then compared with the stored best solution.
Notes:
1. The partial solution contains only the latest required start time for each required batch
without establishing its feasibility.
2. During the subsequent adding of batches to the schedule the actual start time of each
batch may be changed by the shifting due to the constraints.
3. The short-cut algorithm therefore uses an approximation or prediction of the cumulative
curves.
The logic programming formulation is given by defining the predicate possible-plan
in the preceding section "Principles of Solution" as folows :

(possible-plan
(?sched
)
(or (and
;; Solve by finding completion of an empty partial schedule
(schedule-completion NIL ?sched)
)
 720

(schedule-completion
;; Find the completion of partial schedule
(?partial-schedule
?complete-schedule
)
(or
(and (cumulative-curve ?partial-schedule
?cumulative-curves)
(find-hardest-order ?cumulative-curves
?hardest-order)
(batch-alternatives ?hardest-order
?alternative-batches)
;; Select the batch alternative with the
;; lowest max. equipment occupancy
(select-batch-alternative ?partial-schedule
?alternative-batches
?batch)
(better-than-bound ?partial-sched ?batch)
;; Schedule is better than current bound
(add-batch-to-schedule ?partial-schedule
?batch
?new-schedule)
(extend-if-incomplete ?new-schedule
?complete-schedule)

)
)

(extend-if-incomplete
( ?new-schedule
?complete-schedule
)
(or (and (complete-schedule ?new-schedule)
;; No further production required
(is ?complete-schedule ?new-schedule)
)
;; Unsatisfied orders remain
(and
(schedule-completion ?new-schedule
?complete-schedule)
 721

Notes:
1. The enumeration of alternative hardest orders is by the predicate find-hardest-
order which successively selects the alternative orders in the order of increasing latest
possible start time. This predicate currently does NOT backtrack, i.e., the sequence of
products is given by the recursively applied product urgency criterion.
2. The enumeration of batch alternatives is by the predicate select-batch-
alternative which successively selects the alternative batches in the order of in-
creasing maximum equipment occupancy.
3. The predicate schedule-complet ion is recursively called in the second clause of
the predicate extend-if-incomplete.
Optimality in terms of equipment load is guaranteed if the complete search is performed
but the program allows the computational effort to be reduced if necessary by, for example,
reducing the number of alternatives considered at each level or generating a solution which,
while not necessarily optimal, is guaranteed to be within a specified interval from the optimum.
The "short-cut" optimizer returns a list of order - batch pattern pairs, so that the orders
are sorted according to their urgency, from left to right, e.g.,

((ORDER-3 A-BATCH) (ORDER-4 B-BATCH)

(ORDER-3 A-BATCH) (ORDER-2 B-BATCH))
In the partial schedule for which the optimization has been carried out, ORDER-2,
ORDER-3 and ORDER-4 were unfulfilled. Note that ORDER-3 appears twice because one
batch of A-BATCH could not fulfil it completely and with its remaining unfulfilled quantity
it became the 3rd most urgent demand.

4.5 Batch Allocation

The batch allocation procedure· processes the list of order/batch pattern pairs returned by
the short-cut optimizer and progressively extends the partial plan by adding each batch to the
list. After adding each batch, anew, feasible plan results.
The timing of the batch is subject to two types of. constraints :
Plan constraints (global and inter-batch).
Precedence constraints (intra-batch).
The plan constraints are those by which the current state of the plan constrains the
allocation of steps of a batch. They include the partial occupancy of available resources
(equipment, labour, utilities etc.) by already allocated batches and general constraints such as
no overlapping, no preemption, availability of resources, holidays, limits on utilities, labour
etc.
Precedence constraints are imposed upon the sequence of processing steps by which each
batch is processed and are internal to the batch. They may reflect e.g. transfer time and
 722

material stability between successive steps. The precedence relation can be represented as
a connected precedence graph from which the predecessors and successors of each step can
be determined. Steps without predecessors are called start steps, steps without successors
are called terminal steps.
A precedence constraint determines the earliest and latest starting times of a step relative
to the end of a predecessor step. If a step has more than one predecessor then its earliest and
latest relative start times are constrained by the conjunction of all its precedence constraints.
If the interval between the earliest and latest start time for a step is empty then the batch
cannot be allocated.

4.5.1 Feasible Allocation The batch to be allocated is defined as a graph G(S; P) where
5 = {SI,S2, ... ,Sn} is the set of steps to be allocated, P is a subset of 5 x 5 defining
precedences for some pairs of S. If (Sj, Sj )oP then Sj is a successor step of 5j or conversely
Sj is predecessor of 5j. Note: G(S,P) can be represented by a directed acyclic graph
(precedence graph).
Let T = {tl' t2, ... , tn} so that tj is the start time of Sj, U = {u], U2, ... , un} so that Uj
is the duration of 5j.
A precedence constraint vt is a pair (Vii, V1;). Vii and V1; constrain the earliest and
the latest starting time of Sj, respectively, in terms of its predecessor Sj, as follows:

1. Vii: tj ~ tj + Uj + dii
2. V+,·
1)'
t J· <
-
tj + Ui + d+'IJ

where d-:-IJ and d+' are predefined values ' r IJ <- d+IJ .
vt is fulfilled. A predicate
IJ
The relative timing of Si and Sj is feasible if the constraint pair

can be given which is true if and only if the allocation of Sj is feasible in terms of one of its
predecessors Sj. A step S j is feasible in terms of all its precedence constraints iff

4.5.2 The Partial Plan The partial plan is defined as a set of constraints CPo To determine
how the partial plan constrains the allocation of 5 J , a predicate function F P can be given

which is true if and only if the allocation of Sj (i.e., including the starting time) is feasible
in terms of Cp .
It is assumed that the cP is constant, i.e., during the allocation of a batch it does not
change. This assumption allows the separation of the treatment of precedence and plan
constraints.
 723

4.5.3 The Problem The problem is to find a feasible set of timings T' = {t~, t~, ... , t~}
so that

4.5.4 The Solution The algoritlun used to allocate a batch in an existing schedule in a
feasible manner is similar to tho~ presented e.g. in [6] or [5]. The algoritlun searches for the
feasible timing of all steps of a batch to give a completion time which is latest before or earliest
after the given starting position which was determined e.g. by the short-cut optimization. The
feasible timing is obtained by repeated shifting of steps - always in the same direction - by
the minimal amount to obtain a new starting time which makes a step feasible with respect
to all plan constraints and one of its precedence constraints. Because this new location might
violate other precedence constraints of the step considered, backtracking is made until no
constraint is violated. It is, however, possible that no feasible timing exists between the start
and end time of the planning horizon.
If a production unit has to fulfil a given order by some due date then an initial timing
for all of its constituent batches can be given which satisfies the order just-in-time. In the
case that this just-in-time allocation is not feasible, a backward shifting is tried (it is better
to produce too early rather than too late). If the start time of the planning horizon prevents
finding a feasible solution using backward shifting then forward shifting can be tried starting
from the just-in-time allocation.

4.5.5 Implementation of plan constraints Let Si and Sj be two steps already allocated to
the same equipment item so that Si precedes Sj, i.e., tj ~ ti and there is no step between Si
and Sj. The constraint Oij€O can be given as follows:

tj ~ ti + Ui + 0ij
where 0ij is the changeover time required after Si when it is followed by Sj. Normally
0ij "I- 0ji.
The problem is to allocate an additional step Sk between two already allocated consecutive
steps - Si followed by Sj on the same equipment item - so that the changeover constraints
will be satisfied. The feasible location, tk for Sk can be calculated as follows:

tj ~ tk + Uk + 0kj
The above formulation implicitly contains two other assumptions:
Each step of a batch can be allocated to only one equipment item.
After finishing a step the equipment item becomes immediately free for further use subject
to the corresponding changeover time.
 724

The work pattern or holiday constraints H C cP restrict the use of equipment in certain
periods of time. A holiday constraint Hj{H can be expressed as an inequality inhibiting the
overlapping of all steps with a predefined time interval.
The planning horizon represents a hard constraint. If, during shifting, any step of the
batch to be allocated has reached the end of the planning horizon, i.e., if tj + Uj > if then
the corresponding allocation is infeasible.

5. Representation of Scheduling

The main objective was to find a representation which enables a realistic and flexible
description of short-term scheduling problems and which allows flexible manipulation of
schedules both by programs and by the user through the (graphical) interface.
To fulfil this loosely formulated main objective the system has to be capable of

Representing batch production patterns as building blocks of schedules.

Maintaining any number of (partial) solutions of a scheduling problem.
Allowing both graphical and programmed manipulation of the schedule, while maintaining
the equivalence between the two forms of representation. The graphical and programmed
manipulations include creation, modification and deletion of objects.
Representing the constraints as global objects which are to be applied to all scheduling
variants.

With these objectives an entity-relationship schema has been developed (Figure 1). Note
that the actual model in GanttKit contains many attributes and functions and also some object
classes which are not denoted here, however the structure is essentially the same.
The meaning of symbols in Fig.! is as follows :

Rectangles represent object classes.

Rounded rectangles represent relations (among members of object classes).
Ovals represent attributes of the members of object classes.
 725

Figure 1: ER-Graph

The E-R diagram of Figure 1 can be transformed into a set ofAXI assertions as follows:

(ME-BATCHES ME-STEP-R ME-STEPS)

(ME-STEPS ME-STEP-FROM-R PRECEDENCE-CONSTRAINTS)
(ME-STEPS ME-STEP-TO-R PRECEDENCE-CONSTRAINTS)
(ME-STEPS HAS-ATTRIBUTE EQ-ITEM-ASSIGNED)
(ME-STEPS HAS-ATTRIBUTE DURATION)
(ME-STEPS HAS-ATTRIBUTE PRODUCTION)
 726

(PRECEDENCE-CONSTRAINTS HAS-ATTRIBUTE MIN-TIME)

(PRECEDENCE-CONSTRAINTS HAS-ATTRIBUTE MAX-TIME)
(SCHEDULES BATCH-R BATCHES)
(SCHEDULES PSCHEDULE-R PSCHEDULES)
(SCHEDULES HAS-ATTRIBUTE START-TIME)
(SCHEDULES HAS-ATTRIBUTE END-TIME)
(SCHEDULES HAS-ATTRIBUTE OFV)
(BATCHES STEP-R STEPS)
(BATCHES ME-BATCH-R ME-BATCHES)
(BATCHES PBATCH-R PBATCHES)
(BATCHES HAS-ATTRIBUTE CURRENT-START)
(STEPS STMBST-R ME-STEPS)
(STEPS PSTEP-R PSTEPS)
(STEPS HAS-ATTRIBUTE CURRENT-START)
(STEPS HAS-ATTRIBUTE DURATION)
(PSCHEDULES IS-SUBCLASS-OF VIEWPORTS )
(PBATCHES IS-SUBCLASS-OF INVISIBLE.PICTURES)
(PSTEPS IS-SUBCLASS-OF RECTANGLES)

5.1 Schedules
Schedules are defined as a hierarchical structure of the object classes SCHEDULES,
BATCHES and STEPS. The START-TIME and END-TIME attributes of the SCHEDULES
class define the planning horizon. The attribute OFV stands for the objective function value.
BATCHES and STEPS are related to the corresponding production pattern object classes:
MB-BATCHES and MB-STEPS. The relations MB-BATCH-R and STMBST-R have N:l
functionality showing that a batch pattern may have any number of instances.
TIre CURRENT-START of BATCHES means absolute time in the planning horizon,
whereas the CURRENT-START attribute of STEPS is relative to that of BATCHES. The
DURATION attribute of STEPS stands for the occupancy of the corresponding equipment
item. Its value is equal to that of the DURATION of the related MB-STEPS object when
the step is created.
Both the user and the scheduling algorithm can independently and simultaneously manipu-
late any number of SCHEDULES - versions of a scheduling problem. This allows backtracking
to earlier solutions when the version currently pursued proves to be infeasible.

5.2 Batch Production Patterns

The notion of batch pattern plays a central role in the GanttKit system - a schedule
is a set of allocated batches, each of which is derived from one of the batch patterns by
assigning a starting time to each step.
Precedence graphs have been found to be adequate for the representation of batch
patterns. As an example consider Figure 2. The nodes correspond to the operations carried
out in order to transform the batches. Each node has a unique identifier (StepO ... Step8),
 727

an equipment item (e.g. Vessell) and (except the terminal nodes) a duration associated with
it. Terminal nodes describe the sources of raw materials (e.g., StepO) and final destination of
products (e.g., Step7 and Step8). The equipment item associated with a terminal node can
be final or intermediate storage. The components required and produced by a batch are also
described by terminal nodes. The batch pattern in Figure 2. e.g. requires 500 mass units of
A and 500 mass units of B at StepO and produces 100 mass units of Band 900 mass units
of C at Step7 and Step8, respectively.

Step4 stepS

Figure 2: Precedence Graph

The arcs of the batch pattern precedence graph are labeled by the earliest and latest start
times of the.step at the end of the arc relative to the end time of the step at the beginning of
the arc. E.g. the pair [0,0] expresses a zero wait policy between two successive operations.
The precedence graph G( 5, P) is represented by the object classes ME-BATCHES,
ME-STEPS and PRECEDENCE-CONSTRAINTS. The relation ME-STEP-R with function-
ality l:N between ME-BATCHES and ME-STEPS, defines that a batch consists of any
number of steps. A step is always carried out in one equipment item, described by the
EQ-ITEM-ASSIGNED attribute. The DURATION attribute defines the time required to carry
 728

out the step on the equipment item. The PRODUCTION attribute which is used only for
tenninal steps describes the required or produced materials in the fonn of a list of pairs, e.g.
( (A-PROD -500.0) (B-PROD -500.0) )
The first element of each pair is the name of the material, the second one is the quantity
which is negative for feeds and positive for products.
The precedence constraints between batch steps are described by the PRECEDENCE-
CONSTRAINTS object class. MIN-TIME constrains the minimal, MAX-TIME the maximal
time lag between the finish time of a step and the start time of one of its successors.

5.3 Graphical Objects

The upper part of the E-R model in Figure 1 defines the objects supporting the graphical
representation of the schedule in the fonn of a Gantt chart implemented using the KEEPictures
graphical system incorporated in KEE.
The object classes VIEWPORTS, INVISIBLE. PICTURES and RECTANGLES are de-
fined in the KEEPictures system knowledge base. The GanttKit picture object classes
PSCHEDULES, PBATCHES and PSTEPS are subclasses of the KEEPicture object classes;
they inherit the attributes and functionalities of their superclasses.
Only two levels in the picture hierarchy are really visible, namely the objects of the
PSCHEDULES and PSTEPS object classes, whereas the class PBATCHES is defined as a
subclass of INVISIBLE. PICTURES; its role is to define the structure of graphical objects
without being visible.
The SUPERPICTURE relation and its inverse the SUBPICTURES relation are defined
by the KEEPictures system. The structure of the picture object classes corresponds to that
of the schedule.
In accordance with the equivalence between the graphical and internal representation, the
relations between the internal and graphical objects have a 1:1 functionality, i.e., exactly one
graphical object corresponds to each schedule object. The internal representation objects can
be used independently of the graphical ones : a schedule may exist without the graphical
representation and the user may decide when s/he wants to display it graphically.
An important attribute of the grapical objects is POSITION which defines the horizontal
and vertical location of the object picture relative to its superpicture in tenns of pixels.

6. Interactive Operations

The GanttKit Panel is the pictorial representation of a schedule in the form of a Gantt
chart. A panel is implemented as a KEEPictures viewport, or more precisely as an object of
the PSCHEDULES object class which is defined as follows:
(PSCHEDULES IS-SUBCLASS-OF VIEWPORTS)
Similarly to any other viewport, the viewed picture of a GanttKit panel is the topmost pictorial
object of all pictures inside the panel.
 729

..
II -1000

~ -2000
i
t -UOG
'1
-4000

I ._--
'---
._--

.
lfil.
J~ I
I
:: 0 I
DRYER-1 ""

DRYER':
II
I I I I \I I I I I I III I I I II III
o 50 100 150 200 250
15310.08 Time

Figure 3: GanttKit panel with three manually allocated batches

A GanttKit schedule panel (such as the one in Fig.3) consists of the following components:

The panel itself with a frame.

The time axis (x-axis) representing the planning horizon.
Equipment item axis (y-axis) labeled by the equipment items.
Objective function value field appears in the left bottom comer of the panel. In constrast
to other picture objects of the GanttKit system, it is not a KEEPicture object but an active
image.
Cumulative curve plots are displayed in the upper part of the panels. Depending on the
value of the APLOT-TYPE attribute of a schedule which is either demand, or production,
or difference, the cumulative demand, or production, or difference of one or more products
are shown. The cumulative difference curve is the default.
Schedule appears as a set of boxes in GanttKit panels. Section 7 contains a detailed
discussion of schedules.

6.2 Add batches

The Add batches operation is the implementation of the batch allocation algorithm defined
earlier. It allows one or more batches to be added to an existing schedule. Each batch added
 730

is an instance of one of the predefined batch patterns of the current set-up. If more than one
batch is to be added, the operation must be repeated. For each batch to be added a reference
time should be specified (the reference time may be e.g. the due date of an order). The
batches already allocated will not be changed during the batch allocation.

6.3 Reconcile
Manual operations or changing of the constraints may make existing schedules infeasible.
In GanttKit, the exact optimum is not sought but the feasiblity of the schedules is a basic
requirement. Therefore, a mechanism is needed which makes infeasible schedules feasible.
This mechanism in Ganttkit is the Reconcile operation.
There are many ways to reconcile a schedule, e.g., trivially by deleting all of its allocated
batches. GanttKit's Reconcile leaves the number and type of allocated batches unchanged.
Their timing will be modified as follows: first the schedule to be reconciled is emptied and
the batches are temporarily stored (this results in a feasible schedule). Then the batches are
taken one by one, according to increasing original start time and reallocated using the batch
timing algorithm described earlier.
The batch timing algorithm uses the original step timings as initial start times, therefore
the Reconcile is efficient. A feasible schedule is left unchanged by reconciliation.

6.4 Cumulative Curve

The Cumulative Curve menu item allows the cumulative difference curve, which was
defined earlier, to be displayed.

6.5 Fulfil Rest

The Fulfil Rest is the principal operation of the GanttKit tool. It fulfils the currently
unsatisfied product demand using the B&B scheduling algorithm described earlier.

6.6 Panel
The Panel menu item has only one subitem: "Delete". It allows deletion of the panel
(the pictorial representation of the schedule) without deleting the schedule itself.
Any schedule currently not displayed can be displayed again any time later using the
Schedule - Display menu item of the global GanttKit menu bar.

6.7 Schedule
The Schedule menu item has two subitems : Copy and Delete.
Copy makes a duplicate of a schedule under a new name. Delete can delete an empty
schedule, i.e., one which does not contain any batches. A non-empty schedule should first
be emptied (see next section) before it can be deleted.
 731

6.8 Unfulfilled Orders

The cumulative difference curve makes it possible to identify the unfulfilled orders which
can be displayed using this menu item.

7. Operations on Schedules

A schedule appearing in a GanttKit panel is a complex structure consisting of a hierarchy

of SCHEDULES, BATCHES and STEPS objects and all' related objects, including those
implementing the graphical representation.
This structure enables us to navigate among objects. The navigation allows related objects
to be reached starting from a given object along paths composed of relations and inverse
relations. On the user interface only the PSTEPS objects, i.e., the graphical representations
of the STEPS are directly accessible with the mouse.
Because the operations on many different objects are very similar to each other the
following subsections are presented in terms of operations, always noting how the particular
operation applies to different objects.

7.1 Highlight
Highlighting is the way to emphasize selected parts of the structure of a schedule by
graphical means, e.g. by changing the fill pattern of a picture. Any object at each level of
the object hierarchy can be highlighted to reveal the object structure of the schedule.

7.2 Move
The Move operation allows parts of the schedule (STEPS, BATCHES) to be shifted along
the time axis. When moving an aggregated object, all its subobjects are moved simultaneously.
Inherently, many constraints apply to moves along the time axis. The immediate
propagation of their effects is not used for several reasons. Firstly, it may be sometimes
necessary to allow inconsistent states of the schedule which can be reconciled later. Secondly,
the propagation of too many time constraints would slow down the move operation. Thirdly,
many of these constraint violations may be obvious to the user.
A move operation includes the following steps:

1. Identifying the object to move.

2. Collecting all component objects to be moved together.
3. Moving the selected objects with the mouse.
4. Calculating the new start time for each moved object.
5. Recalculating the objective function and updating the product plots.
6. Shading the panel which marks potential inconsistency.
 732

7.3 Move+
The Move+ operation allows any user defined set of steps (of the same or different
batches) to be moved simultaneously. To accomplish this, the steps to be moved together
should first be designated with a Mark operation which is available for batches and steps.
Steps or sets of steps which have been marked by mistake can be "unmarked" with the
Demark operation. The colour of demarked steps will be set back to the colour before the
marking.

7.4 Delete
A complete schedule or any batch of a schedule can be deleted. Because batches are the
smallest independent parts of a schedule, objects lower in the hierarchy, such as steps, cannot
be deleted. Deleting a batch means the physical deletion of the object, whereas deleting a
schedule deletes only its lower level objects and leaves an empty schedule.
After a deletion the objective function value is recalculated and the contributions of the
products are displayed in the KEE Typescript window.

7.5 Verify
The Verify operation, which is available for steps and schedules, carries out the following
tests:
1. Equipment items can carry out only one step at a time (no overlapping).
2. For each step of the schedule :
a. Changeover constraints.
b. Transfer time constraints.
c. Material stability constraints.
d .. Holiday constraints.
e. Equipment break down constraints.

8. Interactive Problem F~diting

Scheduling problems normally require a large amount of data which may be prepared
in an input file or, in the best case, are gathered from a company-wide database. A third
possibility, implemented in the GanttKit system, is to use an interactive data editor. For this
purpose, the interactive data editing DATAIMAGE tool was developed which can be attached
to any KEf; knowledge base.
A DATAIMAGE control unit which can be attached to any KEE object class allows the
user to define which member units of the class and which of their attributes should be displayed
and how they should be treated. More than one DATAIMAGE control unit can be attached
to any object class which allows the same object class to be visualised in different ways.
Figure 13 shows a DATAIMAGE panel which is attached to the ORDERS object class.
Here, all member units are displayed, sorted by product name and due-date.
 733

A DATAIMAGE panel can be used to add or to delete member units, or to change attribute
values. The possibility of changing a unit's attribute is indicated by a heavily lined box. In
Figure 13 e.g. columns 1,3, and 4 are mouse sensitive and can be changed but 2 and 5 are
not. The first column contains the object (unit) names.

8.1 Sensitivity
It is often of interest to know how the optimal (or suboptimal) solution. of a system
changes with variations of the parameters in the formulation. Even if the global optimum is
not found it is meaningful to ask how a suboptimal solution changes if a variable is changed
and a new suboptimal solution is found by the given problem solver, which in our case is
the Reconcile operation.
The DATAIMAGE system allows sensitivity analysis to be carried out using as the
independent variable any numerical data appearing in a mouse sensitive DATAIMAGE box.
A bound value is to be defined for the independent variable and GanttKit reoptimizes the
problem for 10 equidistant values of the independent variable between the current value and
the bound value. A sensitivity plot is made which shows the independent variable on the
x-axis and the change of the objective function value relative to the current value on the y-axis.
The use of the Reconcile operation for sensitivity analysis has important advantages
but also certain limitations. It is advantageous in that the results of the sensitivity analysis
correspond to those obtained by the use of the Reconcile operation. Moreover, the solution
technique is efficient. However, because the Reconcile operation modifies only infeasible
schedules, the problem formulation at the bound value of the examined variable should be
more constrained than at the current value. Otherwise the indicated change of the objective
function will be zero. Because both the GanttKit optimizer and the sensitivity analysis
compute suboptimal solutions, the objective function value obtained may be greater for a
more constrained problem than for a less constrained one. Further details about the sensitivity
analysis are given in the example.

8.2 Versions
GanttKit allows many problem versions to be maintained simultaneously. This feature
allows the "what - if" type of questions which are often of interest in ill-defined and dynamic.
problems such as scheduling.
The problem versions are represented by KEE's KEEworlds system which allows
different versions of a single problem to be built by incremental changes starting from a
root variant (background) and their interdependencies to be subsequently maintained as nodes
of an acyclic, directed graph.

9. Illustrative Example
This section illustrates some of the important features of the GanttKit system but is not
intended to be a· comprehensive numerical example.
 734

A short time scheduling problem should be solved with a planning horizon of two weeks,
i.e., 336 hours, starting on Monday at 0 o'clock and finishing next week on Sunday at 24
o'clock.
The plant represented in a SimKit design window (Figure 4) consists of 3 vessels, 2 filters
and 2 dryers and facilities to supply raw material and store final products. The plant works
in three shifts, round the clock. Two plant shut-downs are defined, the first one from 45 to
50 hours and the second one from 170 to 180 hours. It is assumed that VESSEL-I is not
available between 50-90 and 120-150 and VESSEL-2 between 85-115.

.. ...........
J ... II" I .. , lim;II'lli

>
~~

t!I
""'"
---
~
Efr [JJ
~"'- UEs~T ~

~ -[JJ~/l!!,W' ' ' S':2' 'J!:'.~.,.

DRVERS
111111 t!I

111111 (I-SOURCE:
"'LT~RS

[JJ
I.Ii!S5fLS

VESSEl-S

Figure 4: Plant in SimKit design window

Two products - RED-PROD and B,LUE-PROD - can be produced according to Table 1,

i.e., there is one production variant (batch pattern) for RED-PROD and two for BLUE-PROD.
Table 2 together with Table 3 and 4 define three batch patterns RED-BATCH, BLUE-BATCH-
A and BLUE-BATCH-B. For the sake of simplicity the terminal nodes are omitted and it is
assumed that the products appear always at the end of the last batch step.

Table 1: Batch patterns and production quantities

Product Batch Pattern Batch size [kg]

RED-PROD RED-BATCH 300.0
BLUE-PROD BLUE-BATCH-A 600.0
BLUE-BATCH-B 600.0
 735

Table 2: Batch patterns and step durations

Batch Pattern Step Equipment Dur Successors

RED-BATCH RED-STEP-! VESSEL-! 9.0 RED-STEP-2
RED-STEP-2 FILTER-! 4.0 RED-STEP-3
RED-STEP-3 DRYER-2 !5.0
BLUE-BATCH-A BLUE-STEP-AI VESSEL-2 !1.0 BLUE-STEP-A2
BLUE-STEP-A2 VESSEL-3 7.0 BLUE-STEP-A3
BLUE-STEP-A3 FILTER-2 4.0 BLUE-STEP-A4
BLUE-STEP-A4 DRYER-! 24.0
BLUE-BATCH-B BLUE-STEP-B 1 VESSEL-2 11.0 BLUE-STEP-B2
BLUE-STEP-B2 VESSEL-3 7.0 BLUE-STEP-B3
BLUE-STEP-B3 FILTER-2 4.0 BLUE-STEP-B4
BLUE-STEP-B4 DRYER-2 24.0

Table 3: Precedence constraints

Balch Pattern FROM TO MIN MAX

RED-BATCH RED-STEP-I RED-STEP-2 2.0 7.0
RED-STEP-2 RED-STEP-3 1.0 3.0
BLUE-BATCH-A BLUE-STEP-AI BLUE-STEP-A2 2.0 5.0
BLUE-STEP-A2 BLUE-STEP-A3 0.0 2.0
BLUE-STEP-A3 BLUE-STEP-A4 2.0 3.0
BLUE-BATCH-B BLUE-STEP-B 1 BLUE-STEP-B2 2.0 5.0
BLUE-STEP-B2 BLUE-STEP-B3 0.0 2.0
BLUE-STEP-B3 BLUE-STEP-B4 2.0 10

Changeover times are defined in Table 4 for those batch steps which are assigned to the
same equipment item.
 736

Table 4: Changeover constaints

Equipment From step To step Time

VESSEL-! RED-STEP-J RED-STEP-! 4.0
VESSEL-2 BLUE-STEP-Al BLUE-STEP-A! 1.0
BLUE-STEP-Bl BLUE-STEP-Bl 4.0
BLUE-STEP-Al BLUE-STEP-Bl 4.0
BLUE-STEP-Bl BLUE-STEP-Al 4.0
VESSEL-3 BLUE-STEP-A2 BLUE-STEP-A2 4.0
BLUE-STEP-B2 BLUE-STEP-B2 4.0
BLUE-STEP-A2 BLUE-STEP-B2 4.0
BLUE-STEP-B2 BLUE-STEP-A2 4.0
FIL1ER-l RED-STEP-2 RED-STEP-2 4.0
FIL1ER-2 BLUE-STEP-A3 BLUE-STEP-A3 4.0
BLUE-STEP-B3 BLUE-STEP-B3 4.0
BLUE-STEP-A3 BLUE-STEP-B3 4.0
BLUE-STEP-B3 BLUE-STEP-A3 4.0
DRYER-! BLUE-STEP-A4 BLUE-STEP-A4 5.0
DRYER-2 RED-S1EP-3 RED-STEP-3 2.0
BLUE-STEP-B4 BLUE-STEP-B4 5.0
RED-STEP-3 BLUE-STEP-B4 3.0
BLUE-STEP-B4 RED-STEP-3 3.0

Table 5: Profits and penalties

Product Profit [SFrIkg] Penalty+ Penalty-

[SFr/lOOkg*b] [SFrIlOOkg*b]

RED-PROD 25.0 0.285 2.85

BLUE-PROD 30.0 0.342 3.42

The profit and penalty for over- and underproduction are defined in Table 5, the initial
stock for each product is assumed to be 0.0. The penalty for underproduction is 10 times
higher than that for overproduction.
The orders for products RED-PROD, BLUE-PROD with due dates and quantities are
given in Table 6.
 737

Table 6: Demand in Kg.

Hour RED-PROD BLUE-PROD

20 600.0
60 900.0 1600.0
110 1200.0
220 700.0
310 1800.0

To solve the problem, first a new, empty panel is opened (Figure 5), consisting of a Gantt
chart in the lower part and a curve plot in the upper part.

II I I I III III I I III I I I III

50 100 150 200 250 300
- 40775.5'2 TillU!

Figure 5: Empty GanttKit panel

The time axis of the Gantt chart represents the planning horizon of 0 to 336 hours, the
Y-axis the equipment items. The two vertical shaded bars represent the two plant shut-down
constraints, the horizontal bars represent the three equipment item break-downs. The field in
the bottom left corner of the panel shows the objective function value recalculated whenever
the schedule is modified. The contribution of each demanded product to the objective function
value is typed out in the typeout window.
When an empty panel is created as in Figure 5, e.g. the following message appears:
 738

Current and maximal profit for schedule Dl

BLUE-PRO -25978.32 138000.00
RED-PROD -14797.2055000.00

-40775.52. 193000.00
The maximal profit for BLUE-PROD of SFr 138000.00 results from the total demand of
BLUE-PROD of 4600 kg and the profit per unit amount of BLUE-PROD of 30.00 SFr/kg.
The 2200 kg total demand of RED-PROD with 25.00 SFr/kg unit profit yields SFr 55000.00,
i.e., the total maximal profit is SFr 193000.00. The initial profit for both products is negative
due to the penalty because no orders can as yet be fulfilled.
The upper part of the panel shows the cumulative difference curve for both products.
Because no batches have as yet been allocated, the curve for both products is equal to the
cumulative demand given in Table 6.
Figure 3 is an example of the batch allocation procedure and its implementation by the
Add batches menu item. One batch was allocated for each batch pattern RED-BATCH,
BLUE-BATCH-A and BLUE-BATCH-B. The three dashed boxes represent allocated RED-
BATCH steps, the light grey boxes are BLUE-BATCH-A steps, the dark boxes are BLUE-
BATCH-B steps. At the end of the two blue batches 600 kg BLUE-PROD is produced, so
the CD curve is updated accordingly.

Figure 6: GanttKit panel after manual operation

 739

Figure 6 is an illustration of a manual operation which has been carried out on the panel
in Figure 3. The last step of one of the blue batches has been shifted forward to a time where
it violates some of the constraints, because it overlaps with a holiday and also violates the
precedence constraints. The panel is shaded indicating possible inconsistencies .

..
R -2000
t
i
t
'1

•
Ii
VESSEL..l
I

•••
TER-l
I
FILTER': I
DRYER-l
DRYER':

I I I I I I I I I I I I I II III I I I
o 50' 100 150 200 250 $00
13873.68 TillU!

Figure 7: GanttKit panel after restoration of consistency using Reconcile

The Reconcile operation can be used to remove the inconsistencies. Figure 7 shows
the new state after reconciling the schedule of Figure 6. The step which was previously
manually moved has been shifted backward to the latest feasible time and the steps belonging
to the same batch have also been adjusted in order to satisfy the precedence constraints. The
resulting panel is feasible, the cumulative difference curve has been updated, the objective
function recalculated and the background set to white.
 740

Q
".. -500
t" -1000
i
t
'1
-1500
GJ In----'~I~.~
...
[
-2000

..
III :~II== ::1111
~: I ~ :::1. I
:~ I I ::: I I I
I I: II ... I I I I

••
DI I I I
• III:':
~.II .. 111111111
I I I I I I I I I I I Iii i i i iii I Iii i i i i i i I i
o 50 100 150 200 150 ~oo
151455.68 Ti ....

Figure 8: GanttKit panel with a schedule created by Fulfil rest. Op.

Figure 8 is an illustration of the Fulfil rest operation which is used to complete a partial
plan. It has been obtained from an empty panel after previously deleting the three batches of
Figure 7. It takes about 20 seconds on a SUN Sparc 1+ workstation to obtain this schedule.
The resulting profit distribution is as follows:
Current and maximal profit for schedule 01
BLUE-PROD 129739.33 138000.00
RED-PROD 50392.80 55000.00

180132.13 193000.00
The 8 batches of BLUE-PROD each of size 600 kg produce 200 kg more than the
demanded 4600 kg, the 8 RED-PROD batches each of size 300 kg yield 200 kg of product
in excess of the 2200 kg total demand. The BLUE-PROD CD curve is in the negative region
between about 60 - 260 hours which means that orders with due dates falling in this interval
cannot be fulfilled on time. Also, the last order of BLUE-PROD cannot be completed within
the planning period which is indicated by the negative CD curve between hours 300 to 310,
showing a deficit of 1000.0 at the end of the planning period.
 741

HOLIDAYS .

UNIT-NAME FROM TO

IHOLIDAY-l
II 45
II 50

IHOLIDAY-2 II 170
II 180

EqlllJ,m.el\t BreaJ:-d.OWIu

UNIT-NAME EQITEM FROM TO

IEQBREAK-l II
VESSEL·l
II 50
II 90

IEQBREAK-2 II VESSEL·l
II
120
II 150

IEQBREAK-3 II VESSEL·2
II 85
II 115

Figure 9: Holiday and equipment breakdown constraints

The equipment item break-down DATAIMAGE panel is shown on Figure 9. All fields
enclosed in thick lines can be accessed using the mouse and the corresponding values can be
changed. Also existing units (which is always a row in a DATAIMAGE panel) can be deleted
or new units can be added. All these changes are propagated to the schedules.
A new equipment item break-down is added to FILTER-l from 50 to 100 hours, the
modified DATAIMAGE panel is shown on Figure 10.

UNIT-NAME EQITEM FROM TO

IEQBREAK-267 100

90

150

EQBREAK-3 115

Figure 10: A DATAl MAGE panel after addition ofa fourth equipment breakdown constraint

This change is also propagated to the schedule panel D1, however the newly added
break-down constraint does not violate the feasibility of the existing schedule, so it is left
unchanged (Figure 11.).
 742

Q
"
..
t
i
-100
[
t -1500
'1
-200

• .
II I :===1 1 - ':1 I I I
:::.
VESSEL-l
VESSEL': !!lI: I ~ I I
VESSEL-3
:~ I I ... I I I
I I I I I I I I

••
G I I ... I I I
:::
IIJIIl
l1li111: ;\~ttj;~
l1li11 ... 111111111
...
I I I I II I I I I I I I II I I I II I I I II I I I
o 50 100 150 200 250 300
151438.58 TillU!

Figure 11: New schedule after addition of a fourth equipment breakdown

Sensitivity analysis can be carried out on any numerical value in the active fields of
DATAIMAGE panels. Figure 12 shows the change of the objective function value when the
upper bound of the last added equipment item break-down is varied between 100 and 125
hours with a step size of 2.5 hOUTS. This change, as can be seen in Figure 11 ,affects only
the second step of the third RED-BATCH directly but indirectly the entire third RED-BATCH
is affected. Because of other constraints the batch cannot be moved backward, therefore it is
shifted forward. A small drop of the objective function value can be observed from 100 to
102.5 hours-'because small forward shifts of the third RED-BATCH can balance the increasing
length of the equipment break-down. However, at 102.5 hours these small shifts cannot help
anymore and the batch must be shifted to the end of the planning horizon which results in
a sharp drop of the objective function value.
 743

......... \
\,
\~
\
D ~j

\
e

'.<--------\
1

\
r
o
f
i

\\
t

IF
r \-
)

- '00 O-. 105 110 115

l-,j-r..,.......,.j"'II.
..,.."T""'I"""Ij.......-,jr-r..,...."Tj-,lr-Tj..,...j"Tj""'j'I""'j"T""'I"""Ij......
100 120 125
Quantity

Figure 12: Sensitivity curve

Version management makes it possible to try out different variants of the same basic
problem. In our example we have created a world WI as a descendant of the background
which means that the problem definition in WI is the same as the one defined so far. In WI,
a new problem formulation is developed with changes in the orders for BLUE-PROD. First
the demand of BLUE-PROD is cilanged as shown in Figure 13, i.e., it is assumed that 1000
kg out of ORDER-BI is required no earlier than 320 hours.
 744

ORDERS ill wottd IIIWottd: WI]

UNIT-NAME PRODUCT DUE QUANTIIT ctIMUL

ORDm-Bl
II BLUE-PROD 600.0

ORDm-B2
II BLUE-PROD 1800.0

ORDm-B3
II BLUE-PROD 3600.0

ORDm-363 BLUE-PROD 4600.0

II
ORDm-Al
II RED-PROD 600.0

ORDm-A2
II RED-PROD 1500.0

ORDm-A3 RED-PROD 2200.0

Figure 13: Orders in WI

The facts added to world WI after this modification are shown in Figure 14. Additionally,
the first holiday constraint is deleted and the first equipment item break down constraint is
shortened to the interval 80-90.

III (Output) facts in Wl World

limitivt F~cts:
A WORLD-UNITS Of ORDERS IS ORDER-363)
A QUANTITY Of ORDER-81IS 600.0)
A CUMULATIVE-QUANTITY Of ORDER-81IS 600.0)
A CUMULATIVE-QUANTlTV Of ORDER-82IS 1800.0)
A CUMULATIVE-QUANTITV Of ORDER-83IS 3600.0)
A CUMULATIVE-QUANTITV OF ORDER-363IS 4600.0)

(du/:(d Facts:
O~

Figure 14: Additional facts in World WI

After these modifications an empty panel is opened in world WI which reflects this new
problem formulation (Figure 15).
The objective function distribution of this panel is
Current and maximal profit for schedule 01
BLUE-PROD -17086.32 138000.00
RED-PROD -14797.20 55000.00

-31883.52 193000.00
 745

.." .....,... ··3

---
---

I I I I I I I I II I I II I I I I I I II
o 50 100 150 UO 250 300
-31883.5'2 Ti .....

Figure 15: Empty panel in WI

Figure 16 shows the result of the Fulfil rest operation carried out on the empty panel. The
CD curve of BLUE-PROD has become positive during most of the planning period, i.e., the
first three BLUE-PROD orders can be fulfilled on time (note that a surplus is built up towards
the end of the planning period to do this). However, the last order for the BLUE-PROD (of
1000.0 at 320 hours) still cannot be satisfied before the planning horizon and a deficit of
1000.0 remains at the end of planning period. The RED-PROD orders cannot be fulfilled on
time but it is easy to see that the schedule cannot be improved for RED-PROD under the given
constraints. For the two first RED-PROD orders whose early due dates (20 and 60 hours) are
too restrictive the production starts as early as possible and for their 1500 kg total amount
exactly 5 RED-BATCH batches are allocated each with 300 kg batch size. For the third.
RED-PROD order with due date 220 hours and 700 kg amount, 3 RED-BATCH batches are
allocated, i.e., there will be 200 kg overproduction at the end of the planning period. These
3 RED-BATCH batches are allocated from the due date 220 shifting them backward in time.
 746

2000
Q
u 1000
a
n
t
i -1000
t
'1/ -2000

III <1111
I • ... I I I

.... ....
I I ... I I I
I I I I I I I I
I I I I I I

••••
...

1111 .11
...
1111111111111111111111111111111111
o 50 100 150 200 250 $00
158459.15 Ti.....

Figure 16: Schedule in WI

The profit distribution is as follows:

Current and maximal profit for schedule Dl

BLUE-PROD 137658.68 138000.00
RED-PROD 53461.29 55000.00

191119.97 193000.00

10. Conclusions
As mentioned in the introduction, GanttKit was developed as a pilot project within
a longer-term project - BatchKit. The completion of this work allows the following
conclusions to be drawn :
Batch processing : Only a limited number of problem formulations were solved. This
may be explained by the fact that the work was dominated by the overall project's
goals - more attention has been paid to general aspects and experimentation with new
methodology and less effort was devoted to comprehensive coverage of even a limited
subfield of batch processing (such as scheduling).
General problem solving techniques :
a. The relational representation has confirmed its value as a basic technique.
 747

b. Generalised branch-and-bound algorithm has been investigated but an algorithm of

only limited generality was finally used, mainly for efficiency reasons, however the
principle used allows an easy adaptation to other problem formulations.
c. Version management is currently based on a mixture of techniques. A synthesis and a
more uniform approach is desirable. A necessary pre-requisite of such a generalisation
would be a sufficiently general problem representation framework, so that the versions
can be defined in terms of sufficiently homogeneous increments.
d. The graphical interactive interface of the scheduling system is its most impressive
component, especially considering the small proportion of the total effort actually
spent in its development. This has confirmed the value of KEE as a good development
toolkit.
e. Generalisation of problem solution for short-term scheduling poses a difficult problem
- the program specialisations and transformations which are almost always needed
even in simple cases to achieve efficiency in problem solving are as yet by no means
sufficiently understood.

Knowledge integration : The GanttKit as a pilot project did not attain generality to a
degree sufficient for an integration of a larger body of knowledge from even the limited
field of short-term scheduling, although it offers a meaningful selection of problem-solving
techniques. A more expressive representation is needed to describe further aspects such
as

a. Quantified constraints in terms of complex nested object structures

b. Enumeration strategies in the sequential explication of the search space and in the
application of alternative problem solving methods to subproblems.

Only significant extensions to the representation can help to overcome the limitations of
the current techniques.

The above results appear to be modest considering the ambitious goals of the BatchKit
project [HHR89]. The modest effort invested so far did not lead to achieving the "competence
level" threshold postulated by Feigenbaum [7]- confirming the general recognition that
a large body of application domain knowledge is needed to treat even apparently trivial
problems. The effort which could lead to such a system will probably be larger by orders
of magnitude.
On the other hand, the development has confirmed the importance of knowledge integra-
tion as a goal and also the benefits gained by using the latest software development technology.
The benefits can however only be achieved if the requirements of know-how and interdiscipli-
nary educati~n (e.g. in operations research, computer science, chemical engineering) imposed
by these techniques are met.

Further Work The current basic relational representation technique is being extended to a
fully deductive system. This work has been completed since the presentation of this paper.
The new representation is based on the integration of object and logic programming in the
LinX system which replaces the AXI system described here.
 748

Introduction of plant and process concept (also already completed) will allow other types
of problems to be addressed, such as e.g. the automatic equipment allocation and batch
pattern generation, design, mass balance calculations or multiplant design.
The introduction of constraints will allow the inclusion of further standard constraints
(e.g., more general types of resource constraints).
Further generalisation of branch-and-bound together with the introduction of constraints
will allow easier transition between MLNLP formulations and the default branch and bound
problem solving, while retaining the advantage of user-friendly problem formulation.
This effort will also be supported by the extension of the relational system towards a fully
deductive knowledge base system integrating object-oriented and logic programming.

Concluding remarks The GanttKit system is the first fully interactive system with a
graphical interface for the short-term scheduling of batch chemical plants based on the
presented type of technology. The technology utilised holds a promise that the ease of
operation experienced with GanttKit in short-term scheduling can be extended to other areas.
Due to the variability of problem formulations and to the combinatorial complexity of
the reSUlting optimization problems. the methodology needed to develop useful systems for
the treatment of problems in the batch processing area in general will continue to present
new challenges.
For this research to become practically relevant for industry and for the benefits of
rationalisation to be achieved it is necessary for the information and production automation
systems in industry to reach a critical lowest integration level.
From the point of view of an academic research group, the challenge can probably be met
at the problem solution level by a combination of strategies and at the toolkit development
level by continuously adjusting the knowledge acquired so far to new emerging technologies.
 749

References

1. AkeUa, IL Operations Research Models in F7exibleManufacturing Systems (Archietti F., Ed.). ch. Real
Time Part Dispatch in Flexible Assembly, Test and Manufacturing Systems, Spring, 1989.
2. Bellman, R., Esogbue, A., and Nabeshima, I. Mathematical Aspects ofScheduling and Applications.
Pergamon Press, 1982.
3. Brucker, P. Scheduling. Akad. Verlagsgem. Wiesbaden,1981.
4. Clark, S., and Kuriyan, K. Batches - simulation software for managing semicontinuous and batch
processes. InProc. AIChENat. Mg., Houston (1989).
5. Cott, B., and Macchietto, S. Minimizing the effects of batch process variability using on-line schedule
modification. Computers and Chem. Eng. J3, 1,105-113, (1989).
6. Egli, M., and Rippin, D. Short-term scheduling for multiproduct batch chemical plants. Compo and
Chem. Eng. 10, 4,303-325, (1986).
7. Feigenbaum, E. A., and Lenat, A. On the thresholds of knowledge. InIJCAJ'87, vol. 10, (1987).
8. Fox, M., and Smith, S.F. ISIS - A knowledge-based system for factory scheduling. Expert Systems, 25,
49, (1984).
9. Halasz, L., Hofineister, M., and Rippin, D. A flexible knowledge based toolkit in batch processing. In
AIChE Annual Meeting (1989).
10. IntelliCorp, XKEE User's Guide. IntelliCorp Inc., 1988.
11. Jaffar, J., and Lassez, J.L. Constraint logic programming. In Proc. Con! on Principles of
Programming Languages, 1987.
12. Johnson, S. Optimal two- and three stage production with setup times included. Naval Res.Logistic
Quarterly 1, 61-68, (1954).
13. Kempf, K, LePape, C., Smith, S., and Fox, B. Issues in the design of ai-based schedulers:A workshop
report. AIMagazine JJ, Jan (1991).
14. Klossner, J. Computer-gestuetzter Entwurf von absatzweise arbeitenden chemischen
Mehrproduktanlagen. PhD thesis, ETH Zuerich, 1985.
15. Kowalski, IL Logic for Problem SolVing, North-Holland, 1979.
16. Mauderli, A., and Rippin, D. Production planning and scheduling for multi-purpose batch chemical
plants. Compt. and Chem. Eng. 3, 199-206, (1979).
17. Musier, R., and Evans, L. Batch process management. Chem. Eng. Progress, 66-72 , June (1990).
18. Reklaitis, GV. Overview of scheduling and planning of batch process operations. This volume, p. 660
19. Rippin, D.w. Design and operation of multiproduct and multipurpose batch chemical plants - an analysis
of problem structure. Computers and Chem. Eng. 7,463-481, (1983).
20. Roy, B., and Sussmann, B. Les problemes avec constraintes disjonctifs. Tech. Rep. 9, SEMA Montrouge,
1964.
21. Smith, S.F., Ow, P.S., Potvin, J.Y, and Nicola, M. An integrated framework for generating and revising
factory schedules. J. Operations Res. Soc. 41, 6,539-552, (1990).
22. Sparrow, R. Multibatch: A computer package for the design of multi-product batch plants. The
Chemical Engineer 289,520-525, Sept (1974).
23. Ullman, J.D. Database and Knowledge-Base Systems. Computer Science Press, 1988.
24. Van Hentenryck, P. Constraint Satisfaction in Logic Programming. MIT Press, 1989.
25. Wilkins, D.E. Practical Planning: Extending the Classical Al Planning Paradigm. Morgan Kauhmann
Publ., San Mateo CA, 1988.
26. Yamasaki, Y., Morikawa, H., and Nishitani, H. Production scheduling system for a batch process and
its application to retrofit design. InProc. PSE88, Sydney (1988).
An Integrated System for Batch Processing

S. Macchietto, C.A. Crooks and K. Kuriyan

Centre for Process Systems Engineering

Imperial College, London SW7 2BY, UK

Abstract: The engineering activities accompanying a batch plant development project are
reviewed and the requirements for an integrated batch system environment are detailed. The role
of some of the currently available supporting tools is discussed. Finally, a design is presented of
a system aimed at integrating batch plant and sequence control system design, off-line operations
scheduling and on-line operations management. Preliminary results are presented with particular
reference to a multipurpose batch pilot plant example.

Keywords: Batch plant, design, scheduling, production management, sequence control, modeling,
integration.

1. Introduction

The need for better integration of process engineering functions in the batch processing field has
been discussed several times recently [32,39,43,45,47,48]. Over the last year, within the Batch
Processing Project in the Centre for Process Systems Engineering we carried out a collaborative
study with several industrial partners aimed at identifYing; in some detail, the major technological
and research opportunities, needs and obstacles associated with an integrated system for batch
processing. Three main industrial partners participated in the study, representing a wide range of
industries, backgrounds and present and future demands. APV Baker is a leading supplier of
equipment and processes to the food industry. This industry is characterized by rapid evolution
in market conditions, high product innovation, fierce competition and often tight production
margins. It is not unusual for some food plants to produce 100-200 product varieties with
exceedingly short demand and delivery horizons. The closeness to the end-consumer market
makes food processing representative of the highly dynamic and responsive production
requirements expected in other industries. John Brown is a engineering company with strong
presence in the more mainstream chemicals, pharmaceuticals and biotechnology industries and
 751

with specific experience in batch plants. Both these companies are contractors, rather than
operating companies. As such, they contributed the distilled experience and requirements of many
projects and end-users. BNFL builds and operates plants in the nuclear sector, where many
important operations (fuel preparation, decommissioning, etc.) are discontinuous. Thus, they
reflect both the plant designer and plant operator viewpoints. For obvious reasons, this industry
has particular concerns regarding lifecycle engineering, safety, formal development procedures
and validation. These are already very important in other segments of batch processing (e.g.
pharmaceuticals) and are expected to become ever more central to all batch processing projects,
as regulatory and environmental demands increase and the responsibility is imposed on plant
suppliers and operators to demonstrate compliance. In addition to these major partners,
discussions were held with a large number of other industrial collaborators, in particular fine
chemicals producers.
As part of the project, we conducted a detailed analysis ofthe engineering activities involved
in typical batch processing projects. The goal was to identify the main requirements for a set of
tools and an integration environment to support all such activities. The analysis involved in-depth
discussions with engineering, design and automation teams and included retracing the evolution
of several representative projects from initial preliminary design to commissioning and operation.
To gain first hand experience, we also built our own fully automated, multipurpose batch pilot
plant facility. The design was commissioned to APV Baker and Centre's personnel was seconded
for several months to them to work on the project as integral part of the development team.
In the following, first we will briefly summarize some conclusions of the industrial survey,
and then detail some of the background integration "infrastructure" issues, in particular with
respect to approaches to integration, special modeling requirements posed by batch processing,
data base management systems and modeling environn1ents. The second part of the paper deals
more specifically with some models, techniques and software for preliminary batch
design/scheduling and batch management/control. Some developments are then presented in the
area of automatic synthesis of batch management and control procedures, which have the effect
of bringing together the design and operation aspects. Finally, preliminary results are presented,
with particular reference to our pilot plant.

2. Batch Process Engineering Activities

In spite of the large differences in type of processes and industries covered by the various
industrial partners, the approaches to design and operate batch plants were remarkably similar.
 752

The main engineering activities, schematically shown in Figure 1, consist of:

Preliminary ProcesslPlant Design the goal of which is to identify the major equipment,
simple flowsheets and layouts, descriptions of all processes and major costs. For batch plants, the
demand for equipment is clearly tied to its utilization, hence operational decisions (e.g. operation
in campaign mode, use of distinct lines for different products, operating schedules, etc.) must also
be taken at this stage. Design and operational decision which may have large effect on the
achievable production rates and costs will have to be considered. This includes type, number and
size of process equipment required, their connection network, utility systems and manpower
levels. Broad operations specifications are drawn up and the main control system requirements
are established at this stage. Contracts are typically won and budgets fixed on the basis of this
work.
Detailed Design. Detailed line diagrams (including control hardware) are produced starting
from the sketches generated at the preliminary design stage, together with utility line diagrams,
a detailed plant layout and a more detailed specification of the operating procedures. Mechanical
design (vessels, pipes etc.) is typically done using CAD tools.

I PROCESS I
DEVELOPMENT I
DETAILED
DESIGN

-~ LAYOUT
I

I
-H
PROCESS I
PlANNING I VESSEL
DESIGN

I
I

I
PREUMINARY I unUTIES
DESIGN I
---H SYSTEM
DESIGN

II PIPING
DESIGN
AUTOMAnON II I

II RECIPE
DEVELOPMENT
~ - OPERATION

I, SEQUENCE
DEVELOPMENT
~
- -~
ORDER
ENTRY
I

I SEQUENCE
VAUDATION
I
- ~ PRODUCTION I
PLANNING

I~~~~~NI
,-- 1 SCHEDULING I

COMM~~NING I :- 1 BATCH_I

I
MANAGEMENT

I OPERATOR I I EQUIPMENT
TRAINING I
I ~~~;;~

Figure 1. Lifecycles Activities for a Multipurpose Batch Plant

 753

Automation. Starting from the functional specification and broad operating procedures, the
control software is further specified through various stages of detail, down to individual sequence
control code, and implemented from the bottom up.
Operation. Once built, the usual range of operational activities for a batch plant include
production planning and scheduling, production management, batch execution, etc.
The activities described cover a very wide range of expertise and technologies, typically
involving several organisations over a plant lifecycle. Time horizons may range from two wceks
for preparation of a preliminary design bid to several years between initial design, commissioning
and a major plant retrofit [16,17]. Increasingly, decommissioning aspects will also have to be
considered as part of a plant lifecycle.
Some of the major conclusions of the study are summarized here:
I) The preliminary design decisions by and large determine the eventual plant throughput
range and perfomlance. Techniques which permit exploring in some detail the interaction between
design and operation decisions are required.
ii) Detailed design activities are essentially the same as for continuous plants and are
reasonably well covered by existing CAD systems.
iii) There is considerable overlap between activities of different departments (process,
automation, drafting, etc.) making decisions which are highly interactive. There is much interest
in having a consistent set of information easily accessible to all functions at all stages of
development. This should include the description of "operations" (functional and procedural
aspects) as well as those of a plant flowsheet, from the early stages of a design. These descriptions
should evolve in parallel with equipment design.
iv) The development of batch control and automation systems is an engineering intensive
activity. A major problem area lies in the interface between process engineering and
control/automation groups. As the process and plant are refined many iterations are often required
to clarify an operating sequence or implement it correctly. Apparently, minor changes in
equipment or process detail may have major implications on the automation systems and vice
versa, with costly reconciliation iterations. There is considerable interest in techniques for
improving automation development productivity, reducing errors and validating control
sequences. Interfaces to general purpose simulations are desirable for "what if' studies and
operator training.
v) It is becoming more common for contractors to deliver detailed 3D "design models" to
plant operating companies, but the latter typically have to develop their own operations models
for the supervisory control functions (scheduling and batch management). There is a large interest
 754

in formal but generic "production models" which can be used at the design stage to prove the
overall plant performance and then transferred directly to operations.
vi) The need to frequently revise product formulations (recipes and control sequences),
supervisory production control strategies and plant configurations requires that all techniques
used for grassroots plant and operation design should be model driven and also applicable to the
retrofit case. Particular safety requirements arise in the maintenance and retrofit of control code.
vii) The "models" of the process and plant used in the process engineering, automation and
operations scheduling are often informally defined, incomplete, ambiguous, and incompatible.
A major challenge consists in obtaining consistency between models used for various purposes.
viii) There is a tremendous scope for computer based integration of process design and
process operations activities.
ix) The major economic benefits for the design activity are to be found in shorter design-to-
operation cycles, lower capital and operating costs made possible by increased confidence in plant
operability, lower redesign and commissioning costs associated with increased quality of the
control system development process and lifetime maintenance of control code. On the operations
side, major benefits are associated with tighter integration between scheduling, batch management
and batch control functions. Depending on specific circumstances, benefits are to be gained from
increased plant and raw materials utilization, reduction of stocks and wastes, more competitive
marketing positions achievable through sharing of more accurate and timely information between
order acquisition, planning/scheduling and production execution functions, shorter order-to-
delivery cycles and overall more responsive plant operation.

3. Batch Integration Issues

Forms of Integration

The term integration has several different connotations. For example, Pekny et al. [39] identify
two different forms of integration: problem integration, which refers to the integration of software
tools used to solve different problems, and tool integration, which refers to the integration of
different software tools for solving the same problem. Perkins [40] differentiates between
software integration and the integration of life cycles aspects during process design. An example
of the latter foml of integration is the simultaneous design of a continuous process and its control
system. Here, the following classification will be used:
I) Integration by Communication. Computer programs may be linked to each other by
messages transmitted over a computer network. These messages may be requests for services, or
 755

responses to such requests from service providers. This form of integration is appropriate when
there is a clear division offunctions between the software tools being integrated. Typically, the
volume of information to be shared is small or of a transient nature. The closer coordination of
the modules being integrated should reduce delays and improve the responsiveness of the system.
ii) Integration through Shared Databases. Integration by communication over computer
networks is inappropriate when a permanent, consistent record of shared data is required. This is
the case in process design where, after one group of engineers has finished a design task, they
must pass on the design to another set of engineers using a different set of software tools. In
process operations, sales/marketing, scheduling and batch management functions need to share
data on the current and expected levels of stocks, etc. Use of a common database eliminates
manual transfers and the resulting inconsistencies.
iii) Integration ofAnalytical Methods. This refers to the integration of methods which permit
qualitatively different types of analysis. Some examples are the combined use of mathematical
programming and logical inference in process synthesis [44] and the use of pattern recognition
techniques in process control [54].
iv) Task integration. This refers to the merging of multiple design or operation tasks, e.g.
simultaneous production planning and detailed scheduling, simultaneous batch plant design and
scheduling [6] or simultaneous plant layout and equipment network synthesis. Task integration
can lead to improved designs and schedules by taking advantage ofthe interactions between plant
and operation aspects which would otherwise be considered separately.
For computer integration of batch processing operations, an architecture has been proposed
(ISA, 1992, discussed in more detail in a later section) which is due to become an international
standard by the end of 1993 (Figures 2a, 2b). This architecture could in principle implemented
in a variety of ways. We can envisage a shared database to provide common recipe information
to planning, scheduling and batch management functions. Communication links are required to
coordinate production scheduling and process control in real-time. Since most optimal scheduling
algorithms require a significant amount of computation time, a fast heuristic search method may
be used for on-line scheduling [13], while a more rigorous algorithm may be used for major off-
line schedule decisions. A batch management system may also be required to co-ordinate dynamic
scheduling with process monitoring tasks using, for example, statistical methods or neural
networks [55] for fault detection and classification. Thus, an integrated system for batch
processing may employ all four forms of integration described above.
 756

SCHEDULING

I Schedule I
Generation
• Master

I Recipe
Mana ement
I l Schedule
",chonule

Modification
I
Master
Recipe

BATCH MANAGEMENT

1 Recipe Dynamic I
Selection ~ I Scheduling

I Satch
Batch
Information 1 I Initiation I
t I
I
Process ~ ~ Commands
Information
I Equipment
Related
Control

Figure 2a. SP88 Architecture for Integrated 2b. SP88 Architecture for Batch Management
Operation of Batch Processes

Modeling requirements

This section focuses on the requirements for representing batch processes on a shared database.
Existing process engineering databases [1,14,42] employ a Process Flow Diagram as the focal
point for integration. This, however, does not provide an adequate description of operations and,
therefore, cannot be used to support most batch engineering applications. The adequate
representation of operating procedures is crucial to the design of a batch process engineering
database. Several aspects should be taken into account
i) The flexibility of batch operations results in a variety of plants ranging from dedicated
single product facilities to flexible multipurpose plants and pipeless plants. The use of
multipurpose equipment means that there is not a one-to-one mapping between unit operations
and physical equipment items. Hence, product recipes and operations must be represented
separately. A key issue is the representation of operations networks in a way which reflects the
flexibility of the batch mode of operation. Recent developments in incorporating layout
considerations in preliminary design [27] make it desirable to extend the simple equipment
network models that are commonly used for preliminary design to include information on the
spatial location of equipment.
 757

ii) A process may be successively viewed at different levels of resolution. The determination
of campaign lengths in multipurpose plants is typically carried out by aggregating groups of
batches into cycles with nominal production rates. Papageorgiu and Pantelides [38] show that the
performance of individual cycles may then be improved by allowing the reuse of equipment and
by exploiting intermediate storage. This step requires detailed scheduling within each cycle of
batches. The retiming of cycles can then trigger a recalculation of campaign lengths. Similarly,
Hasebe and Hashimoto [24] describe a scheduling procedure where groups of batches are
successively disaggregated and desegregated to create an overall production schedule.
iii) A further complication is that individual unit operations in a process may themselves be
viewed at many different levels of resolution. A unit operation is typically represented as a simple
elapsed time model for production schedule optimization, as a mixed discrete-eventldifferential-
algebraic model for dynamic simulation and optimal control studies [37], or as a finite state
machine for sequence control purposes.
iv) In batch processing, it is often not possible to completely separate the process
representation from the algorithms used for simulation, scheduling or optimization, since solution
methods often apply to narrow domains. For example, models may have to be simplified (e.g. by
ignoring transfer times, transfer lines, etc.) to reduce computation times for schedule optimization.
This imposes restrictions on product recipes and equipment networks and results in the need to
modify the model when transferring from the domain of one solver to that of another.
v) There are significant differences in the number and nature of database objects between
preliminary and detailed design stages [14]. During preliminary design, a number of alternative
plant and operation networks may have to be considered, with several alternatives maintained for
evaluation. The number of objects is small but the relationship between equipment topology,
recipes and operations networks and models of the unit operations is very complex, in particular,
if we wish to explicitly represent such models and the assumptions on which they are based [52].
Such models are represented by complex semantic networks and cannot easily be stored in
relational databases. During detailed design, on the other hand, the number of objects increases
rapidly as individual piping, instrumentation, electrical wiring, control devices, etc. are added in.
Many software tools exist, such as 3D drawing packages, typically with specialized data storage
formats to speed-up search and retrieval of objects. Again, a database of operations and detailed
control strategies, sequences, control instructions, will have to be defined and tied to the
equipment and control data. While databases for operations/control have been developed for
specialized in-house use (in particular by automation suppliers), there are at present no generic
integrated tools available to support the whole spectrum of batch engineering functions.
 758

Data Base Management Systems

Here, we consider the issue of database management systems suitable for supporting integrated
batch process engineering. Existing process engineering databases use a variety of data models
[5]. PRODABAS [I] uses an extended relational model, while Huang and Fan [25] advocate the
use of a hybrid object-relational model. Motard [36] reports on the relative merits of relational
and object-oriented database management systems. Marquardt [34] proposes the use of an object-
oriented database management system to support a process modeling environment. A
disadvantage ofrelational database management systems (RDBMS) is the difficulty in accessing
and manipulating complex hierarchies of objects, particularly if each object can have multiple
versions. The transaction model that is used to regulate the updating of records in a multi-user
setting is also inappropriate for design environments where transactions can take place over days
or weeks. Object oriented database management systems (ODBMS) have been developed
explicitly to support "non standard" application areas such as engineering design. In addition to
supporting object-oriented programming concepts such as object classes, inheritance and
encapsulation, these database management systems often provide storage for multiple versions
of objects and a locking mechanism appropriate for use in a design environment [29]. They are
usually designed to run on a distributed computing network of multiple file servers and client
workstations. Many ODBMS now provide interfaces to relational store managers, thus allowing
the user to employ object oriented data models while still being able to access data in a RDBMS.
Another option is to structure information as a knowledge base using a frame based system such
as KEE, which was used to implement MODEL.LA [52,53]. Frame based systems provide
excellent facilities for modeling complex relationships between data items through frames, slots,
daemons, etc. and are well suited for preliminary design. However, they are not well suited to
storing the large volumes of information that must be managed during the later phases of design.
Finally, special requirements arise during operations to share large amounts of data between
functions in real time. Now what is "real time" is clearly subjective and application dependent.
A practical definition is "so that results are produced just faster than they are needed". At the
lowest control level, this means milliseconds, at the unit and plant wide scheduling level,
seconds/minutes, at the planning/scheduling from minutes upwards. These requirements imply
that specialized architecture/softwarelhardware are typically needed for the more detailed control
levels so as to ensure such "real time" performance over ever wider domains and transaction
volumes. From a practical point of view, this demands that on-line applications must be adapted
(sometimes extensively) to these architecture. On the other hand, bridges between a real time
 759

environment and general purpose database management systems (e.g. a relational data base) are
often available. It is thus necessary to make a-priory decisions regarding which data and
applications will reside in the real time environment and which will not. These decisions clearly
fundamentally affect portability etc. across systems. There is much work underway in the
computing science and software communities on the development of ODBMS with real time
performance and cross systems portability. Although not much is presently available
commercially, it can be expected that significant progress will be made in this direction over a
five years horizon.

Existing Modeling Environments

Most existing modeling environments for process engineering are oriented towards continuous
processes. Much effort has recently gone into developing frameworks for creating and
maintaining complex process models. Some examples are MODEL.LA [52,53], ASCEND [41]
and VEDA [34]. MODEL.LA is notable for supporting multifaceted modeling [58] and in
particular I) the representation of processing systems at several levels of abstraction ii) the
automatic generation of basic relationship (equilibrium equations, balances, etc.) iii) the
representation of qualitative or semiqualitative relationships and iv) the documentation of
underlying assumptions. This language addresses one of the issues raised in the previous section,
namely the development of process models at varying levels ofresolution. However, for use in
batch process modeling further extensions are needed: I) to allow separate description of recipe
and equipment networks, and the permitted mappings between the two ii) additional semantic
relationships to represent the temporal sequencing of operations iii) representation of production
requirements and schedules and iv) the representation of combined discrete event/continuous
systems.
While modeling environments for continuous processes support both simulation and
optimization, environments for batch process modeling are specialized into two categories.
Packages such as UNIBATCH [21], BATCHES [8] and gPROMS [4] allow the process engineer
to build combined discrete-event/differential algebraic models for simulation studies. On the other
hand, a number of packages provide simplified models without process dynamics for use in
production scheduling and preliminary design. These include GANTT_KIT [22], BatchMaster
[9], SUPERBATCH [12] and gBSS [51]. The first three systems employ constraint satisfaction
strategies to generate a feasible schedule and perform some local optimization. They are
especially useful for performing schedule updates in a dynamic environment [II]. The gBSS
system, on the other hand, creates optimal schedules through the use of mixed integer linear
 760

programming techniques. No system is presently available that can address both design and
scheduling with plant wide batch production models and simulation or optimization with detailed
dynamic models.

4. Models and Software for Batch Processing

In this section, attention is focused on specific models for integrated batch processing,
preliminary design and scheduling, and batch management/control. No attempt is made to present
a comprehensive review of all the possible approaches and solution techniques. Instead, some
ideas and specific tools are described which have been developed over the last few years at
Imperial College and which will be used as "building blocks" for our integration project,
described in the last section.

Reference models for integrated batch processing

A number of organizations and individuals have analyzed the above problems or at least some of
its component parts and produced "reference models" for an integration architecture. For more
general considerations on integrated processing, the book of Rijnsdorp [46] and the work of
Williams [56] are useful. Specifically, for batch processing operation, a very useful analysis and
set of recommendations were proposed by the NAMUR committee. This organization, with
membership by a large set of European companies (mainly users), is aimed at the definition of
standard industrial practices in process measurements and control. A subcommittee formed in
1979 produced a set of recommendation specifically for batch control [49], presently being
revised. Much of their results were taken on board by the Instrument Society of America (ISA)
SP88 Committee (mainly suppliers), which is developing a standard terminology and architecture
for batch control operations. This is due to produce a final report by the end of 1993 with the view
of establishing an international ISO standard. Detailed interim documents have been produced
(ISA, 1992).
The major driving force for both activities is the need to agree standards so as to permit the
development of batch control and application software which is modular, highly flexible, can be
tested, maintained and easily evolve with changes in processes, plant and products. The major
NAMUR recommendations can be briefly summarized in the use of!) a hierarchical structure for
batch management and control systems (definitions of functions and data flows are given) ii)
standard terminology iii) the definition of "operations" independently from equipment, with a
hierarchical organization defined for operations and recipes iv) all application development by
 761

configuration rather than programming. The ISA SP88 committee extends the scope somewhat
at the supervisory control level which include scheduling (Figures 2a, 2b). The analysis produced
by both bodies is much more detailed for the lower level control functions, while the interactions
with off-line planning and scheduling remains a bit fuzzy. Nonetheless, the models produced have
had significant impact with users and vendors, with many commercial batch control systems
already offering compliance and user companies demanding it.
All the above "reference models" concentrate mainly on integrated operations or design of
control systems for a given batch plant with hardly any attention to the design of the plant itself.

Modeling for scheduling and preliminary design

State Task Networks. It is common to employ simple "elapsed time" models of unit
operations for preliminary design and scheduling of batch processes. Typically, the process is
represented as a directed graph where nodes represent unit operations and arcs represent
precedence relationships between Unit operations. The State Task Network representation
developed by Kondili et al. [30] uses a directed graph with two kind of nodes: "state" nodes,
representing unit operations and "task" nodes representing material states. The explicit inclusion
of state nodes in the network allows the description of several characteristic features of batch
chemical processes which are not manifested in discrete parts manufacturing: alternative
pathways for producing the same material; accumulation of material from successive batches;
batch splitting, recycles, etc. State Task Networks have been used as the representation
framework for defining a variety of optimization problems with formulations based on uniform
time discretization. These result in mixed integer linear programming problems. Problems
addressed include short term and cyclic scheduling [51], the determination of optimal campaign
cycles and schedules for each cycle [38] and the preliminary design of multipurpose plants [51].
Maximal State Task Networks. In its original form, the STN models have a number of
ambiguities, resulting from a coarse view ofthe scheduling problem (assignment of main process
vessels to tasks). For example, limitation due to potential connections between main processing
vessels are not considered and "states" are ambiguously identified with storage. An extension of
the concept which resolve these ambiguities was developed by Crooks [15] who proposed a
procedure for generating a maximal State Task Network (mSTN). This includes information
about connectivity and the representation of storage options: storage of input states in a
processing vessel; storage of output states after processing in a vessel; dedicated storage for input
and/or output states; explicit "storage" tasks which can be allocated to processing vessels. The
first step in the construction of a mSTN is to represent any legal movement of material from one
 762

processing vessel to another by a sequence of an output state associated to the upstream vessel,
a transfer task and an input state associated to the receiving vessel. The generation of the mSTN
is completed by defining one instance of each task, along with the corresponding input and output
states, for every piece of equipment in the plant on which the task may be performed (legal
assignment). Thus, the mSTN takes into account both the processing network and the plant
network.
Unit State Task Networks. A description of batch processes must also include information
regarding the state of processing equipment and the effect of carrying out operations on the
equipment. For example, Batch A may leave Unit UI in a dirty state, necessitating a cleaning
operation before Batch B may be processed in the same vessel. Such sequence constraints may
be important for detailed scheduling and are essential for control purposes. Interactions between
batches and processing units could be represented by Petri Nets [23,57] and various finite state
automata. Crooks [15] further extended the above mixed integer linear programming formulation
by introducing the explicit definition the legal states in which a processing unit may be (uStates)
and of the transitions between these uStates caused by processing and ancillary (e.g. cleaning)
tasks. This gives for each vessel a finite state automaton (a unit State Task Network, uSTN) which
may be used to represent efficiently a great variety of state and task precedence constraints, still
formulated as extended MILP problems.
The above scheduling models (described in detail in Crooks et al. [18,19,20]) may be
extended to design by introducing additional discrete variables to represent the existence or
otherwise of individual processing vessels [51], and connectivity networks [2], with continuous
variables representing size (of vessels or connections). With the addition of suitable equipment
and operating cost models (and modifications of the constraints), we can then address problems
related to minimum design (or retrofit) cost with preliminary or more detailed models. All of the
above models based on the STN representation, utilize a discretization of time into intervals of
fixed duration, with any events happening only at interval boundaries. While this may be useful
for preliminary design and scheduling, clearly discretization approximations are introduced,
especially when operations involving widely different time scales have to be considered. Thus,
short operations are normally neglected.

gBSS Overview

A software program (gBSS, for general Batch Scheduling System) which performs both
preliminary design and scheduling based on the STN approach described above was developed
and is described by [51]. It uses a number of techniqucs to improve the efficiency of a branch and
 763

bound method for the solution of the resulting MILP programs which enables the solution of
relatively large problems.

Modeling for batch operation and control

The type of models required for batch execution ranges for approximately the same level of detail
as used for plant wide scheduling (discussed in the previous section) to much more detailed
models for control of individual sequences and steps. A key aspect here is that while for design
and scheduling a single recipe, procedure, etc. define the processing of all batches of the same
type, execution of two instances of exactly the same recipe on exactly the same equipment will
give two "batches" with distinct identities, since they may be made up by distinct supplies of the
same ingredients, will typically have slightly different processing history, quality, etc. In addition,
basic recipes may have to be adjusted for individual batches, for example, to compensate for
known variations in raw material quality, reduced equipment perfonnance, etc. Batch control and
management systems must then be able to execute, track and monitor individual instances of each
"batch", for all batches of all products. The ISA recommendations include the definition of a
batch recipe model with generic master recipes, parameterised according to equipment type and
batch size, and for which the operating procedures are defined in terms of generic control
sequences. Control sequences are similarly defined in a generic, equipment independent fonn and
instantiated for specific equipment. Thus, a transfer from an upstream tank to one of 4 similar
downstream tanks may be described as a generic transfer phase (with associated parameters such
as rate, amount transferred, etc. and a list of upstream and downstream equipment types) of which
there are four specific instances. A master recipe is product and process dependent but equipment
independent. For individual batches, a master recipe is then instantiated into a control recipe,
which includes allocation of specific equipment, batch size, and with the generic control
sequences instantiated into specific sequences for the target equipment (control phases). The
NAMUR and ISA models represent many years of effort by the batch control vendors and
operators towards achieving control code which is structured, modular, easily reconfigured to
reflect recipe, equipment and production plans changes, portable across platforms and
maintainable in the long run. It is worthwhile noting that this effort generated a model based on
concepts essentially equivalent to classes and hierarchical inheritance which make it a natural
candidate for object oriented implementation. It may also be observed that if we defined the
parameters for master recipes and generic control phases in a suitable way, in tenn of required
resources, batch size, control parameters and if we additionally assigned a time estimate for the
completion of each generic control phase, this would allow their use for scheduling purposes as
 764

well as for control. However, models for execution and control purposes must be more
comprehensive and include all processing steps, including short transfers of material, etc.
Additionally, control models must be based on a continuous time representation. Thus, a number
of differences exist between the models required for preliminary design/scheduling and
execution/control.

SUPERBATCH Overview

A system, SUPERBATCH, has been developed at Imperial College over the last few years which
implements the supervisory batch management and scheduling functions ofthe ISA model. Based
on the ideas described in Cott and Macchietto [12], the system is broadly structured as in Figure
3. It is assumed that major economic decisions are made by some higher level
planning/scheduling function, which establishes the (possibly optimal) assignment of batches to
processing units, batch sizes and the sequence of batches. If these decisions have been made by

save changes
allocated
import monitor plan
changes
Plant defn
master recipes , Change
,, Monitor
plan ,
Plan
Initial status ,
changes ,
res. availability

save
Control
: checkpoint '
~------------------~ System ,:
,,
L ____________ I
restore

,,
operator " ,________ ~ _,
displays ,~- - - - - - -:, Alarm list :,
....
, - - -- ------------,
.------ -----
.................... : Audit Trail :
~ ,
1_ _ _ _ _ _ - - - - - - :

I Base Code I :-Sy~t~~- d~p-~~d~~t:

L _________________ I

Figure 3. SUPERB ATCH - system overview

 765

some external scheduler (either manually or algorithmically), they are imported to the system and
verified for feasibility, not just at the initial time but for the entire production horizon, utilizing
the current plant, procedure and constraint models within SUPERBATCH and information on the
state of the current production in the plant, expected availability of all resources, etc. Any missing
decisions or conflicts may also be resolved or changed interactively in an off-line scheduling
mode (using the "off-line planner" in Figure 3), by viewing the resulting schedules, utilization
statistics, etc. until a satisfactory "production plan" is obtained. This can then be passed on to the
on-line portion of the system (the "monitor" in Figure 3) which implements the batch
management functions of the ISA model. The monitor runs in a real-time environment and
processes any new requests as changes to the current schedule. It sets the parameters for and
initiates control phases as and when required by sending commands to (possibly distributed)
control system(s). It then monitors their external execution and revises the schedule for all
remaining production in view of actual as opposed to planned duration, amounts transferred, etc.
The schedule revisions are done at a frequency of, typically, one minute or less utilizing a
completion time algorithm based on minimal algebra [13]. In order to insure fast solutions
(essential for real-time execution) the algorithm deals with only a subset of all possible scheduling
decisions. In particular, it retains (with minor exceptions) the batch sequence order, sizes and
allocations last specified, and revises start and finish times to meet a set of detailed operational
constraints. Thus, it can ensure detailed operational feasibility but not economic optimality. At
anyone time, however, the complete production schedule can be made available to both the off-
line planner and external scheduler for deciding on major scheduling decision changes. The same
algorithms are used in the on-line monitor and off-line planner.
The system is completely driven by a hierarchical model which first defines the plant, recipes
and control phases in terms of generic resources (for both equipment and operations). Each
resource is then parameterised and is instantiated as individual equipment units, control phases,
etc. (some of the terminology used was based on earlier versions of the ISA model, thus ISA's
master and control recipes are called master procedures and batch procedures, respectively, in
SUPERBATCH, etc. - the concepts are however respected). The above design assumes that
phases are delegated for execution to external controllers and this defines the level of supervisory
vs. distributed control and the "granularity' of the supervisory management models (Figure 4).
The interface between supervisory system and external controllers (e.g. the mapping between
phase parameters and specific control tags, devices, etc.) is also part ofthe SUPERBATCH model
definition and therefore external controllers may be easily changed.
 766

An implementation of this design (into an IBM product called ASSYST, for Adaptive
Scheduling System) within a fully integrated plant wide industrial control environment (IBM's
RTPMS) has been produced which permits access to all portions of the models and production
data at any time from a proprietary shared real time database. Thus, multiple users may
simultaneously access the current actual production status (in various forms, induding Gantt
charts) for all executing and scheduled batches, individually examine alternative scenarios in an
off-line mode (changes may range form simple additions and deletions of batches, to changes in
the recipes and plant configurations) and, subject to authorization, promote changes to the on-line
system for execution. It is possible to use all the standard industrial system functions to produce,
for example, user interfaces, detailed history, alarms, etc., to communicate with external relational
databases for corporate accounting or other functions, and to communicate with external "user
programs" running synchronously for heavier numerical work. The external communication
mechanism may be used for interfacing more sophisticated off-line schedule optimisers (to carry
out optimal scheduling functions) and dynamic simulators (so as to enable running a simulation
instead of the actual plant), as described in Macchietto et al. [33]. Some applications ofthis type
were developed by Cott [10] and Kassianides [28], who demonstrated the execution of
supervisory batch control function in conjunction with both Speedup simulations and actual
equipment (an older pilot plant in the Department of Chemical Engineering).
The basic ideas and algorithms were also incorporated into a software product (APV Baker's

manual
expert system
GOALSETIING mathematical optimisation
other??

SUPERVISION

EXECUTION

Figure 4. Control/Management Environment

 767

BatchManager), which provides graphic interfaces for interactive use in a UNIX environment,
with all models definitions held in a relational database and with interfaces to order acceptance
systems. At present, Batch Manager only works in an off-line mode, but work is under way to
integrate it with the ACCOS plant wide automation and control systems which APV supplies
mainly to the food industry.

5. Integration Project

In this section, we describe initial work towards the development of a system which addresses in
particular three of the major problems identified in the industrial survey study: integration of
preliminary plant/process design with operations models, integration of automation and control
system development with plant/process design and integration of batch design tools with batch
management/execution tools. The focus will be on the general aspects and on an example of
recent results. Details on individual aspects will be presented elsewhere.
The approach has been as follows:
I) For preliminary design, the approach based on the SIN representation provides a good
basis for simultaneous consideration of plant, processes and market interactions, for a sufficiently
wide class of problems. In the very early stages, simple models of plant and operations may be
considered (major processing vessels). These can be evolved into rather more detailed models to
include the actual connectivity of the plant, much more detailed operational constraints, etc. using
the extensions of Crooks [15] and Barbosa and Macchietto [2], outlined in a previous section
(mSIN, uSIN, etc.).
ii) For batch operations management/control, the NAMUR and ISA SP88 models define a
design basis which is both appropriate and which cannot be ignored, since it has been and will
be adopted by most industrial batch control systems. Most of the interesting work lies at the
supervisory batch management level, in the dynamic scheduling and in the coordination with off-
line planning/scheduling. For these activities, the approach of Cott and Macchietto [12] is
adopted.
iii) Techniques for the development of automation and control systems should utilize initially
the same plant/process/constraint models used for process/plant design (supplemented later by
details of piping, valves, instrumentation, etc. arrangements and more detailed control objectives,
strategies, constraints) and produce eventually models for batch control and supervision in a form
compliant with NAMURlSP88. This should guarantee that the control structures and sequences
generated can be passed on to and executed by standard industrial batch operation systems. The
 768

required machinery was developed by Crooks [15] and will be briefly described in the next
section [16,17,18,19,20] . A prototype software as developed (CAPS) which implements those
ideas.
iv) Different representation are clearly best for addressing different problems. The
conversations of mode presentations between different domains and data
aggregation/desegregation were given particular attention so as to achieve consistency.

CAPS overview

A functional overview of the CAPS (Computer Aided Procedure Synthesis) system of Crooks
[15] is given in Figure 5. The problem is posed as follows. Given: a description of the plant and
processes involved, operating goals and constraints identify I) the control sequences (phases),
ii) the master procedure networks, made up by those phases and iii) detailed sequence control
instructions for each phase so as to achieve one or more desired operating goal(s), subject to all
restrictions posed by equipment and operations configurations and other constraint specifications.
Where parallel operations are possible, control phases and master procedures should be produced
both in a generic, parameterised form and also in instantiated form for each of the possible
equipment allocations. Since there is often more than one way of achieving a goal , a choice
should be made according to some objective function. Typical goals may be to transfer a quantity
of material from one end of the plant to another while leaving any equipment used in a clean state,

@ -u..,lIfI'¥'a.4odd~
S.~oaJ i ",
(D<uilod Sd.odul;o,l
0-1~1ld"1I

Unit-Stat.;. T ... "

Nc:t-.riu
(USTf<)
E'.qUlttMU\(
N.t.(~ks
(P&ID)
v...--op«ir0«1
S<-qQ~ 'rUo.
Sy"lhc~s o( C ~lrot
SoqIKnCC S~cjr.c:=a«.iOt'ls

Figure 5_ CAPS - system overview

 769

to produce an intermediate product according to a specified recipe, etc. The solution approach
followed consists in a sequence of three main steps. First, the problem is formulated as a detailed
task scheduling problem, using the STN representations and modeling extensions previously
discussed, and solved as a MILP problem. Plant items which are functionally equivalent are
represented at this stage as a single generic unit type (a resource). This produces a generalization
of all possible schedules of the tasks required to achieve the goals. We may note that this can be
done even before the exact number of equipment units in a resource is fixed, i.e. very early in the
design stage. This information is then processed in a second step where generic control phases
and master procedures are identified, according to the SP88 guidelines and parameterised
according to equipment resources, control parameters, etc. The boundaries between phases are
identified so as to enable distributed control and distinct master procedures are chosen so as to
enable monitoring of batch identity (usually destroyed by batch merging and splitting operations).
Tasks may be aggregated into phases in various ways as required for control purposes. For
example, all the tasks involved in a recirculation operation, or in the simultaneous feed of several
ingredients must be carried out together and will constitute a single control phase. Any additional
control phases which were not present in the original gBSS model (e.g. small transfers) are
identified and added. More detailed flowsheet information, when it becomes available from
design, enables identifYing multiple instances of the generic control phases. These first two steps
broadly accomplish what has been called "subgoaling" in the procedure synthesis literature.
Other researchers have approached this problem for continuous plants using a variety of
methods for "planning" the required operation sequences. The main features of the CAPS
approach are the use of an operations model which is sufficiently rich so as to reflect all important
interactions, the use of a numerical technique which is very good at considering complex
interactions and interacting goals, the choice of balance between use of infom1ation about desired
operations which are easily supplied by the user and information which are more easily generated
by the computer, and the use of industrial guidelines to help in decomposing the problem into
sensible components.
Finally, the third step of CAPS utilizes much more detailed information about the plant
network (a process and instrumentation diagram, P&ID) to develop rather detailed sequence
control specifications for each phase in terms of typical discrete and continuous cont~ol primitives
(e.g. open/close valve, check item, put control loop on automatic at setpoint x, monitor logical
condition y, etc.). Because individual control phases are now largely independent (their
interactions having already been identified), very detailed searches and extensive practical control
knowledge and constraints may be applied to this step, which is accomplished using a rule-based
 770

method. The final output from this step is a specification for each control phase. For execution,
individual specifications will have to be translated into actual sequence control code for a target
control system. In principle, this step could also be automated, although we have not yet done it
(all control systems are slightly different, requiring one translator each). The specifications,
however, are sufficiently low level that automatic translation should be rather easy. In practice,
we have manually implemented some of the control phases on two industrial control systems to
which we have access (RTPMS, by IBM and ACCOS, by APV Baker).
The output of CAPS can be produced in principle in a variety of ways. One implementation
which we have developed produces all the definitions needed to define a SUPERBATCH model
ofthe plant, master procedures and control phases, in the required formats. These can be inputted
directly to the supervisory system. It is then possible for an operator to introduce a production
plan (however generated), involving any number of batches for which the corresponding recipes
and control phases have been synthesized, and ask SUPERBATCH to schedule it. These
production plans may of course include different and much more complex operation scenarios
than were initially used in the optimal plant design. If individual control phases have been defined
(by CAPS), but these have not yet been implemented in a control system, then only off-line
scheduling exercises can be carried out. This will enable assessing the actual schedules, detailed
plant operations, performance measures, etc. obtained in a continuous time representation. As
already noted, this can be done very early on in the design stage. When the control phases are
implemented in a control system, SUPERBATCH can then initiate their execution and carry out
all its other on-line supervisory control and dynamic scheduling functions. The controllers can
manipulate actual control devices and receive actual signals from the plants, but, initially, it is
also possible to drive just a simulation (either a simple or a more detailed dynamic simulation,
possibly with random variations also modeled, as described by Colt and Macchietto [18]. The
performance of the actual batch and supervisory control systems can then be assessed early on
with a simulated plant. Later on, when the plant is built, actual end control devices and signals
may be connected to each control phase and exactly the same supervisory and batch control
systems will drive the actual plant. This switch is made very simple since all control variables and
parameters used for each control phase are not hardwired but explicitly declared as part of a
SUPERBATCH model.
Overall, the approach described permits going almost automatically all the way from
preliminary design of a batch plant taking into account operations models, to detailed design of
batch operation sequences and control procedures at various levels of abstraction, to execution
 771

of the same procedures in an on-line system which features supervisory batch management and
dynamic scheduling.
If we now assume that a plant configuration is fixed and tum the attention to the usual
operations problem of generating and executing an optimal production schedule, we may in
principle use again the gBSS plant and operations models already developed at the design stage,
specifY a new production objective, new cost parameters, etc. and calculate an optimal schedule
for the new operation problem. If we go again through step 2 of CAPS, we will have one of two
situations: either the new production can be carried out using master procedures and control
phases which had been already identified, in which case CAPS will just determine the new
parameters (batch sizes, transfer rates, etc.) to be used with the existing procedures, or some novel
processing arrangement is to be done which was never done before (this may be the case if a
product recipe variation is introduced, or an alternative route is proposed as optimal). In this case,
new control phases may have to be developed, and possibly entire new master procedures. CAPS
was designed so that these could be incrementally added to its procedure data base. A similar
situation will arise if the equipment is modified (say, a new reactor is added in parallel to an
existing one). In some cases, it is only sufficient to generate new instances of existing gencric
control phases. In others, it is possible to generate new master procedures as a different network
of existing phases. Finally, for more radical changes, it will be necessary to generate some
completely new control phases and master procedures. It should be noted that even thc simple
addition of a new procedure may require changing several other ones, for example to include
interlocks.

6. Application Example - Batch Pilot Plant

In this section, an example of the ideas outlined above is given, which is used to illustrate the
current progress. The example refers to our own batch pilot plant (described in Macchietto [31]),
a schematic diagram of which is given in Figure 6. This plant is simple in terms of its basic
processing units, but a great deal of complexity is added by the presence of cleaning operations
(partially sharing the same pipework used by the process), operations involving external
recirculation loops, two way flow in some pipes, transfer paths which can be configured through
flexible valve manifolds (so called Flow Separation Plates), as is common in the food industry.
In the initial design stages, a simplified view of the plant may be considered with only the major
equipment items, and representative processes used for preliminary sizing, etc. This is well within
the capabilities of the optimal gBSS software. As the requirements are refined, details may be
 772

Figure 6. Batch Pilot Plant Process and Instrumentation Diagram

added to the model to include connections and more detailed operational constraints. For
scheduling purposes, such a model was developed by Crooks [15] which could be used to
optimally schedule a small number of batches in greater detail. This model was used as input to
CAPS to generate the control procedures for batch management and detailed sequence
specification. One of the basic processes carried out in the plant is an enzymatic starch conversion
to sugars. With a goal specified of producing one batch of glucose, a single master procedure was
identified by CAPS, with 14 control phases (with associated resource requirements, precedence
network between phases, phase parameters, etc.) which corresponded very well with the design
of the actual control system (one of this is an external recirculation of the starch from reactor T3
through heat exchanger HI for temperature control purposes). This exercise was then carried out
for the rather complex cleaning operations, defined by an STN model. Here, the synthesis goal
specified was simply to change the state of a target processing vessel from "dirty" to 'clean". The
initial goal of cleaning T3, for example, resulted in 4 control procedures (preparation of hot
cleaning-in-place (CIP) solution, water rinse T3, detergent wash T3 and final check of cleaning
solution strength). To carry out a T3 clean operation alone, these sequences are scheduled
sequentially as CIP preparation, water rinse T3, detergent wash T3, water rinse T3, final check.
The water rinse sequence is only generated once but used twice with different parameters, for pre
and post detergent wash, as required by the cleaning STN definition. If we continue the exercise
 773

~
j ._ II

P1
12

,.
P2
,l:,'n
P5 +..~
++

. . . . . . . . . . . . . . . . . . . . . . . . .. ..
1<:.
T5

....... .::-:+-::+..... ....................... .. . . . . . . . ._................._......._._....__......__.................

P9

OW_RTN
PROO_ORAN t=~~=. ._._
~.-...~
_.-~--~ ~
_____ •. _.•.. ___~ ~~~____
... __ ................ ~~
.. ___~ -~.______
.. __ -~.-~~ -~-~~. __~-
•. ________ ~.______
._ ~~-~-..
!>WI ..........•_ •••..•.•. _... _.._.•.._ ...•.__......•_ .•

~~of=~~==~~~~====ro~======~OO~======;OO~======~~~

Figure 7. SUPERBATCH Detailed Schedule for Glucose Production and Cleaning-in-Place (CIP)
Operations (on horizonal axis, time in minutes for start)

and specify, as new synthesis goal, the cleaning of Tl-T2 (modeled as a single resource type),
CAPS generates two additional procedures only, water rinse Tl-2 and detergent wash Tl-2. The
CIP preparation and final check sequences are also identified as necessary, but these are
recognized as being already available in the CAPS procedure database. To carry out a Tl-T2
clean alone the sequences are scheduled in the sequence CIP preparation, water rinse Tl-2,
detergent rinse Tl-2, water rinse Tl-2, final check. With a flowsheet showing two vessels for the
resource Tl-2, CAPS will produce two instances for each of the generic water and detergent rinse
phases. We can now prepare production plans consisting of batches of glucose and cleaning
cycles, as required, and produce a corresponding operations schedule with SUPERBATCH. For
example, a detailed schedule (of the flexible connecting paths as well as of the main and service
vessels) for a single batch of glucose with final cleaning of the main reactor T3 is shown in Figure
7. The complex resource contentions during all processing and transfer steps are well handled.
In the next step, the detailed P&ID diagram of the plant is utilized by CAPS to generate the
detailed control specifications required to carry out each of the control phases so far identified.
A typical output of CAPS is shown in Figure 8 for one of the glucose production phases (the
addition of starch through a funnel with external recirculation of the reactor contents). Again, the
specification produced compared well with those for the corresponding sequence in the actual
plant. This example is discussed in detail in a forthcoming publication (Crooks and Macchietto,
1993b). To actually execute the sequence, we would need to translate this specification into
 774

Seq~ SEQ..U]HASES
Com:spoods to Unit Phase U]HASES and Resoun:e Phase R_PHASES
Weight Panm= c 9.9. based on FUNNEL loss
Duration c [10.0. TIME]
Material Transf=
~marrATION~SHRBL~
From: D To: P3 State:: MESS2_I Amount: 89.1
Path: £T3,ABVU.PI_2.FB3.PU.AVI_19.P3]
From: P3 To: D State:: MESSI_I Amount: 99.0
Path: [P3.AVI_21.A VI_35.P2_5.FB4.P2_1,AVI_28.D]
LINEAR TRANSHR BLOCK:
From: Funnel To: P3 State:: STARCH Amount: 9.9
Path: [FUNNEL.BVI_I.P3]
S1N Tasks performed:
Task: REcmC ST Unit: TI Amount: 99.0
Task: ADD_STARCH Unit: P3 Constant Rate from subgoaler. 9.9
The following sequcn= are inhibited by this sequence:
[SEQ..U]HASE34.23.21.21.17.16.15.12.ll.10.9.8.7.6.4.I]
Pn:.Q>eCl:s ror the sequence are:
Ensure nowpla!e bend: FE7 is disconnected.
Ensure flov..plate bend: FE5 is disconnected.
Ensure nowplate bend: FE I is disconnected.
Ensure flowplate bend: FE3 is in place.
Ensure flowplate bend: FB4 is in place.
Unit Funnel minimum level: 9.9
Unit D !="UState #1: CLEAN
Close and Inhibit the following valves:
[AVl_20.AV 1_2.A V 1_32,B V 1_2,SSVl_2]
InIullit the following flowplate bends which must not be used during the sequence:
[FB 1,FB5,FB7]
Inhibit the following items foc use in the sequence:
[AVI_21.A VU5.P2_5.FB4.P2_I.AVl_28.T3,ABVU.PI_2.FB3.P13.AVU9.
FUNNEL,BVU .P3]
Open valves: [A V 1_21.A V 1_35.A VI_28,ABVl_l,AVI_19,BVU]
Stan the following devices:
P3 with a nominal flowrate of: 20
Agitatoc AGI
Tennination Condition(s) foc the sequence:
1. Mta: nominal time: 10.0
Tenninatioo lnstructions for the sequence:
Stop agitator AG I
SlOp flow device; P3
Oose valves: [A VI_21.AVI_35.AVI_28,ABVl_l,AVU9,BVU]
Uninhibit the following plant items:
[AV1_21,A VU5.P2_5.FB4.P2_I.AVI_28.T3,ABVU.Pl_2.FB3.Pl_4,AVU9.
FUNNEL,BVU.P3.FB 1.FB5.FB7,AVI_20.AVl_2.AV U2.BV1_2.SSVl_2]
Post~nditions are:
Unit D post-US tate '1: DIRTY

Figure 8. Control Sequence Specification produced by CAPS for recirculation Addition of Starch

control code. The pilot plant is controlled by an ACCOS system, by APV, with its own sequence
control language (P ARACODE). Assuming this done for all the required sequences, we could
then run the plant from the same SUPERBATCH model. As already mentioned, we are presently
working on an ACCOS/SUPERBATCH interface, and plan to complete this demonstration
shortly. Two other application demonstrations have been developed, where the plants are
simulated by a dynamic model (running under the control of the RTPMS system). Here, we were
able to show the final execution and supervisory management of a complete production plan.
 775

These applications, involving an industrial dairy plant and a multigrade PVC process, are
presented by Crooks and Macchietto [19] and Bretelle, Chua and Macchietto [7]. Some details
of the dairy plant application are also given in Crooks et al. [15,16,17].
Finally, the retrofit design of our pilot plant is discussed by Barbosa and Macchietto [2]. The
MILP model of Crooks is utilized again, with additional design variables, to explore the minimum
cost modification of the plant so as to permit the inclusion of a novel process. The optimal
solution, involving the installation of double seat valves instead of the flow separation plates and
some rerouting of pipes, is presently being implemented in the plant.

7. Discussion and Conclusions

Clearly, proper integration of all batch processing activities is quite a formidable undertaking and
achieving it is still quite a way away. We still need far better and more powerful tools to deal with
each of the individual problems involved. Optimal design and scheduling techniques are still
haunted by combinatorial complexity, although much progress has been achieved recently. On-
line dynamic scheduling could do with more powerful algorithms so as to deal with more (all?)
scheduling and control decisions in real time. Techniques for dealing with the automatic
development of operating and control procedures from basic design information, in spite of my
personal excitement at the results discussed above, are still rudimentary if compared to the level
of sophistication required, say, to pass a strict safety audit. Proper and formal validation
techniques will no doubt have to be used, and some work in this direction is indeed underway
[35]. Finally, the tools required to support a flexible and powerful integration infrastructure
(object oriented data base management systems, modeling environments, etc.) still need further
development before they can be routinely used for the job at hand. More comprehensive models
and representations are still needed at pretty much all levels to cover in particular the boundaries
between classic, well defined but isolated problems.
The collaborative study with our industrial friends confirmed that the separation on-line
management/control and off-line design/scheduling is a major source of problems (or, conversely,
a very rich ground for potential advances). The recent work [15] on procedure synthesis, in
particular, has forced us to take a very hard look at these two aspects of batch processing, at the
underlying assumptions and at development the consistent models. Integration techniques utilized
ranged from the use of shared databases, task integration and integration of analytical techniques.
The results obtained indicate that, perhaps for a limited set of problems and activities, some
interesting steps can already be achieved towards an integrated system for batch processing.
 776

Acknowledgment

This work was supported by SERC/AFRC and a Commonwealth Institute scholarship (CAC).
Industrial partners of the Batch Processing Project, in particular APV Baker, BNFL Ltd. IBM UK
and John Brown Ltd. are warmly thanked.

References

I. Angus, C. J. and P. Winter (1985). An Engineering Database for Process Design. IChemE Symp. Series No.
92, Pergamon Press, Oxford.
2. Barbosa-Povoa, A. and S. Macchietto (1992). Design of Multipurpose Batch Plants - I. Problem Formulation,
Computers Chern. Engng., 17S, 33-38.
3. Barbosa-Povoa, A. and S. Macchietto (1993) Redesign of a Multipurpose Batch Pilot Plant with Cleaning in
Place (CIP) Integration. Proc. ESCAPE 3, Graz (Austria), July 5-7.
4. Barton, P. and C. C. Pantelides (1991). The Modeling and Simulation of Combined Discrete/Continuous
Processes. Proc., PSE'91, Montebello, Canada.
5. Benayoune, M. and P. E. Preece (1987). Review of Information Management Systems in Computer-Aided
Engineering. Computers. Chern. Engng., 11, pp. 1-6.
6. Birewar, D.B. and IE. Grossman (1990). Simultaneous Synthesis, Sizing, and Scheduling of Multiproduct Batch
Plants. Ind. Eng. Chern. Res., 22, 2242-2251.
7. Bretelle, D., E. S. Chua and S. Macchietto (\994). Simulation and on-line operation ofa PVC Process for
Operator, Computers Chern Engng, 18S, pp. 547-551
8. Clark, S. M. and G. S. Joglekar (1992). General and Special Purpose Simulation Software for Batch Process
Engineering. NATO Advanced Study Institute on Batch Processing Systems Engineering, Antalya, Turkey.
9. Cherry et al. (1985). Proc. PSE '85, (F. A. Perris, ed.), IChemE Symp. Series No. 92.
10. Cott, B. J. (\ 989). An Integrated Management System for the Operation of Multipurpose Batch Plants. PhD
Thesis, Imperial College, University of London.
11. Cott, B.J. and S. Macchietto (I 989a). Minimising the effects of batch process variability using on-line schedule
modification Computers. Chern. Engng., 13, Y" pp. 105-113.
12. Cott, BJ. and S. Macchietto (1989b). An Integrated Approach to Computer-Aided Operation of Batch Chemical
Plants. Computers. Chern. Engng., 13, 11112, 1263-1271.
13. Cott, BJ. and S. Macchietto (1989c). A General Completion Time Determination Algorithm for Batch
Processes. presented at AIChE Annual Meeting, San Francisco.
14. Craft, J. (1985). The impact of CAD and Database Techniques in Process Engineering. IChemE Symp. Series
No. 92, Pergamon Press, Oxford.
15. Crooks, C. A. (1992). Synthesis of Operating Procedures for Chemical Plants. PhD Thesis, Imperial College,
University of London.
16. Crooks, C.A., K. Kuriyan and S. Macchietto (1992). Integration of Batch Plant Design, Automation and
Operation Software Tools, Computers Chern. Engng., 16S, pp 289-296.
17. Crooks, C.A. and S. Macchietto (\ 992). A Combined MILP and Logic-Based Approach to the Synthesis of
Operating Procedures for Batch Plants. Chern. Eng. Comm.,II, pp. 117-144.
18. Crooks, C.A. and S. Macchietto (1993a). The Synthesis of Operating Procedures for Batch and Semi-continuous
Chemical Plants - I. The Method. Submitted for Publication.
19. Crooks, C.A. and S. Macchietto (\993b). The Synthesis of Operating Procedures for Batch and Semi-
continuous Chemical Plants - II. Applications. Submitted for Publication.
20. Crooks, C.A., N. Shah, C. C. Pantelides and S. Macchietto {I 993). Detailed Scheduling of General Batch and
Semi-continuous Operations. Submitted for Publication.
21. Czulek, A. J. (\ 988). An Experimental Simulator for Batch Chemical Processes. Computers. Chern. Engng.,
12,2/3, pp. 253-259.
22. Halasz, L., M. Hofmeister and D. W. T. Rippin {I 992). Gantt-Kit - An Interactive Scheduling Tool. NATO
Advanced Study Institute on Batch Processing Systems Engineering, This volume pp. 706.
 777

23. Hanish, H. M. (1992). Coordination Control Modeling in Batch Production Systems by Means of Petri Nets.
Computers Chern. Engng., 16, I, pp. 1-10.
24. Hasebe, S. and I. Hashimoto (1992). Present Status of Batch Process Systems Engineering in Japan, NATO
Advanced Study Institute on Batch Processing Systems Engineering, This volume pp. 49.
25. Huang, Y. W. and L. T. Fan (1988). Designing an Object-Relational Hybrid Database for Chemical Process
Engineering. Computers. Chern. Engng., 12,9110, pp. 973-983.
26. ISA-dS88.0 I (1992). Batch Control Systems: Models and Terminology, Draft 5, ISA, Triangle Park, NC, USA.
27. Jayakumar and G. V. Reklaitis (1991). Graph Partitioning with multiple Property Constraints in Multifloor
Batch Chemical Plant Layout. Paper presented at AIChE National Meeting, Los Angeles.
28. Kassianides, S. (1991). An Integrated System for Computer-Based Training of Process Operators. PhD Thesis,
Imperial College, University of London.
29. Kim, W. (1990). An Introduction to Object-Oriented Databases, MIT Press, Cambridge.
30. Kondili, E., C.c. Pantelides, R. W.H. Sargent (1988). A General Algorithm for Scheduling of Batch Operations.
Proceedings 3rd IntI. Symp. on Process Systems Engineering, pp. 62-75. Sydney, Australia.
3 I. Macchietto, S. (1992a). Automation Research on a Food Processing Pilot Plant, IChemE Symp. Series No. 126,
pp.l79-189.
32. Macchietto, S. (1992b).lnteractions between Design and Operation of Batch Plants. pp. 113-126 in Interactions
Between Process Design and Process Control, IFAC Workshop, J.D. Perkins ed.,Pergamon Press (1992)
33. Macchietto, S., M. Matzopoulos and G. Stuart (1988). On-line simulation and optimization as an aid to plant
operation-the easy way", pp. 669-677, in Computer Aided Process Operations, (G.Y.Reklaitis and H.D. Spriggs
eds.), Elsevier.
34. Marquardt, W. (1992). An Object-Oriented Representation of Structured Process Models. Computers. Chern.
Engng., 16.
35. Moon, I. G. J. Powers, J. R. Burch and E. M. Clarke (1992). Automatic Verification of Sequential Control
Systems using Temporal Logic. AIChE J. 38, I, p. 67.
36. Motard, R. L. (1989). Integrated Computer-Aided Process Engineering. Computers. Chern. Engng., 13, 11112,
pp.1199-1206.
37. Mujtaba,I. M. and S. Macchietto (1992). An Optimal Recycle Policy for Multicomponent Batch Distillation.
Computers Chern. Engng., ICS, pp.273-280.
38. Papageorgiu, L. G. and C. C. Pantel ides (1992). A Hierarchical Approach for Campaign Planning of
MUltipurpose Batch Plants, Computers Chern. Engng., 17S.
39. Pekny, J., V. Venkatasubramanian and G. V. Reklaitis (1991). Prospects for Computer-Aided Process
Operations in the Process Industries. Proc. COPE '91, Elsevier, pp. 435-446.
40. Perkins, J. D. (1992). Computer-Integrated Process Design - Status and Trends. Computers Chern. Engng., 16S.
41. Piela, P. C., T. G. Epperly, K. M. Westerberg and A. W. Westerberg (1991). ASCEND: An Object Oriented
Computer Environment for Modeling and Analysis. The Modeling Language. Computers. Chern. Engng., 15,
I, pp.53-72
42. Preston, M. L. (1990). Integrated Process Plant Design. Proc. FOCAPD III, CACHE-Elsevier.
43. Puigjaner, L., A. Espuna, I. Palou and J. Torres (1991). A prototype Computer Integrated Manufacturing
Modular Unit for Education and Training in Batch Process Operations. Proc. COPE '91, Elsevier, pp. 427-432.
44. Raman, R. and I. E. Grossmann (1991). Relation between MILP Modeling and Logical Inferencing for
Chemical Process Synthesis. Computers Chern. Engng., 15,2, pp. 73-84.
45. Reklaitis, G.V. (1991). Perspectives on Scheduling and Planning of Process Operations. Proceedings 4th IntI.
Symp. on Process Systems Engineering, Montebello, Quebec, Canada
46. Rijnsdorp, J. E. (1991). Integrated Process Control and Automation. Elsevier.
47. Rippin, D. W. T. (1991). Control of Batch Processes. Proc. IEAC Symp. Dycord+ '89, Maastricht, The
Netherlands, pp. 115-125.
48. Rosenof, P. H. and A. Gosh (1987). Batch Process Automation - Theory and Practice. Van Nostrand Reinhold
Co., New York.
49. Sawyer, P. Computer-Controlled Batch Processing. to be published by IChemE, Rugby, UK (Personal
Communication) .
50. Shah, N. (1992). Efficient Scheduling, Planning and Design of MUltipurpose Batch Plants. PhD Thesis, Imperial
College, University of London.
51. Shah, N., C.c. Pantelides and R.W.H. Sargent (1992). A General Algorithm for Short-Term Scheduling of
Batch Operations - II. Computational issues. Computers. Chern. Engng., 17,2, pp 229-244.
 778

52. Stephanopoulos, G., G. Henning and H. Leone (1990a). MODEL.LA. A Modeling Language for Process
Engineering - I. The formal Framework. Computers Chern. Engng., 14,8, pp. 813-846.
53. Stephanopoulos, G., G. Henning and H. Leone (1990b). MODEL.LA. A Modeling Language for Process
Engineering - II. Multifaceted Modeling of Processing Systems. Computers Chern. Engng., 14,8, pp.847-869.
54. Stephanopoulos, G. (1991). Towards the Intelligent Controller: Formal Integration of Pattern Recognition with
Control Theory. Proc. CPC IV.
55. Venkatasubramanian, V. (1991). Recall and Generalization Performances of Neural Networks for Process
Fault Diagnosis. Proc. CPC IV., pp. 647-664.
56. Williams, TJ., (1989) (editor). A Reference Model for Computer Integrated Manufacturing, A Description
from the Viewpoint of Industrial Automation. Instrument Society of America, Research Triangle park, NC.
57. Yamalidou, E. C. and J. Kantor (1990). Modeling discrete-Event Dynamical Systems for Chemical Process
Control- A survey of Several New Techniques. Computers Chern. Engng., 14,3, pp.281-299.
58. Ziegler, B. P. (1984). Multifaceted Modeling and Discrete Event Simulation. Academic Press, London.
An Interval-Based Mathematical Model for the Scheduling
of Resource-Constrained Batch Chemical Processes

M.G. Zenmer and Gintaras V. Reklaitis

School ofChemicai Engineering. Purdue University, West Lafayette, IN 47907, USA

Abstract: This paper is an extension of previously presented work concerned with the use of
. interval analysis and search techniques for the exact modelling of the solution space of
resource-constrained batch chemical process scheduling problems. A mathematical
programming formulation is presented which is based on the notion of interval analysis and
which uses a non-uniform discretization of time derived only from real process events. No
additional discretization variables are introduced. While conventional discretization models
rely on the rounding of problem data in order to reduce the size of the formulation, the base
formulation proposed here does not require such data modification. The importance of this
difference is that, although a conventional discretization model can be solved exactly, an
exact solution is found to an approximate problem; whereas the model proposed here is not
sensitive to the timing of events and therefore always gives an exact representation of the
problem data with minimal discretization. However, added expense is incurred by the
introduction of sequencing variables and constraints. Preliminary comparison with existing
mixed integer formulations indicates promise for drastic reduction in model size for certain
types of problems.

Keywords: Mathematical Programming, Scheduling, Batch Process, Resource Constraints,

Time Discretization, Time Interval, Interval Analysis, Chemical Engineering.

Introduction

Over the past decade, research in the scheduling of chemical processes has evolved from the
treatment of simple, easily understandable problems which were closely related to many
classical problems in operations research (OR) to the treatment of problems of industrial
importance which are more realistic in terms of problem detail. Both the sophistication of
solution techniques being applied to these problems and the level of problem detail treated
have grown in parallel. However, the basic elements of problem formulation still draw
largely upon the ideas set forth by earlier researchers in OR.
 780

In this paper, the concepts behind a different exact modelling scheme for resource
constrained processes which is based on the notions of time interval analysis are discussed.
Not only can interval analysis be used as the underlying logic for a mathematical formulation
of the problem, but also as a means of determining which parts of the problem need not be
included in the formulation.
The next section of the paper will concentrate on previous research in the area. This will
be followed by a discussion of the structure of the specific problem being treated here. Next,
some of the ideas behind interval analysis will be briefly discussed. The details of the
formulation will then be presented. Finally, the size of the formulation will be examined
relative to the MILP size required by conventional formulations.

Background

In the area of OR, process scheduling has been studied extensively for several decades.
Scheduling research in the chemical industry can no longer be considered a new field,
although most research progress has occurred in the past ten to fIfteen years. Only relatively
recently have chemical engineers been studying problems with resource constraints. In this
section of the paper, OR contributions to scheduling will be briefly discussed. Next, some
research in chemical engineering toward problems without resource constraints will be
summarized. The focus then will shift to a more detailed discussion of specific work in the
area of resource-constrained batch chemical processes. Finally, some conclusions about the
current state of research will be given.
A large portion of the scheduling research in OR has concentrated on special cases of
various problems. The focus has been toward developing special purpose algorithms which
result in optimal solutions. Problems for which such algorithms are not found have been
formulated mathematically. There has also been a significant amount of effort directed
toward the construction of heuristics which perform well on many classes of problems. Most
heuristic work is based on job-prioritization and list scheduling, often coupled with
neighborhood search techniques. A general overview of this area is provided by Baker in [I]
and a further examination by French in [7]. Surveys of various approaches are given in
[8,9,10,14]. In [2], Blazewicz, et al, provide an extensive summary of OR approaches to
several special cases of resource-constrained problems. For problems of very simple
structure, exact algorithms are presented. For problems of a more complex nature,
mathematical formulations are provided. Finally, an important line of research in the area of
resource-constrained problems was initiated by Fox [6], who presented one of the first
approaches which attempted to reason about the problem in terms of its constraints.
 781

Much of the early work in scheduling in the area of chemical engineering focussed on
various special cases of batch processing problem structures in the absence of resource
constraints. Overviews of the various structural features present in batch processes as well as
of some of the approaches to problems of these various structures are given in [16,18,22].
Most recently, Reklaitis [17] revisits earlier reviews in order to provide not only more precise
classifications of currently treated problems, but also to outline current needs for further
research.
More recently, researchers have recognized the importance of resource constraints in
chemical processes. Both heuristic and exact modelling approaches have been presented,
however, the exact approaches have yielded the most productive results toward better
understanding of the problem. Lazaro and Puigjaner [12] allowed for the incorporation of
resource constraints in problems with unlimited intermediate storage. A more realistic case of
a general batch process was treated by Egli and Rippin [5]. An enumeration scheme was
employed in order to solve relatively small sized resource-constrained problems. More
recently, a stream of research based on a formulation proposed by Kondili, et al [11], has
resulted in solution approaches to larger, more realistic problems. The formulation is based
on the uniform discretization of time [3,13,15], and thus is limited in its capability to exactly
represent problems without the rounding of problem data. Sahinidis and Grossman [19]
recognized sub-structures of the formulation in [11] which resembled lot-sizing problems. By
disaggregating some of the continuous variables in the formulation, they were able to propose
a larger formulation which had tighter LP relaxations than the original formulation, resulting
in reductions in solution time ranging from 10% to 81 %. However, the structure of the binary
variable set was not changed. In addition to exact approaches, some research has been
focussed toward heuristic algorithms. Cott and Machietto [4] and Tongo [20] propose
schemes which resemble simulation type logic, using a "fIrst available" time slot heuristic.
The multiproduct batch plant with resource constraints was treated by Tsirukis and Reklaitis
[21] using a decomposition approach. The planning problem MINLP was solved using a
generalized Hopfield network. Subsequently, the resource allocation problem was solved by
restricting the search space with a feature extraction algorithm prior to detailed optimization.
Finally, work presented by the authors [23,24,25] studied the resource-constrained problem
from a different perspective. A framework was constructed which was based on a non-
uniform discretization of time defmed by real process events. The decisions were handled by
a search framework which produced a time interval for each task and a set of precedence
constraints. The intervaVconstraint problem was then solved using an LP formulation.
The common theme throughout the modelling of resource-constrained scheduling
problems is the uniform discretization of time. The diffIculty with such a scheme is that exact
modelling of processes is quite dependent on the timing of events. In order to accurately
 782

model a process, the greatest common divisor of all event times must be used as the width of
the discretization intervals. Clearly, when problem data is expressed in terms of real
numbers, rounding must be done in order to obtain a reasonable interval width. Furthermore,
in practice the rounding is often more severe than simply converting real numbers into
integers. For example, since an hourly discretization of a week would result in a large
formulation, often the discretization must be more coarse. This effectively limits the
scheduling of many tasks which would require an hour or less. As a result, for certain
scheduling problems, a formulation which is not sensitive to the timing of events may prove
advantageous.

Problem Definition

The formulation presented in this paper addresses the modelling of resource constrained batch
process scheduling problems which are quite general in structure. However, since the level of
the scheduling problem treated here is that of the detailed timing of batches, some restrictive
assumptions are made. It is assumed that input information will be known about the structure
of the processing equipment network, resource availabilities, product recipes, assignments of
units to tasks, and production requirements. Other aspects of a more general problem will be
treated in future work on this formulation. The discussion in this section will address plant
structure, recipe structure, and production requirements.

Plant Structure

The proposed plant structure allows for the most general form of equipment connectivity.
Equipment items may be connected to each other or to storage vessels in any fashion. In
addition, the dynamic nature of connections can be modelled through the use of resources,
that is, in order for a transfer to occur, the units must be connectable and a resource which
represents a transfer line must be reserved. The connectivity of the various resources to the
processing equipment can also be specified.

Recipe Structure

The recipe structure will be discussed in terms of three issues. First, the basics of product
recipes will be defmed. Then, the allowed storage modes will be covered. Finally, the
interaction between resources and product recipes will be outlined.
 783

The recipe structure considered encompasses all of the customary attributes of chemical
product recipes. Specifically, the following attributes of each recipe step - processing unit
pair are accommodated: sequence dependent setup times, transfer times both into and out of
the unit from and to other specific units, processing times, and clean out times.
However, this treatment goes beyond the conventional level of detail in that each recipe
task can be subdivided into multiple sub-tasks, called steps. For example, a recipe task may
have two distinct phases, heating and stabilizing, which could have vastly different resource
requirements and rate expressions. In general terms, the decomposition is fIrst into recipe
tasks, each of which is assigned to one processing unit and which are related by precedence
constraints. Each recipe task can be further divided into operations. The operations may
include set-up, transfer-in, processing, transfer-out, and clean-out Each operation can more
fmely be composed of individual recipe steps, as in the case of the multiple processing steps
for a recipe task which was mentioned above. Depending on the processing unit chosen for a
task, the structure of the steps necessary for the task may be different.
Recipe steps can have several attributes. Each recipe step can have its own resource
requirements, which may be dependent on which processing unit is chosen for the recipe task.
Each recipe step can have a fmite allowed wait time which may follow the step. Such wait
times can be used to represent product instabilities. It may also be specified that during the
processing or all steps of a given task, an aggregate amount of idle time may be incurred.
This is an amount of idle time that may be distributed among the individual steps such that the
total idle time incurred during the processing of the steps does not exceed the aggregate
allowed idle time. The fIrst step of the recipe of a product may specify several necessary
precursor products, thus representing the combination of intermediate materials during
processing.
The handling of storage in this approach is more comprehensive than most of the past
treatments, 3.Ithough it still is not completely general. The aggregation of batches in storage,
for example, where three batches of G enter the storage vessel and exit as two larger batches
of G, is not detennined, rather it must be specified by the user as problem input. However,
several aspects of storage are modelled here that have not been extensively studied in past
approaches.
The fIrst issue deals with the fed batch mode. If material in an upstream unit is ready to
transfer out of -the unit and into the next stage, which is operated in fed batch mode, two
scenarios are possible. The slow (fed batch) transfer can take place directly from the unit,
occupying the unit until the fed batch phase of the downstream recipe step is complete.
Another option is that the material can be quickly transferred to storage, clearing the upstream
processing unit, and allowing the batch fed operation to proceed from the storage tank. Such
options can be described in the framework presented here.
 784

The second feature is the allowance for [mite wait times for a recipe task. The current
accepted description language for specifying storage is insufficient for representing batches
that are only stable for a limited amount of time. A decomposition of the storage availability
and the allowed storage time into separate vectors is proposed as an expansion of the present
terminology. Specifically, these two vectors can be "multiplied" in order to defme complete
storage policies as shown below:

[~ ] x [NIS FIS MIS UIS]

-[ zw
FW*NIS
UW*NIS
zw
FW*FIS
UW*FIS
zw
FW*MIS
UW*MIS
FWz:rrrS
UW*UIS
]

The letters ZW, FW, and UW represent the amount of time a product may stay in storage:
Zero Wait, Finite Wait, or Unlimited Wait, respectively. Of the three, FW has not been
extensively treated in previous approaches. The letters in the other vector represent storage
conditions. NIS means that while no storage vessels may be used, the batch may remain in
the processing unit for storage. FIS indicates that the batch must leave the processing unit but
may be stored in one of a [mite number of storage vessels. MIS indicates that both NIS and
FIS apply. Finally, UIS implies that there is an unlimited number of storage vessels. The
meaning of the product of a storage time and availability attribute should be self evident:
FW*NIS means that the product may be stored in the processing unit for only a finite amount
of time, and so forth. For obvious reasons, any availability multiplied by ZW reduces to ZW.
The above types of storage policies are important when dealing with sensitive intermediate
products with limited stability, for example, biochemical intermediates.
Also related to storage is the use of idle time, which may be incurred in a processing unit
prior to the initiation of processing, between the steps associated with a processing operation,
and following the completion of processing. This allows a process to be interrupted between
phases when different resources are required but are not yet available as well as to be delayed
prior to processing and/or held after processing.
The fmal aspect of the recipe description outlines how resources are to be treated.
Resources may be of two types, renewable (RR) or non-renewable (NRR). The renewable
resource is one which is replenished after use, for example, manpower, a transfer pipe, or
electricity. Every level change of a renewable resource is treated as an absolute change and is
not cumulative. Non-renewable resources are those that are consumed, for example, raw
materials. Each specified level change of a non-renewable resource has a cumulative effect.
That is, if there was I ton of raw material A and the level is changed by the arrival of 3 more
tons, the level of raw material A would be 4 tons, whereas in the renewable case, the new
 785

level would have been 3 tons. The availability (or arrival) patterns of resources can be
specified to be following a cyclic pattern or an acyclic pattern, as in weekly manpower levels
or raw material levels, respectively.
Each recipe step may require several resources which serve different roles. As previously
discussed, each task may be split into several steps, which allows for the specification of
different resource demands for different phases of the task. For example, transfer into a unit
may be divided into initial rapid charging of the unit followed by a slower fed batch transfer,
each requiring different levels of electric power.
In the above discussion, it was shown that the recipe structure defmed in this approach
encompasses not only the consideration of the precedence of recipe steps, but also
consideration of a variety of storage policies involved for each recipe task and treatment of
detailed resource demands for each recipe step. The recipe model is sufficiently general to
represent the vast majority of product recipes.

Production Requests

Each required product may have a due date and a release time. The release time of a product
is the earliest point in time at which production of the product may be initiated. While both
due dates and release times seem to impose the same type of constraint at opposite ends of a
schedule, they are distinct Specifically, a due date may not be met but a satisfactory
schedule may still be obtained, whereas a release time is a hard constraint and in most cases
can not be violated.
A restrictive assumption made in the current model is that the batch sizes will be known
in advance, since this decision is somewhat higher in abstraction than the notion of detailed
scheduling. This assumption implies, then, that batch aggregation will not be determined by
the model presented here since the logic for determining when to aggregate and split batches
would be nearly the same as that required in order to size batches at the onset. However, the
assumption does not restrict the specification of aggregation and splitting by the user in the
description of the product recipe through the renaming of products after aggregations and
splits. This assumption also implies that the user must specify the number of batches to be
produced in addition to their sizes in order to describe the overall production scheme. The
amount of a product specified in a product order is assumed to be one batch.
A fmal ramification of assuming that batch sizes are fixed is that task durations and
resource requirements are fixed as well. These stipulations are fairly restrictive, but one
should note that using a constant level of resource requirement by a task is at best an
approximation anyway. As a task is processed, the level of resource required can vary
significantly over the duration of the task due to external influences and controller actions.
 786

Summary of Problem Structure

The proposed scheduling problem includes a number of scheduling features that have not
been treated before, many of which are specific to chemical processes and thus typically are
not regarded in operations research studies. To recapitulate, the new ideas are the treatment
of complex recipes including mixing of intennediate products, an allowance for splitting
major tasks into many steps, the intelligent accommodation of storage with fed-batch
processing mode, a new defmition of storage modes allowin.g for finite stability times of
intennediates, and a complete treatment of resources. The salient features of the problem can
thus be outlined as follows.
General equipment connectivity
Division of recipes into tasks
Division of tasks into operations
Division of operations into steps
Finite allowed idle times
Aggregate idle times
Merging of intennediates
Complex storage policies
Non-renewable and renewable resource constraints
Fixed batch sizes
Fixed unit and resource assignments
Product release times
Product due dates
Many of these individual features have seen little or no attention in the literature. Some of the
assumptions with regard. to fixed data in the problem description are restrictive, however,
these aspects of the problem can be modelled by the addition of more variables and
constraints to the fonnulation proposed here.

Definitions

Before develo~ing the fonnulation, it is necessary to introduce some additional concepts.

Specifically, the concepts of resource confluence intervals (RCI's), RCI elimination criteria,
and event free intervals (EFI's) will be used in order to pre-analyze the problem so that only
pertinent parts of the problem are fonnulated.
A resource confluence interval is an interval of time for a step such that all resources
required for that step are present in sufficient amounts so that the step can be processed. In
addition, an RCI spans as wide a time interval as possible. This means that, for a given step,
 787

40- ----------------------------~----____r----------------------------------------

35 ---------------------------- ----------------------------------------

3(}- -----~----.--------- ---------------------------------------- ----------------------------------------

~ 25 ------------------------------------------------------------------
e
~. 2(}-f-- ------------ - -----------------------------------------'---------,---------------
~
15 ----------------------------------------------------------------------------------------------- ---------------

1() ----------------------------------------------------------------------------------------------- ---------------

5- --------------------------------------------------------------------------------------------------------------

Q.
o 5 10 15 20 25 30 35
lime

Figure 1 - Resource ProfIle for Stearn

two ReI's never will exist end-to-end_ For example, consider the resource profiles in Figures
I and 2. If a step required 25 kW of electricity and 25 kBTU of steam, ReI's for that step
would exist over the intervals [2,6], [10.5,17.5], and [20,22]. During all other times, at least
one of the two resources is present in insufficient quantity. Also, note that none of the above
intervals overlaps with any of the others, or touches end-to-end. This is because the intervals
are created over as large a time interval as possible. Intervals which are shorter than the
duration of the step are not considered valid ReI's and receive no further consideration.
Further details of the construction of ReI's are developed in more detail in [23].
After preliminary analysis, there can be several ReI's for each step. Based on the set of
ReI's and on pair-wise consideration of steps, some ReI's may be shortened or altogether
eliminated. For a given pair of steps i and j, where step i precedes step j, the duration
parameters Dij,min and Dij,max can be computed from recipe data. The durations of all steps
which must occur between the start times of steps i and j indicate a minimum amount of time
which must elapse between the start times of the two steps, D ij.min- The durations of the steps
between i and j coupled with their allowed idle times defme the maximum time which may
 788

4C

35 - - - - - - - . .------,-------------------------

3.G -- --------------------- ------- --------------------- ------- ---------------------- ------------------------

~ 25 -- --------------------- ------- --------------------- ------- ---------------------- ------------------------

e.
€ 2(} -- --------------------- ------- -----------------:--- ------- ---------------------- ------------------------
-g.
'1::

iIi 15-~ --------------------- - --------------------- ' - ----------------------'--- ---------------

10- ----------------------------------------------------------------------------------------------- ---------------

5 ----------------------------------------------------------------------------------------------- ---------------

o 5 10 15 20 25 30 35
Time

Figure 2 - Resource ProfJIe for Electricity

elapse between the start times of the two steps, Dij.max' These two parameters are used in
Figure 3, which illustrates one criterion for eliminating an RCI. Consider the pair of steps i
and j where i precedes j. Further, examine one fIxed RCI for i, RCI i , against several RCl's for
j, as shown in Figure 3_ The point labelled A on the axis is the earliest time step j could end if
RCIi was used for step i. Similarly, the point labelled B is the latest time step j could start if
RCIi was used for step i. If there is no RCI for j with end time RCIj,ET ~ A and start time
RClj.sT ~ B, then no portion of RCI i is useable, since its use will preclude a feasible RCI
assignment for j. Several other such criteria for shortening and eliminating RCI's exist and
are detailed in [23]_
From the discussion of RCl's, it is apparent that several changes in resource level can take
place during an RCI, as long as the resource level is not exceeded by the demands of its step
at any point during the RCI. This implies that the discretization of time must be more
detailed for the summation of resource consumption_ For this purpose, the notion of an event
free interval (EPI) is introduced_ An EPI is an interval of time for a step (during one of its
 789

Time A B

Figure 3 - An RCI Elimination Criterion

RCl's) during which no resource required by that step changes leveL Depending on resource
requirements, the set of EFI's may be different for the various steps in the problem.
The tools for resource interval creation and processing presented in this section serve as
valuable pre-analysis tools prior to the generation of a detailed formulation. Through the
generation of RCl's and subsequent elimination and shortening, the overall set of EFl's which
must be considered in the final formulation can be considerably reduced.

Mathematical Formulation

The discussion will now be focussed on the set of mathematical formulae used for the
representation of the non-uniformly discretized formulation. First, the representation of time
and the constraints associated with it will be presented. Then, the concepts and constraints
involved with the sequencing variables will be discussed. Next, the manner in which
renewable resource constraints are handled will be given. These three aspects of the
formulation are those which allow it to be different in structure than past approaches.
Following the discussion of these three key aspects, the rest of the formulation, which largely
consists of constraints for the incorporation of additional problem features such as linkage
with a continuous formulation, optional use of storage vessels, non-renewable resource
constraints, and sequence dependent set-up times, will be developed.

The Formulation of Time

The mathematical description of the time domain for resource-constrained scheduling

problems, which is the simplest concept to represent using conventional approaches, typically
 790

is the root of the intractability in problem solution. A large number of decision variables is
usually necessary in order to divide the scheduling horizon into small, uniform increments.
Although time discretization will be used in this work, that discretization will be based upon a
non-uniform set of time intervals, which allows a significant reduction in the number of
variables. In the interest of following convention, the variable Xij is introduced where Xij = 1
if an interval is assigned to step j which encompasses the ith event free interval (EFI) for step
j.
The number of variables introduced by such a defInition is not as obvious as it might at
fIrst seem. Depending on recipe structure, the discretization for each step may not be the
same as that for every other step. For example, if there is a transfer step whiCh uses
electricity and a pre-heating step which uses steam and if the sets of EFl's defmed by steam
and electricity are different, the discretization for the transfer and pre-heating steps would be
different. If, however, the pre-heating step uses both steam and electricity, the discretization
for the two steps must be identical (taken from the combination of the two resource profIles)
in order to sum up resource usage. Even though the transfer step does not use steam, it
inherits the discretization from all steps which share electricity since it must, at some point,
be considered in conjunction with steam utilizing steps. In addition, application of interval
narrowing and elimination criteria may make some Xij variables unnecessary.
Example: Consider a case where two steps, Al and Bl, which use the resources steam
and electricity from Figures 1 and 2, are to be processed. Step B 1 uses steam, and step Al
uses both steam and electricity. Considered by itself, step Bl would result in discretization
points at times 0, 2, 6, 9, 22, and 30. However, since both Al and B 1 use steam and since Al
also uses electricity, both Al and Bl will inherit a discretization based on both resources. All
points of discretization will then be 0,1,2,6,8,9,10.5,17.5,20,22,27, and 30.
The qualitative description above will now be cast into a more precise form. Let 5 be a
set of steps which share at least one resource with step j and which are not otherwise
constrained with respect to step j because of recipe considerations. Let 1( be the set of all
resources required by every step E 5. Every step", 5, which uses at least one resource in 1(
and which is not constrained with respect to j by the recipe, is added to the set 5. The
resources required by the new members of 5 are then used in updating 1(, The defmition of 5
continues in a recursive fashion until the set can no longer be expanded. Every step j E 5
derives its discretization from the resources in 1(,
In the worst case, all steps will consider every resource in the time discretization. Let s be
the number of steps, r the number of resources, and e the average number of events per
resource. Then, the number of Xij variables will be s(er- 1). Through pre-analysis, some of
these variables can be eliminated. For example, if there is a release time specified for a
product, all Xij variables where the ith EFl occurs prior to the release time are not considered.
 791

One can further envision several elimination criteria based on the values of the parameters
Dij.min and Dij,maX'
In many cases, an interval assigned to a step may span several EFI's within a given
resource confluence interval (RCI). Only certain pairs of EFI's can be used as the end points
of an interval assigned to a step. As discussed earlier, each RCI for a step can be composed
of several sequential EFI's. Each pair of EFI's allowed to serve as end points must satisfy two
criteria. First, the time spanned between the start time of the fIrst EFI and the end time of the
last EFI must be greater than the duration of the step. It is also important that the time
spanned by the EFI pair is not larger than necessary. The allowed pairs of EFI's for step j can
be determined by sliding a time interval which represents the duration of step j along the time
horizon~ Initially, the start of the interval will coincide with the start of the fIrSt EFI for step j.

One allowed pair is defmed which starts with the fIrSt EFI and ends with the EFI in which the
duration interval ends. Then, the interval is moved forward in time. Each time either the end
or start of the interval crosses the boundary between two Xij'S, a new pair of allowed EFI's is
discovered. In order to estimate the worst case number of these pairs, recall that there are
(er- 1) Xij variables (intervals) for each step. This means that the start and end of the duration
interval can each cross at most (er- 1) interval ,boundaries. There are also the initial position
and fmal position of the duration interval, which results in a total possible of
.s(2(er- 1) + 2) - 2ser pairs in the worst case. Many pairs will never be necessary because the
step duration is wider than the duration of the single EFI's. Further, considerations of Dij,min
and Dij.max can be used, as discussed earlier, to restrict the number of allowed EFI pairs by
using interval elimination ideas.
It is important to be sure that all Xij'S within the assigned interval have a value of 1 (an
assigned interval for a step must be continuous). When the two EFI's, EFI. and EFI h, are
chosen as the end points for the interval, for all i where a:5: i ;5; h, then Xij - 1. The constraint
below enforces this condition.

h
h - a + 1 - (h - a + 1)[(1 - Xh) + (1 - X.)] ;5; ~ Xij (1)
I-a

Note that this constraint need not be created when a - h or when a = h - 1. When a given pair
of X variables is chosen for the end-points of the interval for a step, not only must all of the
X's between the end points be set to 1, but the X's outside of the chosen interval must be set to
O. This includes all X's in the RCI in which the interval is chosen and all X's outside of the
chosen RCI. The following constraint accomplishes this.

,·1 z
(a + z - 1 - h)[(1 - '4,) + (1 - x.) + Xa• l •j + Xh+ l •j] ~ I Xij + I Xij (2)
i-I i-h+1
 792

Since one instance of constraints 1 and 2 must be created for every allowed EFI pair for a
step, the worst case estimate for the number of each of these constraints is 2ser.
It is also necessary to ensure that each step is assigned a time slot. For this purpose, the
following constraint is created for each step.

(3)

When choices for steps are present, as in the case of multiple options for the use of storage or
sequence dependent set-up steps, constraint type 3 will be modified to include X's for all
options, The size of constraint set 3 is s.
It is also necessary to ensure that enough· consecutive EFI's are reserved in order to
accommodate the duration of each step. Allow TBij and TEmj to represent the start and end
times of the ith and mth EFI, respectively. For all Xij and X mj of every step j where
TEmj - TBij < D j, iSm, the following constraint is created.

Xi.I,j + ~+Ij;:: i - m + I Xaj (4)

a-I

Essentially, when all EFI's between i and m are being used, the constraint forces an additional
EFI to be used. Otherwise, if only part or none of the EFI's between i and m are used, the
constraint has no effect. In the worst case, all EFI's will exist within the same ReI for each
step. For a single step, then, the worst case number of constraints from type 4 is the
summation from 1 to (er- 1). For all steps, the number of constraints is ser(er- 1)12.
However, if this is the case, note that the number of constraints of types 1 and 2 will be much
less than the worst case since there would have been only one allowed pair variable in the
worst case for type 4.
Finally, since the notion of allowed pairs is being used for the definition of constraint
types 1 and 2, it is necessary make sure that no additional start/end pairs of EFI's are chosen
which span wider time intervals than the allowed pairs. In other words, it is either necessary
to enforce constraint types 1 and 2 for all EFI pairs or create a constraint which ensures that
only allowed pairs of EFI's will be chosen as interval end-points. For every allowed pair of
EFI's h and m such that there is no allowed pair {i,l:(i<:h A km) v (i>h A ISm)}, the following
constraint states that if h and m are the chosen allowed pair, no EFI's outside of the time
spanned by h and m will be chosen.
 793

Xh-,j + Xrn+lj::; 2[2 - (Xhj + Xrn)J (5)

The size of constraint set 5 is equal to the number of allowed EFI pairS in the worst case,
which results in a possible 2serconstraints.

Sequencing Variables and Constraints

In order to ensure that resource levels will not be violated, it is necessary to know whether or
not two steps are sequenced with respect to each other. Further, it is important to propagate
the effects of sequencing two steps to the rest of the steps in the problem. For these purposes,
several constraint types are needed.
First, the variables Sij and Sji are introduced for every pair of steps i and j which use at
least one common resource and which are not already constrained because of recipe
considerations. If step x follows step y, then Syx - 1, Syx - 0 otherwise. It follows that:

(6)

which states that the two steps can only be constrained in at most one direction with respect to
each other. More strongly, if steps i and j share a processing unit:

(7)

In the worst case, every step could be constrained with respect to every other step except
itself, resulting in s(s-l) variables and s(5-1)/2 constraints.
Constraints are also necessary in order to propagate the effects of a sequence decision.
For example, suppose that two steps, j and I, are constrained such that j precedes I (S)l = 1).
The implication is that all steps which precede j also precede I (Sxl = 1 if Sxj = 1). Similarly,
but in the opposite direction, Sjx = 1 if six = l. In order to model this situation, for every
unique triplet of steps x, j, and I (note that {1,2,3} and {1,3,2} are unique triplets) which share
some common resource, the following non-linear constraint represents the propagation of
constraints:

(8)

Using a simple linearization construction, the above constraint translates into the following
pair of linear constraints:

(9)
(10)
 794

In the worst case, every step will have the potential to overlap every other step, though in
practicality this would never happen. There are .s(s-I)(s-2) unique triplets, giving a total of
2.s(s-I)(s-2) linear constraints in the worst scenario.
Finally, additional constraints arise from recipe concerns. Specifically, no task can be
pre-empted on its processing unit in preference to the completion of another task on the same
processing unit. Recalling that a task is constructed of individual steps, this implies
additional· constraints among the steps of tasks assigned to the same processing unit. Let T A
and TB be tasks assigned to the same processing unit. Further, allow aET A and bETB to
represent steps which belong to tasks A and B, respectively. The symbol ITxl is used to
indicate the number of steps involved with task X. With these definitions, constraints can be
formulated which state that if at least one step of task A is related to one step of task B, all
other steps of both tasks must be similarly related. For every sba where bET Band aET A:

Sba 2: (lilTAliTBD I ISba (11)

aeTA beTB

In order to get an estimate on the number of constraints of type 11 necessary in the worst
case, let t represent the number of tasks and assume all tasks are assigned to the same
processing unit. In such a case, there are t (t - I) possible unique task pairs. With each pair
of tasks, there are potentially (sf t)2 step sequencing variables, sba' resulting in a total of
s2(t- I)/tpossible constraints. Again, it should be noted that this would rarely be the case in a
real process environment.
Of the three types of step sequencing constraints proposed, the first seems obvious and is
conventional, while the second and third are more specialized to this approach and work in
conjunction with more of the constraints presented in subsequent sections. It was also shown
that if one is not careful in deciding which pairs of steps do not need sequencing variables
during the formulation of a problem, much extra information can contaminate what might be
an otherwise concise formulation.

Renewable Resource Constraints

A major cause· of complexity in mathematical formulations of scheduling problems is the

consideration of shared resources. Conventionally, with the fine, uniform discretization of
time, the levels of resource usage during each interval are summed and compared to the
available level of resource. A similar approach applies to the non-uniform grid presented
here; however, it is important to remember that in this formulation, constraints apply only to
the potential for various steps to share resources over an interval, not the absolute fact of
resource sharing.
 795

In order to represent the potential for steps to overlap, an expression can be written in
tenns of the variables presented so far. If there are n steps which share some resource and
none of which have any recipe-derived constraints with respect to each other, they have the
potential to overlap during a given EFIe if:

II Xci II (1 - (Sij + sji» = 1 (12)

i-l j-i+l

Equation 12 states that steps i through n can overlap if they are assigned during the given EFI
and if none of the steps have any constraints with respect to any of the other steps. In order to
constrain the actual problem, it would be easy to enumerate all possible overlapping sets for
each EFI and state which ones are allowed and which are not; however, this would produce an
undesirable number of constraints.
Further examination of the problem suggests a logical point for constraint building based
upon the concept of maximal overlapping sets. A maximal overlapping set for an EFl is a set
of steps which all could overlap in time and not cause any resource level violation such that
there exists a step that can be added to the set which would cause a resource level violation.
Specifically, since a maximal overlapping set will cause no violation, no constraint is
necessary for any maximal or sub-maximal overlapping set. An algorithm for the generation
of maximal overlapping sets is given in [23]. A constraint is necessary, however, for each
combination of a maximal overlapping set and a step which is not a member of the set and
which would cause a resource level violation. These constraints will represent the boundaries
of allowable resource allocation, making additional constraints unnecessary. The fonnat of
the constraints will be that of equation 12, except that, since it is desired that the set of steps
not overlap, the right hand side of the constraint will be 0, not 1. This fonn can subsequently
be linearized.
Example: Consider a case with 4 steps (1, 2, 3, and 4) which share some resource and
can exist during EF1e. If the level of resource during EFIe is 2 and the demands of steps 1,2,
3, and 4 are 0.5,0.5,0.75 and 1.0 respectively, then the maximal overlapping sets are {1,2,3},
{3,4}, and {1,2,4}. The set of constraints that arises is then:

X 1e (1-(SJ2+S21»( 1-(s!3+s31»(I-(sI4+s41»X2e
(H ~3+S32) )(1-( ~4+S42) )X3e(1-(S34+S43) )X4e = 0 (13)
Xle(1-(S!3+s31»(I-(sI4+s41»X3e(1-(s34+s43»X4e = 0 (14)
X2e(l-( S23+S32»( 1-(S24+S42))X3e(1-(S34+S43) )X4e = 0 (15)

Constraint 13 arises from the flrst maximal overlapping set combined with step 4. Constraints
14 and 15 arise from the second set combined with steps 1 and 2. The third maximal
 796

overlapping set combined with step I would yield a constraint identical to 13. The set of
constraints is complete in that it disallows all resource violating combinations but does not
eliminate any valid overlapping combinations. The set above can be transfonned into the
following linear set:

Xle+(l-(SI2+S21»+(l-(SI3+S:31»+(I-(SI4+S41»+X2e
+(I-(s23+s32»+(l-(~4+s42»+X3e+(I-(s34+s43»+X4e S 9 (16)
Xle+(1-(S13+S31»+(1-(SI4+S41»+X3e+(1-(S:34+S43»+X4e S 5 (17)
X2e+(I-(~+s3~)+(1-(~+s42»+X3e+(I-(S:34+s43»+~ S 5 (18)

The number on the right hand side of constraints 16-18 represents an integer 1 less than the
number of 0-1 tenns on the left hand side. Effectively, the constraints indicate that all of the
events associated with each of the left hand side tenns cannot simultaneously take place.
Constraints of the type 16-18 can be expressed in a general fonn. First, let an augmented
maximal overlapping set (AMOS) be fonned by the addition of one step, which is not a
member of a maximal overlapping set (MOS) and which causes a resource level violation, to
the MOS. Therefore, from each MOS, several AMOS's can arise. For each unique AMOS,
the following constraint is written:

L (X.i +
ieAMOS
L
(1
jeAMOS
- (Sij + s.0» SIAMOSI- 1 (19)
pi

The term IAMOSI indicates the number of 0-1 terms on the left hand side of the inequality.
Within one EFI, the worst case would be when, if there are n total steps which could overlap,
nl2 steps belong to each MOS. This would give rise to n!/«nl2)!)2 MOS's. Let the number of
MOS's then be K. To form an AMOS, there are nl2 steps possible to add to each MOS.
However, since each AMOS would be produced twice, this number is divided by two,
resulting in Knl4 constraints (AMOS's).
The number of resource violation constraints introduced by constraint type 19 is seen to
be quite problem dependent. Although the constraint was originally a highly non-linear
decision constraint, exploiting the structure of the product of tenns and the physical meaning
of the constraint results in a rather simple linear f9nn of inequality. No additional variables
were necessary in order to account for resource usage. However, some pre-analysis of the
type alluded to concerning the generation of MOS's is necessary for the construction of a
minimal set of resource usage constraints.
 797

Linkage with the Continuous Formulation

The discussion of the time horizon has been centered on a wide, non-uniform discretization of
time. In contrast to the conventional form of discretization, the selection of an interval
(X ij = 1) in this formulation does not necessarily dictate a start time for a step. Instead, the
selection of an interval indicates that the start time of a step may take place somewhere within
the interval. The variable Bj and the parameter Dj respectively represent the start time and
duration of step j. A continuous formulation which relates the various Bj's is discussed in
[25]. The formulation includes straight-forward constraints on idle times and step durations
which are based on recipe concerns. Also included were constraints which dealt with the use
of storage vessels in order to ensure that a series of fill and empty steps did not cause the
vessel to run dry. Additional constraints which can be used to link the continuous
formulation in [25] to the discrete portion of the formulation presented here.
The relationship between the continuous variables and the binary interval variables is
given by the following two constraints.

B j ~ (Xij - X i .l.j)TB ij (20)

Bj + Dj ~ (Xij - Xi+1.)TE ij + H(l - (Xij - X i +1)) (21)

The first constraint forces the start time of a step to be greater than the start time of the first
chosen interval, TB ij. The second constraint forces the start time of a step to be less than the
end time of the last chosen interval, TEij> minus the duration of the step. The symbol H
represents the time horizon. The size of this constraint set will be the same as the number of
X variables, s(er- 1).
More constraints are needed for every possible sequencing decision. Specifically, for
every Sij variable, the following constraint is important:

(22)

Constraint 22 indicates that if two steps are sequenced, a space of at least the step duration
must be left between the start times. These sequence constraints are in addition to those
implied by product recipes, which would already be included in the balance of the continuous
formulation.
The constraints presented in this section are intended only to link the discrete formulation
presented here to the continuous formulation for exact time assignment. All of the constraint
types presented in [25] apply to this expanded version of the continuous formulation in order
to form a complete mathematical formulation.
 798

Optional Use of Storage Ve/isels

Several constraints are necessary in order to model the optional use of storage. The fIrst
constraint is a modilled form of constraint 3, given below:

~I Xi.Tl(j-+k) + .L~Xi.Tl(j-+S)
5 I
~1 (23)

The arrow notation indicates that transfer is from a task to its successor task, k, or to a storage
vessel, s. The T1 indicates litat it is only the fIrst transfer step of a transfer operation. The
constraint indicates that at I~ast one mode of transfer between tasks must be chosen, using
storage. or a direct transfer. The flJ'St summation covers the direct transfer while the second,
double summation covers all possible storage units. The number of such constraints in the
worst case is t, where t is the number of tasks in the problem.
Constraints of the type of constraint 1 remain applicable. However, constraint type 2
must be modilled in order to exclude all other storage vessels once a mode of transfer has
been chosen. For each allowed pair of start and end EFl's Xi.TI(j-+X) and Xm,Tl(j-+x), where x
represents k or s (regular transfer or storage), for every task which can use storage, the
following constraint is necessary.

i-I z z
.L X.,TI(j-+x) + .L x.,TI(j-+x) + .L .L X ••TI(j-+y) :s;
a-I a-m+ I )"'X .-1

(24)

The symbol NT represents the number of terms on the left hand side of the constraint. The
size of this constraint set is dependent on the number of transfer options (allowed storage
vessels) and on the number of resource events, giving 2tyerconstraints, where yrepresents the
average number of storage choices per task.
Once a transfer mode is selected by choosing an interval for the flJ'St transfer step of that
transfer mode, the rest of the steps associated with that transfer operation (additional transfers
and storage vessel clean-outs, indicated by (T,C)x) must be selected as well. If a transfer
mode is not selected, then none of the subsequent steps of the operation should be selected.
The following two constraints ensure these two properties.

(25)

.L
(f,C)x
Ii Xi,T1 :s; NT I
i
Xi,TI (26)
 799

An instance of constraint 25 will be created for every (T,C)x for every mode of transfer,
giving a total of tzy constraints, where z represents the average number of transfer and c1ean-
out steps associated with the use of storage. An instance of constraint 26 will be created for
each mode of transfer between two tasks, giving a total worst case number of constraints of
ty.
The combination of the above constraints effectively handles the use of optional storage
vessels. Storage of material within a processing unit is handled by the idle time constraints of
the continuous linear program presented in [25].

Non-Renewable Resource Constraints

It is necessary to introduce an extra binary variable for handling non-renewable resources.

The variable Zij is used to represent the starting point of an interval for step j. It is only
necessary to introduce the Z variables when a step uses a non-renewable resource. In the
worst case, every step would use non-renewable resources, giving s(er- 1) possible variables.
In practice, a small percentage of the total number of steps will actually require non-
renewable resources.
1f Xij = 1 and X i _1j = 0 then Zij = 1, otherwise Zij = O. In order to enforce this rule, the
following two constraints are necessary for each Zij'

Zij ~ Xij - Xi.1,j (27)

Zij::; (1 - X i_1j + X;yl2 (28)

In the worst case, there would be s( er - I) instances of each constraint. In practice, this will
rarely be the case.
Over the scheduling horizon, non-renewable resource arrivals defme events in time. Each
non-renewable resource may have its own event calendar, where an event in non-renewable
resource N will be referred to by eN' For each non-renewable resource N, a constraint is
created for every event eN of the form:

(29)
susesN TBa<eN

The tenn RU&N refers to the required level of resource N by step s. The tenn RLe.o,N refers to
the level of resource N just prior to the event eN' The constraint accounts for a cumulative
sum of all tentative reservations for resource N prior to the event eN' The size of this
constraint set in the worst case is er. As mentioned earlier, this will rarely be the case.
 800

Sequence-Dependent Set-Up Times

In order to consider the effect of sequence dependent set-up steps on each processing unit, an
extra set of constraints in place of constraint type 3 is necessary. This is due to the fact that Sij
is used not only to represent a direct precedence relationship here, but indirect relations as
well. Also, since, in some cases, there may be no set-up required because of production
sequence,it is not as simple as stating that at least one X = I for some set-up operation.
For each task, there are several potential set-up operations, the choice of which depends
on the task that immediately preceded the task on the processing unit. Each of these potential
operations for the set-up of the unit has an initial set-up step, SUI, and may have subsequent
set-up steps, SUx. Let X..SU1(G-+H) and Xt,,sUl(G-+H) represent an allowed pair of start and end
EFI's for a string of EFI's from EFI. to EFIb for the fIrst set-up step required if task G
immediately precedes task H (where G and H are not necessarily of the same batch) on the
processing unit. Further, allow the subscript X-PI to represent the fIrst processing step of
task X. Then for every potential initial sequence dependent set-up step on a unit U, the
following pair of constraints are written (where T represents the number of tasks assigned to
U):

I+ r
X#.x,,1
SJ-Pl.x.Pl - SI.PI.x-PI :5 r
a-I
X•.SU1(l-+l) + (1 - SI.PI.l.Pl)(T - 1) (30)

I + (lI1XI) r
X#.x,,1
SJ.PI.x.PI - SI-Pl.x.PI ~ X•.SU1(l-+J) (31)

The first constraint ensures that the proper set-up operation will be initiated when task I
immediately precedes task I on the same processing unit. The second constraint states that if
task I precedes task J, but not immediately, the sequence dependent set-up operation between
I and I is not chosen. When task I precedes task I, the constraints are switched on, otherwise,
they contribute no effect on determining set-up decisions. The summation subscript X is used
to indicate all tasks except I and I assigned to the same unit to which tasks I and I are
assigned. If task I does precede task I, the term l: SJ.PI.x.Pl - Sl-Pl.x-PI will take on a value of 0
if no tasks are between I and I, otherwise, it will be negative. The term IXI indicates the
number of tasks assigned to the processing unit, excluding I and I. If not every pair of steps
requires a set-up, the value of IXI must be suitably altered in order to match the number of
terms in the summation. Constraints 30 and 31 will then indicate that an interval for the
appropriate set-up must be chosen when I immediately precedes I, otherwise no interval
should be chosen for the set-up. In the worst case, every task is assigned to the same unit.
The result would be t(t- 1) instances of constraint 30. Also in the worst case, there will be
 801

er- 1 Xij variables for each step. This, multiplied by the possible number of sequence-
dependent set -ups gives t( t - 1)( er - 1) instances of constraint 31.
It is also necessary to restrict the binary variables when task J precedes task I, as in the
following constraint which disallows sequence dependent set-ups when the precedence is in
reverse order.

I X ••SUl(l~J) ::; nSI.Pl).PI

a-I
(32)

There are as many constraints of type 32 as there are possible worst case sequences, t(t- 1).
When the initial set-up step for a sequence dependent operation is selected, the rest of the
steps in the operation must be carried through. Similarly, if the initial step of an operation is
not selected, none of the subsequent steps should be selected. The following two constraints
enforce these rules.

n m

I X..SUl (I~J) ::; n a-I

a-I
I X •.SUx(I....J) (33)

q n m

I I X~SUx(l~J) ::; qn i-I

x-I i-I
I Xi.SUl(I~J) (34)

The fIrst constraint is created for each set-up step, SUx, where x> 1. In the worst case, this
will give qt( t - 1) constraints, where q is the average number of set-up steps in a set-up
operation. The second constraint is created for each possible sequence, resulting in t( t - 1)
constraints.
As with other formulations, the consideration of sequence dependent operations in this
formulation requires additional constraints. However, each of the constraints is strictly an
integer constraint, and the number of such constraints is polynomial in size.

Evaluation of the Formulation

Although the dimensions of the formulation presented here have been discussed in terms of
various problem attributes, it is helpful to view the relative size of an instance of the
formulation in contrast conventional formulations. Specifically, our formulation will be
compared to that of Kondili, et al [11]. First, a small example problem will be discussed.
Then there will be brief comparisons of the sizes of the formulation of [11], the formulation
of [11] with intelligent interval analysis, and the formulation presented here.
 802

Table 1 - Example Problem Data

Electricity Steam
Task Step Duration reguired CkW) required CkBTUl
AI Process 5.0 25.0 10.0
A2 Process 4.0 12.5 10.0
CI Process 10.0 7.5 25.0
C2 Process 6.0 20.0 7.5

The example to be considered involves four processing steps for the production of two
products. Each of the steps requires various amounts of the resources steam and electricity
from Figures 1 and 2 in the amounts shown in Table 1 along with the step durations. No
transfer steps or due-dates are considered.
The first element of problem size in the formulation of [11] is the discretization of time.
In order to exactly model the solution space, it is necessary to use a discretization width equal
to the greatest common factor of the step durations and resource event times. In this case, that
interval is of width 0.5. Since there are 4 tasks with processing units specified, the scheduling
horizon of 30 time units results in a total of 240 binary variables. Since the formulation of
[11] also determines batch size, and since there are 6 states of material in the example, there
are a total of 360 continuous state variables. An additional 120 continuous variables are
necessary in order to model the storage of material in a processing unit during idle time. In
order to ensure that no unit begins more than one task during any interval, 120 binary
constraints are necessary. In order to ensure that no step begins processing on a unit while
another step is still in process, 240 binary constraints are needed. In order to force the amount
of material undergoing a task to 0 when the task is not in process, 240 mixed
binary/continuous constraints are used. Material balances result in 60 continuous constraints.
Since there are two resources and 60 time periods, 120 mixed binary/continuous constraints
are necessary in order to assure that resource levels are not violated. The consideration of
storage of material in an equipment item adds 240 mixed binary/continuous constraints which
ensure that no material is stored in an equipment item while processing is taking place, and
120 more mixed constraints which ensure the forward flow of stored material.
Using intelligent interval analysis as presented in the earlier portion of this paper, the size
of the formulation suggested in [11] can be drastically reduced. The ideas of the creation of
RCI's and the RCI set elimination criteria can be used in order to compute intervals of the
scheduling horizon during which some of the steps will never take place. The result is two-
fold. First, the number of discretization variables is reduced for some of the steps. Second,
the reduction in the number of variables implies that fewer constraints are necessary. The
number of binary variables can be reduced from 240 to 98. In addition, the total number of
continuous variables can be reduced from 720 to 496. Because of the reduction in the number
of binary variables, the number of binary constraints is reduced from 360 to 148 and the
 803

Table 2 - Fonnulation Size Comparisons

Kondili, et al Reduction Proposed Reduction

Item Kondili, et al with Analysis Factor Fonnulation Factor
Binary
Variables 240 98 0041 30 0.13

Continuous
Variables 720 496 0.69 4 0.006

Binary
Constraints 360 148 0041 88 0.24

Continuous
Constraints 60 58 0.97 2 0.03

Mixed
Constraints 720 362 0.50 30 0.04

number of mixed constraints from 720 to 362. There is very little change in the number of
continuous constraints.
The fonnulation presented in this paper coupled with intelligent interval analysis results
in a significantly smaller formulation. Intelligent pre-processing of intervals results in 22
binary Xij variables. Consideration of step sequencing results in 8 unspecified sij variables. A
total of 20 binary constraints are introduced concerning step sequencing. Renewable resource
level considerations give 6 binary constraints. Note, however, that as the number of steps
increases, this number has the potential to rise quickly, depending on problem structure.
There are 30 mixed constraints required for linking the discrete and continuous portions of the
problem. In order to ensure correct use of the time horizon, 62 binary constraints are
introduced. Only 4 continuous variables are introduced to model the start times of each of the
steps. Finally, 2 continuous constraints are used in order to separate the start times of
successor and predecessor steps as specified in the product recipes.
The resulting fonnulation sizes are summarized in Table 2. The column labelled,
"Kondili, et aI," contains estimates on the formulation size as presented in [11]. The column
labelled, "Kondili, et al with Analysis," gives an estimate of the size of the formulation in [11]
with the application of the interval elimination criteria alluded to in this paper and described
in [23]. The fIrst reduction factor column is a decimal which represents how large the
Kondili, et alJ6rmulation is with interval analysis compared to the raw formulation. The next
column presents the size of the fonnulation presented in this paper. The [mal reduction factor
column gives the size of the proposed fonnulation as a fraction of the raw Kondili, et al
fonnulation. When examining the reduction factors, it is important to remember that the
reduction in solution complexity is exponentially, not linearly, related to the number of binary
 804

variables. Therefore, a reduction of 50% in the number of binary variables reduces the
solution complexity by significantly more than 50%.
In fairness, it is important to mention that the Konctili, et al formulation includes
considerations of batch sizing, which can not be eliminated from the formulation even when
such decisions are not necessary. The result is an increased number of continuous variables
and mixed constraints (in this example, 240 continuous variables and 240 constraints). Batch
sizing decisions are an important consideration in practical problems, but are not included in
the current form of the proposed formulation, which only involves the detailed scheduling of
known tasks.
It is equally important to mention that no additional variables or constraints are added to
the proposed formulation in order to model material storage in the processing unit, whereas
the formulation of Kondili, et al required 120 extra continuous variables for this purpose.
Further, the proposed formulation allows the storage of material to take place prior to a task,
after a task, and between the various processing steps of a task. The formulation of Kondili,
et al allows idle time only prior to the initiation of processing. Specifically, whatever
material is being stored at some time in a unit must later be processed in that same unit. This
implies that unless two successive recipe tasks are assigned to the same unit, material must
vacate the unit as soon as processing is completed. The formulation proposed in this paper
allows more flexibility in fitting the various resource demands under resource profiles
through the use of idle time when external storage vessels are not available.
The question remains of which formulation is ultimately the best. The information in
Table 2 seems to indicate that the formulation presented here is not just marginally better than
previous work, but that it is a quantum leap. While this is clearly true for the example
problem discussed here, it is not universally the case. When problems can be conveniently
approximated with large, uniform time discretization intervals, and when there is a large
number of steps which must be sequenced, the conventional formulation could be a better
choice. However, the user of a conventional formulation must decide when the round-off of
task durations and resource events (necessary in order to get a coarse time discretization)
begins to represent a problem different than reality. The formulation presented here will
exactly model the process regardless of the values of task durations and event times. The
approximate conventional formulation may yield an .exact solution, but the result may be an
exact solution fo an approximated problem and not an exact solution to the exact problem.
This limitation becomes especially important when dealing with tight scheduling applications
in which data rounding may exclude feasible solutions.
 805

Conclusions

In this paper, several aspects of the resource-constrained batch chemical process scheduling
problem were analyzed in an exact mathematical form. First, it was demonstrated that, by
judicious choice of modelling variables through the use of RCI's and RCI elimination criteria,
the size of a formulation can be drastically restricted. The formulation was then presented
with respect to all aspects of the detailed scheduling problem discussed in the problem
defmition section. The flavour of the formulation was motivated by the scheduling method
outlined in earlier work [23,24,25]. Finally, comparison with the currently accepted
formulation on a small problem indicated promise for further study of enhanced formulations
based on the notions of interval analysis.
All variables in this formulation can be mapped to some exact physical cause. No
artificial constructs have been added to the problem in order to generate the model. In
comparison with the traditional discretization model, the model presented here can be either
better or worse in terms of solution complexity, depending on problem structure. A properly
structured problem can always be used to demonstrate that one method is better than another,
but such demonstrations seldom result in gaining any fundamental understanding about the
problem. What is important here is that a new formulation is given which is, in essence, an
exact formulation, whereas previous modelling approaches have effectively rounded the
problem data and give an approximate model. While the solution of such approximate
models produces an "exact" solution, it is important to realize that an "exact" solution to an
inexact formulation does not necessarily give an optimal solution to the original problem. In
terms of analogy, consider the area of process modelling. It is customary to generate an exact
model of the process in terms of a set of differential equations and boundary conditions.
Typically, this model cannot be analytically solved, so time and space are discretized and a
numerical solution is generated. For practical use, an empirical model is as useful as the
numerical solution; however, modelling is a valuable exercise in further understanding of the
process. The same is true of building models for scheduling problems, except that here,
sometimes even discretized models are not solvable. The model presented here can be useful
both for further understanding of the process constraints as well as for obtaining practical
solutions for small to medium sized problems. The reader should now be aware that there
exists more than one means of mathematically modelling a problem, that the choice of the
best paradigm for a given application requires some analysis, and that by investing time
toward problem analysis, order of magnitude or greater savings may be realized.
 806

Nomenclature

AMOS augmented maximally overlapping set.

start time of step i.
the maximum amount of time which can elapse between the start times of steps i
andj.
the minimum amount of time which must elapse between the start times of steps i
andj.
Duration of step K, where K is one of the step notation subscripts.
duration of step x.
a small decimal number.
event free interval.
e average number of resource events per resource.
eN event time of non-renewable resource N.
H length of the scheduling horizon.
K the number of augmented maximally overlapping sets.
M a large number.
MOS maximally overlapping set.
NRR non-renewable resource.
NT number of terms on a given side of a constraint.
P number of RCl's per step.
q average number of set-up steps in a set-up operation.
'R.. a set of resourC'es used by potentially interacting steps.
RCI resource confluence interval.
resource level of resource R available at time e.
renewable resource.
resource level required of resource R by step s.
r number of resources in a problem.
s a set of potentially interacting steps.
s number of steps in a problem.
sequence variable for step i preceding step j.
number of tasks assigned to a given unit.
the.siart time of an EFI.
the end time of an EFI.
number of tasks in a problem.
a binary variable indicating the use of EFI j by step j.
y number of storage options following the completion of a task.
~j a binary variable indicating that Xij represents the fIrst EFI used by step j.
z average number of transfer and clean-out steps associated with a use of storage.
 807

References

I. Baker, K.R. Introduction to Sequencing and Scheduling. Wiley, New Ycck, 1974.
2. Blazewicz, Jacek, Wojciech Cellary, Roman Slowinski, and Jan Weglarz. Scheduling Under Resource
ConstIaints - Detenninistic Models. Annals of Operations Research, Ed. Peter L. Hammer, J.e. Basltzer
AG Scientific Publishing Company, Switzerland, Vol. 7, No. 1-4, December, 1986, pp. 1-226,
3, Bowman, E,H. "The Schedule Sequencing Problem." Operations Research, Vol. 7,1959, pp. 621-624.
4. Con, BJ. and S. Machietto. "A General Completion Time Determination Algcrithm for Batch Processes."
AIChE Annual Meeting, San Francisco, November, 1989.
5, Egli, U.M. and D.W.T. Rippin. "Short-Tenn Scheduling for Multiproduct Batch Chemical Plants."
Computers and Chemical Engineering, Vol. 10, No.4, 1986, pp. 303-325.
6. Fox; Mark S. and Stephen F. Smith. ''ISIS - a Knowledge-Based System for Factory Scheduling." mm
Systems Jrurnal, Vol. I, No. I, 1984, pp. 25-49.
7. French, S. Sequencing and Scheduling: An Introduction to the Mathematics of the Job-ShqJ. Horwood,
Chichester, 1982.
8. Graves, S.c. "A Review of Production Scheduling." Operations Research, Vol. 29, No.4, 1981, pp. 646-
675.
9. Gupta, S.K. and J. Kyparsis. "Single Machine Scheduling Research." Omega International Jonnal of
Management Science, Vol. IS, No.3, 1987, pp. 207-227.
10. King, 1.R. and A.S. Spachis. ''Heuristics for Flow-ShqJ Scheduling." International Journal of Production
~, Vol. 18, No.3, 1980, pp. 345-357.
11. Kondili, E., C:c. Pantelides, and R. W.H. Sargent. "A General Algorithm for Scheduling Batch
Operations." Third International Symposium on Process Systems Engineering, Sydney, August, 1988, pp.
62-75.
12. Lazaro, M. and L. Puijaner. "Simulation and Optimization of Multi-Product Plants for Batch and Semi-
Batch Operation." I. Chern. Symp. Series, 92, 1985, pp. 209-222.
13. Manne, A.S. "On the Job-Shop Scheduling Problem." Operations Research, Vol. 8, 1960, pp. 219-223.
14. Park, Yang Byung, C. Dennis Pegden, and E. Emory Enscore. "A Survey and Evaluation of Static
FlowshqJ Scheduling Heuristics." International Journal of Production Research, Vol. 22, No. I, 1984, 127-
141.
15. Pritsker, A.A.B., LJ. Watters, and P.M. Wolfe. "Multiproject Scheduling with Limited Resources: A
Zero-One Programming Approach." Management SCience, Vol. 16, 1%9, pp. 93·108.
16. Reklaitis, G.V. ''Review of Scheduling of Process Operations." AICHE Symposium Series, Vol. 78, No.
214,1982, pp. 119-133. .
17. Reklaitis, G.V. "Perspectives on Scheduling and Planning of Process Operations." Proceedings of the
FaJrth International Symposium of Process Systems Engineering, Montebello, Canada, August, 1991.
18. Rippin, D.W.T. "DeSign and Operation of Multiproduct and Multipurpose Batch Chemical Plants - An
Analysis of Problem Structure." Computers and Chemical Engineering, Vol. 7, No.4, 1983, pp. 463-481.
19. Sahinidis, N.V. and I. Grossman. "Reformulation of Multiperiod MILP Models fcc Planning and
Scheduling of Chemical Processes." Computers and Chemical Engineering, Vol. 15, No.4, 1991, pp. 255-
272.
20. Tongo, G.O. Scheduling of Batch Chemical Processes with Resource Constraints. Diss. Purdue
University, West Lafayette, IN, August, 1990.
21. Tsirukis, T. and Reklaitis, G.V. "A Comprehensive Framework fcc the Scheduling of Resource
Constrained Multipurpose Batch Plants." Proceedings of the Fourth International Symposium of Process
Systems Engineering, Montebello, Canada, August, 1991.
22. Wellons, M.C. and G.V. Reklaitis. "Problems in the Scheduling of Batch Chemical Processes."
TIMS/ORSA Joint National Meeting, Paper no. MA35.3, April, 1988.
23. Zentner, Michael Gerard. "An Interval-Based Framework for the Scheduling of Resource-Constrained
Batch Chemical Processes." Diss. Purdue University, May, 1992.
24. Zentner, M.G. and G.V. Reklaitis. An Interval-Based Approach for Resource Constrained Batch Process
Scheduling Pan I: Interval Processing Framewak. Computer Oriented Process Engineering Proceedings
ofCOPE-91, L. Puigjaner and A. Espuna Eds., Elsevier, Amsterdam, October, 1991, pp, 151-157.
25. Zentner, M.G. and G.V. Reklaitis. "An Interval-Based Approach for Resource Constrained Batch Process
Scheduling Pan ll: Assignment and Adaptive Stccage Retrofitting." AlChE Annual Meeting, Los Angeles,
Paper No.14Od, November, 1991.
 Batch Processing in Textile and Leather Industries

L. Puigjaner, A. Espufia, G. Santos, and M. Graells

Dept Enginyeria Quimica, Universitat Politecnica de Catalunya, Avda. Diagonal. 647. E-08028 Spain

Abstract: Incorporating new technologies to the batch process industries constitutes a difficult
task essentially dependent on the specific characteristics of the type of industry contemplated.
Present trends in the methodology used to solve the efficient production scheduling problem can
be grouped into two categories: the general p~ose scheduler that employs the rigorous Mixed
Integer Linear Programming (MILP) optimization strategies approach and the development of a
heuristic-based customized approach which focusses on particular industrial sectors.
This paper describes how the combination of optimization and expert systems has been used
to solve the production scheduling problem in the textile and leather industrial sectors. Detailed
plant simulation is essential for production plant evaluation and eventual validation. which is
facilitated by the data base management system. Mixed Integer Linear Programming is used to
solve the optimization problem in the case of the leather industry, while the textile manufacturing
scheduling problem is solved using ad-hoc hierarchical expert rules together with the plant
manager supervision in a highly interactive mode. The reported industrial experience in both
cases will be presented and results from specific case studies will be analyzed.

Keywords: Batch processing, multiproduct plant, production scheduling

Introduction

Batch processing is receiving increasing attention from a wide variety of industrial sectors due to
the potential offered by this kind of production network: flexibility in the assignment of
resources to multiple products. quick adaptation of capabilities to product market demand
fluctuations, easy acceptance of alternate raw materials and/or manufacturing recipes. These
potential attributes must be consolidated by the integration of adequate management tools and
techniques that will enhance the productivity of batch operations, minimize idle times and ensure
uniform product quality.
Computer powered batch processing optimization tools have for some time been developed at
 809

academic institutions with little or no impact on the industrial world. They attempt to solve partial
aspects of the problem with varying degree of realism, thus generating an adverse gap between
promise and the practice. On the other hand, software houses offer a variety of batch
management commercial packages mainly devoted to material resource planning or production
enhancement through simulation, without considering to actual process optimization.
This situation is not alleviated by the fact that batch processing cannot be easily generalized
and that the level of detail required to describe industrial scale operations may become very
complex and often problem-specific, which may involve significant amounts of speciality
products [8].
To obviate the present circumstances two basic approaches are currently underway. On the
one hand, the rigorous treatment of batch process operations optimization is contemplated, even
for the most complex case of multipurpose network configuration [9,5]. On the other hand,
problem specific solutions are attempted within a general framework [1,2]. The rigorous
approach to the comprehensive treatment of the general resource constrained scheduling problem
involves the solution to very large scale non-linear mixed integer programming models that can
only be honestly contemplated in the perspective of developing parallel and distributed
algorithms using highly parallel and distributed computers [7]. This assumes that present
expectations of progress in computing technology will become a reality and a new generation of
low cost supercomputers will be accessible to medium/small sized enterprises.
In the meantime, and even in the long range, the second approach may prove to be a better
alternative in many cases, provided that heuristic formulations of specific problems are
constructed from more informal problem descriptions and/or user intervention using AI
methods. In this paper, this second approach has been used to solve the production scheduling
problem in the textile industry. The strategy used to find good but SUboptimal solutions is
compared with another industrial case study leather manufacturing where strict branch and
bound techniques are implemented to reach the optimum production plan.

Modelling Framework

Computer aided production scheduling of batch plants requires a comprehensive modeling

framework representing the detailed batch/semicontinuous individual process operations, the
connectivity aspects of the process units and the available resources. An open representation
framework has been devised that facilitates the treatment of the general resource constrained
scheduling problem and its application to specific environments [3].
Modeling of plant operations considers mixed product campaigns made-up of the tasks
 810

carried out on each of the m=I,2,3 ... ,M available units. Such modeling includes the
identification of task j=I,2,3, ... ,Jjbelonging to batch n=I,2, ... ,N being processed in equipment
m=I,2, ... , M that is used for the Ilh time. The binary variable Xjlmn describes the assignment of
tasks to individual equipment, and the binary variable Yin is used to associate product i=I,2 ... ,1
and batch number. In turn, any task is made up of five elementary subtasks s=I,2, . .. ,S (S=5),
which may include (Fig. 1) waiting time (1W) only after the operation subtask.
s=1 s=2 s=3 ~ s=5
SCI-UP load operation 1W unload clean-up
F~W ~W74
TrIm TF lm

Fig. 1. Set of subtasks considered in every task starting at time TIhn and futishing at TFhn

Once calculated the tIms, the initial and final times for the operation of equipment m for the IIh
time are related by:

S=5

TF lm = TI lm + Ltlms + TW lm (1)
s =1

The batch or semicontinuous mode of operation of the unit is described by (see Fig. 2):

Zm E (O,l) (2)

b/z.
a)

t ·
W .i??ZZZ/l
.
T 1 1m

VZ///~

b)

Z m= I
fOK""""
T 1 All

Fig. 2. Unit operation: a) semicontinuous operation; b) batch operation

 811

and the time relation between consecutive stages of the same batch n is given by:
M L.

LL
m=11=1
Xj+l.lmn (TI 1m - tlml ) =
M ~ 3

LL Xp.lln [ TI All + (l-z~ {LtAlls + TWAll} +

/1=1 A.=1 s=1
ZIl tAlll] (3)

The relationship between 1 and 1+ 1 successive jobs is given by:

TI I+ I •m ~ TF lm m = 1, .... , M; 1=1, .... , Lm-! (4)

The modeling framework just described considers the plant operation for a certain short-term
period q=1.2, ... Qp of a specific mid-term period p=1,2, ... P. Total demand Di to be covered
during the available time horizon H is specified for each mid-term period p (Dip).

The objective function and constraints

The objective function f considers the penalty costs per product i and production period (P)
associated to production (PTip), inventory (Sip), delays (Rip) and unfilled demand (Dip) [3].

I P

f= L L PTip+ Sip+ Rip+ U ip

_I p.1
(5)

to be minimized under the "Finite Wait" (FW) constraint, which expands the concept of mixed
intermediate storage (MIS) to all possible policies, form Zero Wait (ZW) to No Intermediate
Storage (NIS):
N I J1

LLLY
.,..1 i-I j-l
in Xjlmn TWij: ~ TW lm ~ 0 '\:I I,m (6)

and the time horizon constraint:

T \XI = max
I,m
{TF1m } (7)

Utilities constraints are given by:

M N S =5 K ij.,

W u(t) = LL LLW njskuC t) ~ W:ax ; -00 < t < 00 '\:Iu (8)

j=1 n=1 s=d k=l

where the contribution of consumption njsk to general utility u pattern can be expressed by
equations:
 812

Wklmsu (t) =~
N

ttI I;

X jImn"Y in" CO ijsku {O( t - TI~msu) - O( t - TF~msu) }

"i/ k,l,m,s,u (9)

s -1 N I I;

TI~msu = TI Im + I. tima + TWIm"O(s-4) + tims I. I. I. XjImnYin OtijkmSU (10)

0=1 n=1 i=1 j=1

s -1 N I I;

TF~msu = TI lm + I. tIma+ TW Im"O(s-4) + tims I. I. I. XjImnYin Otijkmsu (11)

0=1 n=1 i=1 j=1

being:

t S; 0
t> 0

Kijsu = times utility u is used during subtask ijs; k = 1, """, Kijsu

COijsku = utility u power used the kill time during subtask ijs
OtOijkmsu = starting time, related to subtask time, of consumption Wklmsu
Otijkmsu = fmishing time, related to subtask time, of consumption Wklmsu

Additionally, intermediate storage considerations are also being incorporated in the actual
formulation and will be presented elsewhere [10].
The preceding formulation of the problem of optimum production scheduling of a general
processing network is a mixed integer non-linear programming problem (MINLP). Solutions to
this problem can be obtained in reasonable computing time using rigorous approaches like
reformulation techniques applied to linearize it in the space of the integer variables [12], and
specific problem decomposition techniques [11,6] when the number of products is relatively low
(i. e. less than 50).

Case study

The rigorous approach has been used to solve the optimization of process operations in the
leather industry [4]. Detailed simulation has been carried out, covering the different
manufacturing stages from pickled skins to "finflex" operation, and totalling 25 steps that are
differently shared or bypassed according to the individual recipes of the final products (Fig. 3).
The general modeling framework has been accommodated to accept the specificities of this type
of industry, leading to the following working hypotheses taken from industrial practice:
 813

Temperature Ind humIdity controled store

C TANNING

STORE: TANNED SKINS <l.ASSIFIED FOR DIFFERENT USAGES

r+ RETANNING
'- DYEING
FAn.IQUORlNG

IAIR HANGINg. I HOf AIR TUNNEL DRYE3 • DRYING

I MAN' fAI TACiNd • I AI nTlMATICTAClCTNd •

(0) Product dependent optional operation

Fig. 3. Flowsheet of process operations in the leather industry

Every production step is carried out in batch/semicontinuous equipment.

Set-up and clean-up times are considered for each unit.
Task transfer mixed policy. is permitted (FW).
Sharing manpower is the main constrained resource.
The jobshop network configuration is considered.
Batch processing advancement or postponement is allowed.
 814

Then, the optimization of process operations is attempted by minimizing the total elapsed time
required to produce all products in the required quantities (rnakespan) for a specified time horizon
and under particular demand and utilities constraints. While the time horizon provides for a
continuous period of time, flexible stop and starting production times can be introduced according
to manufacturing needs.

A specific characteristic of this type of industry is the shared use of limited resources at
discrete levels. Such is the case of labor: several operations may share different tasks at different
times. As a consequence, the production scheduling problem becomes acrylic (aperiodic), leading
to an increased source of bottlenecking situations. Production planning also becomes especially
complex due to the multi-site characteristics of the production facilities that require the adequate
management of utilities at a particular site, the allocation of production to tasks, and the inter-site
transportation requirements.
Manpower limitation has been analyzed in detail in the overall process operation
optimization. The optimization procedure behaves very well, finding satisfactory solutions to the
aperiodic constrained scheduling problem caused by shared limiting resources [4].

The heuristic approach

An alternative approach to solving the constrained scheduling problem is implicit enumeration.

Branch and Bound methods are generally very useful in discrete enumeration. But the bounding
procedure depends on two key elements: a lower bound on the value of the performance index
associated with the solution of each subproblem and a trial solution to the original problem.
Heuristic rules are useful for providing the trial solution or upper bound in the procedure.
General heuristic procedures are of limited use in industrial practice when the number of
products and units increases significantly (over 50 jobs and 50 machines).
A user-guided heuristic methodology has been developed for specific scenarios. Basically, a
branch and bound methodology is used to solve the optimization problem but a set of ad-hoc
heuristic rules are provided for upper bounding and lower bounding and to conduct the search
tree towards good solutions in less computing time. Experience-based heuristics have proven to
be very helpful in finding realistic industrially valid solutions with reasonable computing effort
even in complex production networks, as will be discussed later. A catalog of priori zed rules can
be automatically used to construct the tree, but a semi-manual mode is also possible, making use
of the knowledge and experience of the expert engineer in two ways: by adding and/or
modifying the existing catalog and by influencing the tree search using qualitative information to
decide which variables should be fixed at any level during the tree branching (Fig. 4). The
algorithm includes "manual optimization tools" of graphic interactive characteristics that can be
actuated via mouse using the Gantt chart representation of the occupancy of processing and
 815

Marketing Constraints

New Specific Objectives

Simulation Results

Fig. 4. Simplified flowchart of the heuristic approach

storage units. The level of branching and the number of alternative solutions to be kept or
fathomed -if wished- are under user control, but if left under the automatic mode they will be
handled in such a way as to minimize computer memory storage.

Case study

Heuristic .procedures are especially attractive for finding problem specific solutions, mainly
when the complexity involved in the manufacturing task representation and the very large variety
of interrelated products and other organizational characteristics require tailor-made strategies that
cannot be dealt with in general purpose algorithms, even with sophisticated knowledge
representation and high performance optimization solvers.
Such is the textile industry. Producing textiles requires a complex manipulation of a large
number of variables related to fabric properties such as twist, sizing, dying, printing and
finishing characteristics. In order to study the complexity of the problem we wiII concentrate on
the weaving stage, one of the most often difficult and economically penalizing steps in the textile
manufacturing process (Fig. 5). Here, different fabrics and designs are produced using diverse
types of machines which can be shared by some products.
This situation leads to a very complex problem of task assignment and optimization of
production lines for the best use of existing equipment in order to meet some specified demand.
Typical figures (Table 1) indicate that even for qualified and experienced personnel making the
 816

Fig. 5. Production recipe for the test case

production plan of such an industrial facility is a burdensome task if it has to be done manually,
even without considering production scheduling optimization.

Table 1. Flowshop characteristics

Problem Dimensions

Total Number of Processing Stages ............................................... 8

Total Number of Processing Units in the Main Stage ....................... .450
Number of Types of Processing Units in the Main Stage ..................... 21
Total Number of Different Products ......................................... 34000
Number of Product Families .................................................... 800
Total Number of Compatibility Restrictions ................................ 12000
Number of Compatibility Aspects ................................................ 30
Total Number of Production Priority Aspects .................................. l0

An ad-hoc solution strategy has been developed using two independent modules. The fIrst
module deals with:
• Simulation of the plant under diverse production conditions.
• Economic evaluation of such production plans according to an objective function that includes
all elements allowing quantification, raw materials costs, operation time costs for each
production line, utilities and manpower requirement costs, penalty costs for delivery time
delays, etc. This evaluation implies the execution of alternate production plans that meet the
desired demand.
• Checking for all production restrictions. The availability of the plant resources indicated
above will determine the flexibility of each production plan.
• Verifying the discrepancies against the objectives described above.
 817

The second module generates alternate long and short term production plans according to the
general objectives described above and additional information coming from the plant status at the
current time. The optimization procedure is indicated in Figure 6, the subplant calculation is
shown in Figure 7 and the short term algorithm flowsheet is given in Figure 8.
As it is essentially an order-driven system, an
additional list of general priorities taken from practical
experience, which can be conventionally weighed up by
the plant manager, has been included in the objective
function.
1. First serve children's size demand (as the market
requires).
2. Some articles must be produced fIrst in order to permit
subsequent specifIc processes.
3. Assorted lots must be produced simultaneously.
4. There is a set of preferred customers that must be
served fIrst.
5. Due dates must be guaranteed for each product.
6. Individual product campaigns are preferred over
alternating products in each unit. There are machine
dependent penalty costs associated with product
changeover.
7. Assign priorities to each unit, taking into account its
production characteristics.
Fig. 6. The optimization procedure. 8. Establish the appropriate products to be produced for
each unit set-up.
Some of these constraints may be found to be contradictory, and then appropriate penalty
functions are evaluated to choose the prevalent one, and fInally assign the production order to the
specifIc unit. As the optimization procedure only considers single variable objective functions,
all constraints and penalties have to be expressed in economic terms.
During day-to-day work, the short-term production planning module provides information
necessary to modify current production plans, which are manually generated or suggested by the
long-term optimization module, and conveniently adapts them to changing production patterns
not considered in the long-term planning.
Finally, the simulation module allows the production manager to easily modify long and
short term production plans, and make an objective evaluation of the consequences of such
modifIcations (or of the ones suggested by the short-term production planning module). This is
very useful when market pressures introduce unexpected changes in a long term production
planning policy.
 818

Fig. 7. The subplant calculation specified flowchart

The software package has been implemented in a workstation and provided with over 100
custom-mode colour windows to facilitate input and user interaction, thus obtaining high
versatility and ease of training in the use of the program. Extensive use of graphics provides
useful information to the user via Gantt charts, histograms and the like. It is worth note that
although the input/output interfaces and the solution structure itself, are specific to the textile
sector, this software framework can be easily extended to cover other industrial applications that
may have common similarities.

Final considerations

For firstly thing, at the process topological level, it would be desirable to have an interactive
scheduling system provided with an adequate interface to input the user's model that accurately
represents the process plant. Detailed deterministic simulations can be carried out using existing
simulators [13] that provide the expected performance of the plant for a given schedule and
alternative configurations. Unfortunately, a general purpose simulator should be able to model a
broad range of process industries, making it too costly in hardware and software for the
smalVmedium enterprise due to the variety and specificity of this type of production facilities.
 819

Fig. 8. The short tenn algorithm flowchart

However, it should be understood that one way or another plant modeling must be performed at
the highest possible level of accuracy. This is a hard reality that requires further investigation,
since it is not clear either on economical or on technical grounds that more powerful interfaces
incorporating object-oriented structures capable of representing any plant configuration would be
a better alternative to specialized process engineering houses making specific plant applications
of standard low-cost scheduling optimizers ready for interfacing with the specific application.
Two industrial applications have been presented that show how both, exact solution
methodologies and heuristic implementation can be valid for solving the constrained scheduling
optimization problem, provided that detailed framework modeling is used. But at the same time
they point out the need for custom-made solutions oriented towards a limited range of
applications with common characteristics. Diversification of commercial schedule optimizers
may be a good policy both in terms of economy and quality.

Acknowledgements

Part of this work was carried out under the auspices of the CIRIT (Comissi6 Interdepartamental
de Recerca i Investigaci6 Tecnolbgica, projecte QFN 89-4006). Support received from the CEC
(JOULE 43) is also thankfully appreciated.
 820

References

1. Abad, A., Espuna, A., Puiganer, L.: Computer Simulation and Optimization of Textile Fibers Manufacturing
Process Operations. Computer Oriented Process Eng. (ed. by L. Puiganer and A. Espuna), pp. 177-184 (1991)
2. Espu/\a, A., Puigjaner, L.: Solving the Production PI anning Problem for Parall e I Multiproduct Plants. Chern. Eng.
Res. Dev., 67, pp. 589-592, 1990
3. Graells, M.: Working Paper. Universitat Politecnica de Catalunya, Barcelona 1992
4. Graells, M., Espulla, A., Puiganer, L.: Optimization of Process Operations in the Leather Industry. In: Proc.
ESCAPE-I Symposium. Elsinore: Denmark 1992
5. Lazaro, M., Espuna, A., Puigjaner, L~ A Comprehensive approach to Multipurpose Batch Plants Production
Planning. Compul. & Chern. Engng. 13, 1031-1047 (1989)
6. Lazaro, M., Puigjaner, L.,: Solution of Integer Optimization Problems subjected to Non Linear Restrictions: An
improved algorithm, Compul. & Chern. Engng. 12,443-448 (1988)
7. Pekny, J., Venkatasubramanian, V., Reklaitis, G.V.: Prospects for Computer-Aided Process Operations in the
Process Industries. Computer Oriented Process Engineering (L. Puigjaner and A. Espuna, eds.), pp. 435-447.
Amsterdam: Elsevier 1991.
8. Puigjaner, L.: Improving the Design and Management of Production Chains in the Process Industry. Multi Suppliers
Operation (W. van Puyembrocek, ed.), pp. 10. I - 10.20. Bruxelles: lOS Press 1992
9. Rapacoulias, C., Shah, N. and Pantelides, C.C.: Optimal Scheduling of Order-Driven Batch Chemical Plants.
Computer Oriented Process Eng. (L. Puigjaner and A. Espuna, eds.), pp. 145-150. Amsterdam: Elsevier 1991
10. Santos, G., Espulla, A., Graells, M, Puigjaner, L.: Improving the Design Strategy of Multiproduct Batch Plants with
Intermediate Storage. A1ChE Annual Meeting, Florida, 1992.
II. Shah, N.,Pantelides, C.C., Sargent, R.w.H.: Efficient Solution Techniques for Optimal Scheduling of Batch
Operators. Submitted to Compul. & Chern. Engng ( 1991)
12. Shainidis, N.V., Grossmann, I.E.: Reformulation of Multiperiod MILP Models for Planning and Scheduling of
Chemical Processes. Compul. & Chern. Engng. 15,255-272 ( 1991)
Baker's Yeast Plant Scheduling
for Wastewater Equalization

Neyyire (Renda) Tiimsen l , S. Giray Velioglu2 and Oner Hortaysu l

I Department of Chemical Engineering, Bogazi9i University, Istanbul, Turkey

2 Department of Civil Engineering, Bogazi9i University, Istanbul, Turkey

(present Address: Halk Sigorta AS., Istanbul, Turkey)

Abstract: A Baker's yeast plant was modeled using data obtained from the normal operation
of the plant via Monte Carlo analysis. The model which was first validated with respect to the
normal plant operations was then used to predict wastewater discharges from two scheduling
scenarios. The results showed that the plant wastewater discharge may be equalized if a 25
% reduction may be considered to be economically acceptable since wastewater equalization
will not necessitate additional facilities for wastewater equalization external to the existing.
However, if only one fermentation vat may be added to the process then wastewater
equalization would be realized with 24 % increase in the capacity. The additional cost of the
new vat is considered to be off-set by the increased plant capacity.

Keywords: Scheduling, wastewater equalization, baker's yeast

1. Introduction

Equalization is among one of the most widely used practices for volume and strength reduction
of wastewater. Normally wastewater equalization is achieved by directing the wastewater into
tanks where they are held for a specified period of time. However, it can also be achieved by
good timing and scheduling of operations, especially when wastewater is discharged from a
series of successive activities. In this paper, an approach aimed at the realization of the latter
goal is presented, through the development of a Monte Carlo simulation model as applied to a
baker's yeast production plant.
 822

Monte Carlo methods are well established tools of numerical analysis that aid the decision
making involving the design and operation of complex systems subject to uncertainty. The
name Monte Carlo is derived from the fact that the method includes random sampling from
statistical distributions. In the Monte Carlo approach random numbers are used to represent
the status of each component in the system at an instant of time. As numbers are generated
through random sampling, the entire set of numbers is modified, according to the logic of the
model, in order to represent the new status of the system [3 and 7]. In carrying out a Monte
Carlo simulation, the values of the variables are obtained by random sampling from specified
distributions defined by the characteristics of these variables.

2. Simulation Model

As previously stated, this study aims to present an approach to achieve wastewater

equalization via mathematical simulation. The wastewater discharged from a batch process
show variations in concentrations of pollutants and quantities of wastewater discharged from
batch to batch. Moreover, time and duration of discharges are generally not regular, therefore
are not predictable except in a probabilistic sense. Hence, for industrial activities involving a
series of batch processes Monte Carlo simulation is well-suited, for the analysis and evaluation
of the effects of different scheduling schemes. Clearly, the simulation of wastewater discharges
require the determination of the activities that contribute to the wastewater streams along with
their precedence relationships, followed by data collection. The data should then be examined
and evaluated to define the probabilistic nature of the important attributes of the activities (e.g.
duration and idle time of the activities, quantities of wastewater, etc.)
As Monte Carlo simulation experiments require "data generation" mechanisms by
producing a large number of random variates obeying the desired probability distributions, the
use of Monte Carlo simulation has become more restricted to computers. Most computers,
today, are equipped with a random number generating algorithm. The most commonly used
algorithm, in fact, produces a nonrandom sequence of numbers (pseudo-random numbers) each
being completely determined by its predecessors and consequently all numbers being
determined by the initial (seed) number [ 4].
Upon completion of the evaluation of the random behavior of the components, the validity
of the assessed probability distributions has to be checked. This can best be done by comparing
the observed and simulated behavior of the system. Means of achieving wastewater
equalization can be initiated only after it is shown that the probabilistic nature of the activities,
of the system, are adequately described by the derived distributions and the system is well-
represented by the simulation model. Different scenarios for timing and scheduling of activities
can be experimented with and their "degrees of equalization" can be quantified. Degrees of
 823

equalization stand for the extent of peak load smoothing, which can be quantified by an
"equalization metric", such as the maximum absolute deviation from the mean. It must be
noted that the evaluation of the scenarios should not solely be based on the degree of
equalization. Other important factors, such as the production level, should also be taken into
account. Hence, the simulation experiment should be able to lead to a trade-off analysis
between the degree of equalization and other important factors (i.e. the production level).

3. Application of the Model

Application of Monte Carlo simulation to chemical and environmental engineering problems is

not too common. Tayyabkhan and Richardson [8] have analyzed a complex batch blending
operation by the Monte Carlo technique. Berthouex and Brown [1] used Monte Carlo
simulation to investigate the patterns and quantities of wastes discharged from a tanning
industry. Peak load smoothing by altering the schedules of waste discharges from specific
tanning process batches have been evaluated. However, in their study the Monte Carlo analysis
is focused on evaluation of wastewater characteristic only. They considered the duration of
operations to be deterministic which greatly simplifies the timing and scheduling of activities.
In this study, the duration and the idle time between activities (operations) are viewed within
the perspective of Monte Carlo analysis in addition to the evaluation of wastewater
characteristics. In doing so, Berthouex and Brown's [1] approach is significantly extended.
Although the resulting system is more difficult to deal with, its selection over that of the
system considered by Berthouex and Brown [1] is justified, as the duration and the idle time
of the activities significantly affect the discharge pattern and characteristics of the
wastewater.
The simulation model outlined in the previous section, is applied to a baker's yeast
production plant. It should be noted that most simulation models are system-specific,
therefore, the emphasis should be on the concepts and the logic of the model. Hence, within
this framework; the application below should be viewed as a specific illustration of the
concepts and logic of the model rather than a unique solution to the problem of wastewater
equalization.

3.1 Process Description

Baker's yeast production process consists of rapid multiplication of the yeast microorganisms,
in a controlled medium of nutrients, followed by separation and washing and filtration of the
fermented wort. The flow diagram of a typical baker's yeast production plants is shown in
 824

Fig.I. Yeast plant wastewaters are highly contaminated with organic compounds and thus have
a specific yeasty odor, high color and turbidity with a pH of 4.3 - 6.5 [6].

yeast germ

seed yeast
cooling water
lermenter

ashing
clarified entriluge
molasses

wastewater
seed
yeast
storage
water nutrients

cooling water

separation and
washing
wastewater
centrifuges

wastewater

Figure 1. Flow Diagram of a Typical Baker's Yeast Manufacturing Process

 825

3.2 Data Collection and Evaluation

Data collection was carried out in a yeast plant which had five fennentation vats. In the plant
Vat No 1 and 2 operate alternately, following one another, to produce the seed yeast while the
other three (Vats no. 3,4 and 5) are used in manufacturing the commercial product. (The first
two vats will be referred to as "seed vats", and the last three vats will be referred to as
"commercial vats" from hereon.). The seed vats were of the same size and specifications. A
complete cycle of seed yeast fennentation (i.e. fennentation, separation and washing of the
fennented wort, equipment washing and preparation for the next cycle) was observed to take
about 25 to 26 hours. Fennentation in a seed yeast vat starts approximately when the cycle is
completed in the other seed vat. The total amount of seed yeast required for one batch of the
commercial vats could be produced in a single batch in either of Vats No: 1. and 2. The
commercial yeast vats were not of the same size. Vat No: 5 was somewhat larger and its
fennentation duration is about sixteen hours as opposed to 13-15 hours for Vats No: 3 and 4.
Once the fennentation is completed in one of the vats the fennented wort is separated and
washed. Two centrifuges are used for this purpose, with Centrifuge No. 1 used for the seed
yeast and Centrifuge No. 2 for the commercial yeast. Since the seed vats operate alternately
(one at a time) their cycle completion times do not coincide, and thus the fennented wort can
be directed to Centrifuge No.1 without delay. However, the commercial vats can operate
concurrently, therefore, fennentation may be completed in anyone of them while Centrifuge
No.2 being used for separating fennented wort from another vat. In such a case, the separation
of the fennented wort from the most recent cycle of the commercial vats would be delayed
until Centrifuge No.2 is free.
As the separation and washing operation continues, the yeast milk, the concentrated liquor
of the fennented wort is directed to cool in the storage tanks. It is gradually drawn from the
storage tanks to the filter presses. If any yeast milk is already present from previous batches,
then the most recently washed yeast milk is held in the vessel until the ongoing filtration
operation ends. Otherwise, filtration starts within 10-20 minutes following completion of
fennentation. Finally, the yeast is compressed and packed as the final product for sale.
In the application of the model only the discharges from separation and washing and
filtration operations were considered. In the plant, the cooling waters from seed vats were
discharged into the sewers, while the cooling waters of the commercial vats were reused. The
cooling waters, together with the effluents from equipment washing, though large in quantities,
are not highly contaminated. As suggested by Koziorowski and Kucharski [5], dilute effluents
from fennentation plants are isolated and treated separately, thus concern was focused on the
discharge behavior, patterns and equalization of concentrated effluents. We will note here that
should the importance of wastewater streams not considered in this study become apparent,
they can also be included in the analysis with no difficulty.
 826

Throughout the study, COD (chemical oxygen demand) was used to characterize the
organic strength of wastewater. BOD (biochemical oxygen demand) tests were also carried out
to cross-check the consistency of the values with those reported in the literature. Grab samples
which were collected from the effluents of the centrifuges approximately at the middle of the
discharge periods, proved to be satisfactory as evidenced by running COD tests for two
different samples collected from the same discharge at different time periods. The 15%
difference in the results was considered insignificant in view of the possible experimental errors
for the COD range in question. Composite samples were collected from filtration effluents. All
samples were preserved according to the methods and recommendations given by the EPA [3]
until the COD tests were carried out. Weekly averages of the COD and BODs concentrations
are summarized in Table 1. As can be observed from Table 1, the obtained COD and BODs
concentrations are in generally good agreement with those reported in the literature.
Flow measurements from the filter presses were made manually, by collecting a certain
volume of liquid waste for a known period of time. On the other hand, separation and washing
centrifuges discharged large quantities of wastewater, prohibiting manual measurement. For
the latter case, the height of the fermented wort was measured at the end of the fermentation
period to compute its volume. The volume of yeast milk was computed in similar manner. The
volume difference between the two volumes was assumed to quantify the centrifuge effluent,
and the average flow rate was estimated by dividing the volume difference by the discharge
duration.
Data obtained from the survey, covering a period of two months are given in Renda [7].
Probability distributions for the relevant variables were determined empirically from the
collected data.
Frequency diagrams were constructed from the data for each variable. The probabilistic
behavior of the variables were then visually determined to be described by either normal or log-
normal models by comparison of the frequency diagrams with the respective density functions.
In addition, the data were plotted on normal and log-normal probability papers. When both
distributions appeared to be plausible the preference was made on the basis of chi-square
goodness of fit test. For the data sets that did not fit into either of the above distributions or
into any other theoretical distribution, a "table look-up" method was used. In the "table look-
up" method the cumulative probabilities are computed from the data and the value of the
variate is determined in accordance with its relative frequency. A detailed discussion of data
evaluation is given in Renda [7]. Table 2 summarizes the relevant process variables and their
associated probability distributions. In Table 2 normal and log-normal distributions are
designated by N (m,s) and LN(I,§) where m and s are the mean and standard deviation of the
variate and I and § are the mean and the standard deviation of the natural logarithm of the
variate, respectively.
 827

Table 1: COD and BODS Concentrations for Filtration and Separation and Washing Effluents.

Process COD BODs

Filtration 6880 5525

7200 4100
10560 7250

(Nemerov [6]; <;iler et al.[2]) 7000 5000

--------------------------- ...... _---------------------_ .. --------------- ... -------
Separation - Washing 20000 14050
19200 13800
21600 14250

(Nemerov [6]; <;iler et al.[2]) 21500 16800

3.3 Assumptions

In delineating the time horizon for the analysis, the start and completion times of the activities
are of interest. The 24-hour operating day is divided into 15 minute intervals to retain the detail
and sensitivity of the overall process. Such a time interval was chosen because the shortest
activity duration (i.e. separation and washing were assumed to be adequately covered by 15
minute increments.
To begin a simulation run, the start time of fermentation vats are drawn from a uniform
distribution U (0,96). In other words, it was assumed that any vat could start operation,
independent of others, at any 15 minute interval during the day, noting that there are ninety-
six 15 minute intervals in a 24-hour day. The operational scheme of any activity was assumed
to be independent of the duration of other activities. For example, the separation and washing
begins right after the completion of the fermentation regardless of the duration of the
fermentation. Each vat was assumed to operate once a day, therefore, every seven simulation
runs describe a working week. That is, the number of weeks was increased by one after every
seven cycles (i.e. group of batches) and simulation is stopped when the desired simulation time
in terms of working weeks is reached.
 828

Table 2. Relevant Process Variables and Their Associated Probability Distributions

COD(mgl It)
Filtration (SF!) N ( 8290, 3670)
Commercial Yeast Separation and Washing (SSE) N (19930,2990)
Seed Yeast Separation N (23800, 6390)
Seed Yeast Washing N ( 8400, 3640)
Flow, Q (m 3 115 min)
Filtration (QFI) (QSE) LN ( -l.5966, 0.2090)
Commercial Yeast Separation and Washing LN (2.2819,0.2055)
Seed Yeast Separation LN ( 1.9659, 0.0773)
Seed Yeast Washing LN ( 1.3942, 0.0902)
Duration(hrs)
Separation-Commercial Yeast & Seed Yeast (TSE) - Table look-up method
Filtration - Table look-up method
Fermentation
- Seed Yeast Vats No.1 and No.2 - Table look-up method
- Commercial Yeast Vat No.3 LN (2.7391,0.0627)
Vat No.4 LN (2.7272,0.0555)
Vat No.5 LN (2.8203,0.0356)
Idle Time Till the Next Batch (hrs)
Seed Yeast Vats No.1 and No.2 (TID) N(O,I)
Commercial Yeast Vat No.3 LN ( 1.9636, 0.4425)
Vat No.4 LN( 2.1088, 0.4515)
Vat NO.5 LN( 2.4801,0.1641)
Yeast Produced (kg! batch)
Vat NO.3 N (4940, 313)
Vat No.4 N (6150, 572)
Vat NO.5 N (11800,1170)

N - Normal Distribution
LN - Log Normal Distribution
 829

As mentioned in Section 3.2, filtration starts 10-20 minutes after the completion of the
fermentation operation. Since variations within 10-20 minutes can not be differentiated within
the time scales used in this study, the start times of filter presses were assumed to be 15
minutes following fermentation (i.e. one time unit).

3.4 Mathematical Formulation

The first activity of the process is fermentation. A fermentation activity for the i th
batch( cycle) in the i th vat is defined as AFE ij . The completion time of AFE i,} denoted by
FEC ij is simply obtained by adding the duration of AFE ij (i.e. TFE i,j) to the start time (i.e.
FES ij) of AFE ij:

FEC"=FES"+TFE"
IJ I,J I,J (Vi,j) (1)

The completion ofthe activity AFE i,j calls for the separation and washing operation (i.e.,
activity ASE ij). Thus, the start time of ASE i,j (i.e., SES i,j) is set equal to the completion
time of AFE i,r

SES"I,J =FEC"I,J (Vi,j) (2)

As mentioned in Section 3.2, it must be noted that for commercial vats, j = 3, 4 and 5, a
check has to be made regarding Centrifuge No.2. If Centrifuge No.2 is busy, SES i,j
has to be delayed accordingly. Therefore, the completion time of ASE ij , i.e., TSE ij , is set
equal to the start time of the previous separation. Thus, in general
SEC ij = SES ij + TSE ij (V i, j) (3)

where, the duration of separation TSE i,j is computed by the table look-up method.
As the separation process continues, the yeast milk is directed to the filter presses, the
start time of which (i,e., FISij) is one time unit after the completion of AFEi,j (i.e. FECi}

FIS ij = FEC i,j + 1 (V i,j=jX) (4)

It must be noted that Equation (4), as well as any other expression for the filtration
operation, is only valid for commercial vats, since seed yeast is not filtered. This is reflected in
Equation (4) by index jX fjx is a subset of j denoting commercial vats only). The completion
time of the filtration operation (i.e. FIC i,j) is obtained by adding the duration of filtration (i.e.,
TFI ij ) to the start time of filtration (i.e., FIS i,j) :
 830

FIC .. = FlS ..
I,J I,J + TFI IJ
.. (5)

The above series of activities define a complete process cycle (batch). The computations for
the start of the next cycle is initiated by addition of the idle time (Le., TID i,j) to the
completion time offermentation (i.e., FEC ij):

.. + TID ..
.. = FEC IJ
FIS I,J I,J (V ij) (6)

and the computations are carried out until the stopping criteria is met.
Throughout the above computations, whenever separation and washing (ASEi,j) and
filtration (AFIij) activities are ongoing, a wastewater stream is generated having a flow rate
(Le., QSEi,j and QFli,j) and a COD (i.e., SESij and SFIij) concentration in accordance with
their probability distributions.

3.5 Computational Scheme

The average period for the restart of a batch being over 24 hours suggested that a daily
simulation run would not be meaningful. Therefore, a week was chosen as the minimum
simulation period. Since the processes were not continuous waste discharges were tracked by
means of clock index numbers. The flow rates and COD values, from all operations that fall
into the same interval (clock index number) were superimposed. A computer code was
developed following the formulae of Section 2.4. The probabilities and the associated values
for the variables for which table look-up method is used, are included in the program
statements. All other relevant information, such as the mean and standard deviation for a
specific variable obeying normal or log-normal distributions including the total simulation time
are specified within the READ command. For convenience in keeping track of the clock index
numbers, the computations for the commercial vats and seed vats are carried out separately but
the commercial vats (i.e. j=jX) are handled first in the calculations. Details of the
computational scheme and the computer code are given in Renda [7].
Computations are initiated with the generation of three uniformly distributed numbers
between 0 and 96, designating the fermentation start times of commercial vats (i.e., j= jX). As it
is practical and sometimes necessary to delineate the activities in progressive order,
computations are first carried out for the commercial vat having the minimum start time. For
example, if the generated numbers turned out to be 57, 6 and 94, corresponding to the start
times of Vats No 3, 4 and 5 in terms of IS-minute time units, respectively, then the
computations proceed with Vat No.4 since its start time is the minimum. Negative values for
the normal variates were not allowed throughout the computations.
 831

3.6 Validation of the Model

The model is validated by testing the simulation outputs with the observed behavior of the
system. This is often quite difficult since both the simulation results and observed behavior of
the system are of probabilistic nature. Thus, a single simulation outcome, being a random
sample, should not be expected to agree with survey data in every detail. However, it should
agree with the important aspects if the system is properly simulated and if the input data are
realistic.
To validate the model, simulated patterns of the fermentation durations for all the vats are
compared with corresponding survey data in Figure 2. Since the fermentation operation is the
keystone of the overall process an acceptable degree of agreement between the simulated and
observed data is necessary. Figure 3 gives the volumetric flow variations with time as
determined from the model and the flow behavior of the real system. The figure is plotted
discontinuously because the observed data were available only for the day time hours.
Figures 2 and 3 show that the derived distributions adequately describe the probabilistic
nature of the system on the whole. The good agreement between the weekly averages of the
amount of yeast produced, flow rates and COD values obtained from a single simulation run
for seven weeks as given in Table 3 indicate the consistency ofthe model on a weekly basis.
Actually, the outcome of a single simulation run should be viewed as a single observation
of an experiment. Since the analysis in a simulation model is carried through using random
numbers, different simulation outputs are expected to show variations. The sensitivity and the
implicit variability of the model mainly depends upon:
i. the initial state (start-up) ofthe system
ii. the sequence of random numbers used in simulation, that is, the initializers
(seed numbers) used in the random number generation process.
Figure 4 assesses the effect of the former issue. As can be observed from Figure 4, the
simulated and the observed fermentation durations generally agree well after a day. Hence, the
conclusions drawn from the very first day of the simulation run were disregarded. In other
words, the start up period was considered to be one day and the system was considered to be
in its normal state of operations after the first day of the simulation run.
To evaluate the effect of the initializers, 20 simulation runs were made, each with a
drastically different initializer. The 95% confidence intervals for the important variables of the
system, obtained from these runs, were narrow enough to conclude that the system behavior
was virtually independent of the initializers used. Hence it was believed that the system is
satisfactorily simulated and attention was focused on the appropriate timing and scheduling of
activities to achieve wastewater equalization as discussed in the next section.
 832

r-T_Y_P(-+__V_AT-+__________________~24r_----------------~2~4-----------------:24~ (t~~)

......
'" ------=
1==----- _:::I _
--==
1-------=_=-=_=______ _
~-----

F.bruary 10 Fc~ru.ry 11 February 12

- - - - Si""lated
----------- Observed

Figure 2. Fennentation Durations

100

ao ~
II

60
r II
II
I'
,I
,I
I'
n
(1
I'
I I l
40 I \ I'
( I I'
,
I
I
I
I
I
I
I l, \I
20 ,
.I
I
I
I
I
I
I
I , I
, I
, I I

.
1
I
I 1 1 1 , ' I
I
I I I
I Ill'
12 24 12 24 12 24
ti.,. (hrs)

- - - - - - . Silluhted
------------ Oburved

Figure 3. Variations in Wastewater Row Rates with time

 833

Table J: Simulated Weekly Averages for tbe Amount of Yeast Produced, Flow Rate and COD

Week Yeast Production Flow rate COD *

(kg/week) (m3 /week) (kg/week)

1 136425 2321 43658

2 132628 2575 49587
3 138554 2517 51310
4 136261 2374 43596
5 141886 2714 50287
6 136334 2512 48336
7 139960 2137 36717
---------------------------------------------------------------------------------
Mean 137535 2450 46213

* COD kg/week - COD mgllt x Flow m3Jhr x Conversion factor

TYPE YAT ~yl ~y 2 ~y l

1----------- -------
.
-,;.
~
'"
2
-r-- --
______ -~_~_=.~_i!=,.-
-----
4 ____ _

5 1-.

Si ..ula ted ------------- Observed

Figure 4. Duration of Fennentations Following Process Start-up

3_7 Timing and Scheduling of Successive Activities

Apparently the problem of timing and scheduling of activities to achieve wastewater

equalization is very complex because of the large number of possible combinations. Thus, the
solution should be sought through an optimization analysis which is difficult to implement. In
this study since the probabilistic behavior of the operation start and end times do not yield to
 834

the deterministic scheduling algorithms available an intuitive approach was attempted. This
mode of solution will more easily satisfy the system-specific observations.
It was observed that separation and washing operations were largely responsible for the
large deviations in flow and consequently in the stream concentrations. Therefore, it was
thought that distributing separation and washing periods throughout the week would
somewhat result in equalization. The wastes discharged from separation and washing can be
regarded as slug discharges due to their short discharge durations. Thus, if the discharge
durations could be extended then the volumetric discharge rate would necessarily decrease.
Starting from this point, the main aim should be to obtain sequential discharge patterns for the
vats within a given period of time. In other words, the discharges should be arranged such that
when discharging from one of the vats ends, the next vat should have completed fermentation
and start discharging its own waste. Within this framework, the suggested operational scheme
requires attention to the timing and scheduling of fermentation operations in all five vats in
order to fix the start and completion times for separation and washing processes.
From the simulation runs the daily discharge of the wastewater was observed to be around
350 m3. Therefore equalization schemes should aim to distribute the 350 m3 of wastewater
uniformly throughout the day (i.e. at a rate ofl5 m3/hr). In view of the wastewater volumes a
discharge pattern is foreseen at a rate ofl5 m3/hr. However, the complete cycle of Vat No.5,
namely fermentation, separation and washing and preparation for the next cycle, takes at least
26 hours. Thus it is physically impossible to discharge wastewater from Vat No.5 every
22-24 hours. Therefore, it was concluded that either the operational schemes of other vats
have to be arranged so that Vat NO.5 would be able to discharge wastewater after completing
its cycle, i.e. 26-27 hours, (Equalization Scheme I) QI another vat would have to be introduced
into the system to absorb the burden (Equalization Scheme II).

Equalization Scheme I In the present operational schedule, when the seed yeast is being
prepared in Vat No.1, the other seed vat, Vat No.2, is idle and vice versa. This equalization
scheme is based on the idea of making beneficial use of the idle seed vat by allocating it for
commercial yeast production when idle. However, if only one vat is used for seed yeast
production, the daily seed yeast production would not be adequate for the remaining four
commercial vats. Therefore, it is proposed that vats No.1, 2 and 3 are to be scheduled in such
a manner that they could be used interchangeably for both seed yeast and commercial yeast
production. Although such on arrangement may seem to be operationally difficult, it is thought
to be manageable provided that sufficient time is allowed for the interchange between batches.

Equalization Scheme II: In this scheme installation of a new vat (Vat No.6), having the
specifications of Vat No.5 is proposed. This leads to a discharge sequence of vats in order of
 835

1, 3, 5, 2, 4 and 6. The cost of adding a new vat is expected to be offset by the increase in the
overall yeast production.

VAT

-
--
24 24
a) Equalization Schelle 1

---

-I--

--
d.y I 24 d.y 2 24 day 3
b) ~qual ization Sche..e II

- - - C....... rci.1 yeast discharge

---------- Seed ye.st discharge

Figure 5. Fixed Discharge Periods to Provide a Relatively Continu...

When the discharge periods, that is, the start and completion times, for the separation and
washing operation are fixed as depicted in Figure 5 such that they follow one another the
resulting flow and COD variations appear as given in Figure 6 and 7, respectively. As can be
observed from these figures both equalization schemes definitely achieve a noticeable degree of
equalization, both in flow and COD when compared with the present operational practice. It
should be observed that the mean values of flow in Figure 6 and COD in Figure 7 for
Equalization Scheme J, Equalization Scheme II, and the present operational practice, although
close to each other, are different. This is due to the random behavior of the output variables as
well as due to the fact that different operational schemes produce different quantities of yeast
and consequently discharge wastewaters having varying characteristics.
 836

To quantify the degree of equalization in a given operational scheme, three different

expressions (i.e., equalization metrics) were used:
1) average of the absolute deviations from the mean
2) average of the squares of the deviations from the mean,
3) maximum absolute deviation from the mean
The three equalization metrics for the different schemes are summarized in Table 4, as
obtained from a seven week simulation run. Apparently, Equalization Scheme I results in a
satisfactory wastewater equalization when compared to the present operational practice at the
expense of a reduction of about 35 tons/week, i.e. 25%, in yeast production. On the other
hand, Equalization Scheme II achieves a better degree of equalization while increasing the
yeast production by about 33 tons/week, i.e. 24%, at the expense of the installation of a
new vat.

Table 4. Yeast Production and degree of equalization

Scheme Yeast COD

Produced Max. D Avg. D· Avg. D2 Max. D
(ton/week) (rn 3/hr) (rn 3/hr) (rn 3/hr)2 (rn 3/hr)

Present Scheme 137.5 72.9 41.3 77.5 1264 739 1487

Eq'tion Sch'm I 103.8 16.9 12.0 17.6 397 253 331
Eq'tion Sch'm II 170.1 5.0 3.1 4.3 210 127 163

* D = absolute deviation from the mean

4. Conclusions

A system consisting of a series of successive activities each possessing random behavior is

shown to be satisfactorily simulated by a Monte Carlo model. Such a model was used with the
aim of achieving flow and strength equalization of wastewater. The proposed approach
involves appropriate timing and scheduling of the activities, and thus the discharges, based on
system-specific observations. An application of the proposed methodology to a baker's yeast
 837

production plant revealed that the method is indeed feasible in achieving a satisfactory degree
of equalization.
The smallest time increment considered in the analysis was 15 minutes. It is thought that
shorter time increments would lengthen computations with unnecessary details. Apparently, the
computations with 15 minute increments resulted in simulation clock index numbers in the
orders of thousands. This would mean more computational load and would not improve the
precision ofthe method appreciably.
It should be emphasized that the proposed approach can be adapted for other equalization
purposes, such as idle machine hours, storage area for products, process stream, etc.. In all
cases the simulation results can lead to a trade-off analysis between degree of equalization and
product level and the decision maker can choose hislher operational policy in view of his/her
personal preference.

References

1. Berthouex, P.M. and L. C. Brown, 'Monte Carlo Simulation of Industrial Waste Discharges', ASCE,
Journal of Sanitary Engineering Division, V. 95, Oct. 1969.
2. Ciler, M., Cetinkaya, N., Dumlu, G., I1han, R, Kinayyigit, G., MutIu, H., Oktem, O. Timur, A., and
Topal, H., "Pakmaya Fabrikasi Aritma Tesisleri Proje Calismalari Ara Raporu ", TOBIT AK, Gebze (
Turkey),1981 (in Turkish)
3. EPA-U.S. Environmental Protection Agency, Manual of Methods for Chemical Analysis of Water and
Wastes, Techoology Transfer Series, National Environmental Research Center, Cincinnati, Ohio, 1974.
4. Fishman, G.S., Concepts and Methods in Discrete Event Digital Simulation, Joho Wiley and Sons, Inc.,
New York, 1973.
5. Koziorowski, B. and J.Kuchorski, Industrial Waste Disposal, Wydawnictwa Naukowo-Techoiczne,
Warsaw, 1972.
6. Nemerow, N.L., "Liquid Waste ofIndustry: Theories, Practices and Treatment", Addison-Wesley Pub.
Co., Reading, Mass., 1971.
7. Renda, A.N., "Wastewater Equalization by Monte Carlo Simulation", Masters Thesis, Dept. of Chemical
Engineering, Bogazi9i University, Istanbul, Turkey, 1981.
8. Tayyabkhan, M.T. and T.C. Richardson, 'Monte Carlo Techniques', Chemical Engineering Progress,
V.61-1, Jan. 1965.
Simple Model Predictive Control Studies
on a Batch Polymerization Reactor

Ali Karaduman and RIdvan Berber

Department of Chemical Engineering. Ankara University. Tandogan. Ankara. 06100 - Turkey

Abstract : A single step control algorithm has been developed and implemented for the
temperature control of a batch styrene polymerization reactor. The algorithm is based on the
energy balance of the system with assumed linear dependency of the overall heat transfer
coefficient on feed flow rate through the cooling jacket. The heat generation term in the energy
equation is treated as an inferentially measured process disturbance and predicted by on-line
process measurements. The algorithm was tested in real-time runs and resulted in a good
performance for maintaining the reactor temperature at its set point during the isothermal reaction
stage.

Keywords : Process control. model predictive control, batch polymerization

Introduction

Batch chemical reactors are extensively used in polymer, chemical and pharmaceutical industries
because of kinetic advantages in some systems. For radical chain polymerization, for example, they
are often preferred for high conversion compared to moderate conversion that may be obtained
with continuous systems with recycle of monomer and solvent. However, because of the fact that
the batch system has no steady state, the dynamic nature is much more complicated and this makes
temperature control difficult. There are still cases in which a batch reactor is simply controlled by
human operator.
It has been well established that there exists close relationships between the operating
conditions of reactors and the quality of the polymer produced. Therefore, the first problem of the
batch polymerization control is to determine the set of time dependent controls, the most
important being the optimum temperature profile. This has been subject for a great deal of
research work [1,5, 14, 18, 19].
In polymerization systems, the primary control objective is to control the polymer quality. This
implies some structural properties of the polymer such as the average molecular weight, the
molecular weight distribution and the chain length distribution which depend on the reaction
conditions and affect the end-use of the polymer. The second objective of the control problem is
to maintain the polymerization temperature as close as possible to its desired profile in spite of
time variant and often nonlinear changes. A servo control strategy is, thus, required to move the
system along the predetermined temperature vs. time trajectory like the one shown in Figure 1.
In fact, these two control objectives (MWD and temperature) are correlated.
Temperature control in batch reactors, therefore, can be divided into three phases: a heating
phase to raise the reactor temperature to its target as fast as possible without overshoot, and a
 839

Time

Figure 1. Desired Temperature Trajectory of a Batch Polymerization Reactor

stabilizing phase to maintain the reactor temperature at a set point or set point trajectory, and a
cooling down period. Different control algorithms are normally required for the heating and
cooling process. Implementing the change in control mode is critical.
For the heating phase, the control policies are bang-bang in nature. The problem during
stabilizing control is the very uneven heat load on the cooling system because the polymerization
reaction exhibits an auto-acceleration rate in the heat generation phase. The change in this phase
occurs so quickly that maintaining a constant temperature in the reactor by an automatic controller
becomes very difficult. As the reaction progresses, the heat generation rate decreases as a result
of the monomer consumption.
The number of studies on control of batch polymerization reactors relative to that of
continuous reactors is limited in the literature. Therefore, there are still unresolved problems and
the area provides an interesting challenge.
MacGregor [11] gave a literature survey on temperature control of polymerization reactors.
Despite of the fact that the most widely used controllers are PID type, the complicated control
requirements for batch reactors may not be fulfilled by classical controllers, because of the time
varying characteristics and other nonlinearities in the process. The standard PID algorithms may
be sufficient only when the polymerization time is large and polymerization heat effects are small.
Kiparissides and Shah [10] used, what they called, an extremely well tuned PID controller in their
simulation run and observed more than two degrees overshoot during the transition from the
heating to the stationary phase under PID control. PID controller tuned to operate in a stochastic
environment gave an oscillatory performance when deterministic measurements were used. They
concluded that a well-tuned PID controller would only perform satisfactorily if the signal to noise
ratio did not change significantly. On the other hand, the evolution of digital control computers
allowed one to produce much better designs without considerations for hardware realizability.
What is required, therefore, for batch reactors is an adaptive control technique which can easily
be implemented in digital computer environment. Kiparissides and Shah [10] evaluated two
adaptive control techniques for self-tuning control of a batch PVC reactor and reported better
performance compared to classical PID case. MacGregor [11] later reported that all of various
adaptive controllers appeared to be robust and superior to fixed parameter PID controllers.
An alternative approach to the problem was to investigate the control strategies employed by
human operator and design "fuzzy" logic controllers [17].
 840

To overcome the difficulty associated with sharp increase in the rate of polymerization at
reduced polymerization time, feedforward control was used by Amrehn [1] to take action in
advance by predicting the amount of polymerization heat in a given reaction stage. Later Jutan and
Uppal [9] employed a feedforward- feedback control algorithm.
"Parametric control" strategy was employed by Jutan and Rodrigez [8] for control of a batch
reactor heating system which involved expressing the manipulated variables of the process in terms
ofa new set of variables, called parametric variables. They established the parametric relationships
experimentally to satisfY the requirements. In Ponnuswamy et a1. [14]'s work, different open loop
policies for the batch reactor were derived and optimal temperature policies calculated from the
optimal feedback control law were implemented on an experimental reactor system through a PI
controller.
In recent years, "predictive control" techniques have been proposed based on a nonparametric
model, namely an impulse response or discrete convolution model [12]. Predictive control
techniques are advantageous because of usefulness for processes with unusual dynamics where
specifYing the model structure for a parametric model would be difficult as well as inherent time
delay compensation and ability to accommodate process constraints. This approach is applicable
theoretically to any system which can be described by a set of linear differential equations. Two
predictive control techniques have evoked a great deal of interest recently: Model algorithmic
control (MAC) and dynamic matrix control (DMC). Model predictive control approach and its
comparison to IMC systems is well described by Cheng and Brosilow [4] and, Hidalgo and
Brosilow [7]. N, for the polymerization systems, MPC combined with coordinated control
strategy, has been applied to a continuous styrene polymerization. They simulated the styrene
reactor and control system using the complete fourth order model for the process and the reduced
third-order system (ignoring the equation for initiator concentration) for the model and, have
demonstrated that stable control and exact tracking of the set point can be accomplished.
Most recently, Westerholt et a1. [20] described a time-optimal algorithm for second order plus
dead time systems and tested it via simulation for the startup of an exothermic batch reactor as
well as for some other systems. They reported that the algorithm had performed well on all of
physical systems tested.
This paper aims to develop a simple, easy-to-implement control algorithm and establish its
feasibility for temperature control of a batch solution polymerization reactor at the isothermal
stage.

Process Description

The process model for the free radical solution polymerization of styrene in a jacketed continuous
stirred tank reactor has been given in early investigations [3,5,7,19].
Reaction steps for free radical polymerization are represented by the following equations:

(1)
I 2R (initiator decomposition)
 841

(2)
(initiation)

(3)
(propagation)

(4)
(termination)

(termination) (5)

Following these kinetics, the model for a batch process, is composed of material balances for
initiator and monomer, and energy balances for the reactor and jacket as given by equations 6-9.
Perfect mixing, constant holdup and constant physical properties are assumed in the reactor.

d[1] = -k [1] (6)

dt d

d[M] = -k [M] [P] (7)

dt P

(7a)

dT
pc v -
P dt
= (-tili)vkP [M][P] - UA(T-T)C (8)
 842

(9)

It is known that, polymerization of certain monomers in concentrated solution is accompanied by

a marked deviation from first order kinetics in the direction of an increase in reaction rate and
molecular weight, termed "gel effect". However, benzoyl peroxide-initiated polymerization of
styrene is reported to be accurately first order up to quite high conversions [2]. In virtue of this
fact, the viscosity effects have not been included in the above model at the present stage of the
work.
In the energy balances, the assumption that for a sufficiently rapid coolant flow rate the inlet
and outlet temperatures would not vary by much holds and, therefore, an average temperature for
the cooling jacket was used. Literature values were used for process parameters [3].

Control Algorithm
The approach used here for temperature control of the batch polymerization process in general
is a simple model predictive control. MPC is explained in terms ofa sequence of tasks [4]. The
tasks define the MPC structure just as block diagrams define the structure ofIMC [6].
Task 1. Use process model and current measurement to estimate the disturbances entering the
process. If there is a process dead time, then disturbance estimates are extrapolated forward in
time by the amount of the dead time. This task explicitly estimates the disturbance so that the
model can be used to predict the future process outputs after the dead time.
Task 2. Predict the process state one dead time into the future based on current measurement,
current and past controls and extrapolated disturbances.
Task 3. Compute the desired future values ofthe process state beyond the dead time starting
with the process state predicted in Task 2. Desired future state should evolve as a dynamical
system whose relative order is the same or greater than that of the process model. Modeling errors
in MPC will often be compensated by assuming that the errors between the model and process
outputs were caused by unmeasured disturbances.
Task 4. Compute the controls which force the model to track the desired future states
computed in Task 3 as closely as possible over some time horizon into the future.
In some versions ofMPC such as Model Predictive Heuristic Control [16] and Quadratic
Dynamic Matrix Control [15], control is calculated by minimizing the integral of the square
deviation between the desired trajectory and the model trajectory over a preset time horizon
subject to control effort and process constraints.
An alternative approach, is to select the current control so that the output or the state of the
model matches that of the reference dynamical system at some future time. This is computationally
much less intensive than the previous method. We, therefore, want to use this approach here in
order to develop an easy-to-implement and effective control algorithm for the isothermal stage of
the batch polymerization reactor.
The overall objective of polymerization control can be accomplished by controlling the reactor
temperature and monomer and initiator concentrations. However, because of its large effect on
the structural characteristics of the polymers, the temperature control bears probably most
importance. This was particularly true for solution polymerization where no gel effect was present
 843

because of relatively low viscosity polymerization. Therefore, we have chosen to control the
reactor temperature by manipulating the flowrate of cooling water into the jacket. We assume that
the only disturbance to the system is the heat of polymerization, represented by the first term on
the right hand side of reactor energy balance, eqn. (8). We treat this nonlinear heat generation term
as an inferential disturbance (to be predicted from process measurements) and, denote it by D.
Thus, the energy equation becomes
dT
pc v -
P dt
=D - UA(T-Tcl_ (10)

The exothermic reaction enthalpy, 6H is a complicated function of the concentrations. If one

wishes to obtain a theoretical expression for it, the necessary detailed information is often not
available. As we do require information about 6H in order to control the process, our suggestion
that its value be predicted from on-line process measurements would be a practical and realistic
approach. Such an approach was, in fact, previously used by Jutan and Uppal [9] in their
feedforward control studies via simulation, but was not implemented in a real-time control
algorithm.
However, what we end up with is still a nonlinear equation because of the overall heat transfer
coefficient. U is the essential term in the above equation that relates the manipulated variable to
the process variable. We calculated U by evaluating the film coefficients ~ and It" for the reactor
side and jacket side respectively. For ~ the correlation given by Perry [13] for heat transfer from
agitated liquid contents of vessels to jacketed walls was used while It" was calculated for laminar
flow in annuli according to a correlation again provided by Perry [13]. The resultant equation was
highly nonlinear as follows,

U=------ (11)
336.44 + 1310.1 fc- l13

However, ifU is plotted against t;, for the range of operation in our cooling water feeder in
the experimental system, one notices that after a fiowrate of 20 cm3 /s the graph exhibits a linear
behavior as can be seen in Figure 2. As the experimental results later demonstrated, the linearized
portion of this relationship mostly fell into the actual operating range of the controller. Therefore,
we suggest that the following linear equation, thus obtained, may be used for simplicity

(12)

If we substitute this equation into equation (10), then we get a linear first order system on
which we can establish a model predictive control algorithm fairly easily. We, consequently,
suggest the following control algorithm:
1) Using measured values ofT and Tc ' solve the suggested form of the energy equation (eqn.
10 coupled with 12) for D, the disturbance estimate. For a small enough sampling time, we treat
this disturbance estimate constant.
2) Calculate the desired value of the temperature in the control horizon as follows.
 844

lE-4 lE-5
16 160

U 14
-- 140

12 120

..:..:
10

8
100

SO
I
.......
III

60
Ie
6
u
40
u '"
2 20

0 0
0 5 10 15 20 25 30 35 40 45 SO
fclcm3s-11

Figure 2. Dependency of Overall Heat Transfer Coefficient on Cooling Water Flow Rate

--H --H (13)

Td (tt +H) = e E T(tk) + (l-e E )T,

This reference trajectory was obtained by applying a first order filter on the current reactor
temperature to reach the desired value. The sampling time for the temperature measurement was
5 seconds. The control horizon was chosen two times the sampling time for stability reasons, as
suggested by Hidalgo and Brosilow [7] as a consequence of their numerical studies.
3) Fmd the value, T(ft +H), that process variable would reach ifthere were no control starting
from T(tk ) again by making use of the linearized model above. Here, we assumed that there were
no dead time.
4) Determine the step response of the system for a unit step change in control, S(H)
5) Calculate the control effort from the control law such that
Td (tt +H) - T(tk + If)
u(tJ = S(If) (14)

As the first implementation of this control calculation resulted in some oscillations, we used
a first order digital filter on this calculated control effort to stabilize the control action and smooth
the operation. The final form of the actually applied control effort, therefore, was
c{tJ = cxu(tJ + (l-cx)C(tk - At) (15)
 845

As for the filtering constant, first experiments indicated that a = 0.3 would be a good choice
in our case.
This algorithm was implemented in an interactive real-time monitoring and control software
programmed in Turbo Pascal version 5.0.

Experimental

The bench scale experimental system under investigation in this work consists mainly of a I-liter
batch stirred tank reactor with cooling jacket and necessary auxiliary equipment for on-line
microcomputer control. A diagram of the process is shown in Figure 3.
The reactor is made out of glass and surrounded by a metal jacket which is insulated by a thick
layer of g1asswool. To avoid any pressure build-up in the system, the reactor was operated under
a condenser that was fitted at the top. The control strategy developed in this paper manipulates
the flowrate of cooling water through the jacket to maintain the reactor temperature at desired
level during the isothermal, heat-generating phase of the reaction.
Prior to the experiments, styrene monomer was distilled under vacuum and stored in a
refiigerator until use. As for the initiator, pure benzoyl peroxide (B~O:J crystals freshly prepared
by crystallizing from a solution in chloroform were used. Toluene was chosen as solvent for the
reaction medium.
Experiments were started by loading the styrene monomer (6 moVl) and solvent into the
reactor. Initially, the reactor was purged with nitrogen gas and was kept under nitrogen
atmosphere until the end of the experiment. The initial temperature to start the operation was first

1---1 Converter ~--------J Thermocouple ~----:

: :-1 Converter +--------1 Thermocouple I--i i Condenser
l I I 1
I I
I

r---
I
---1 DC motor I

I
I
I
I
I
I

Cooling
vater
hnk

Figure 3. Experimental System

 846

reached by heating the contents of the reactor with an external electric heater under manual
control. This was a few degrees below the set point temperature for the reaction. The initiator
could not be used for the initial warming up period because of the fact that the reaction rate of
the initiator (BZzOz) was very low at temperatures below 70°C. This implies that an external
source for heating was needed.
In industrial applications, the polymerization of styrene is usually conducted at 70 °C where
the completion of reaction takes long and the rate of heat generation is relatively low. As was
noted on the literature survey section, the classical control algorithms work well under these
conditions. We have chosen a slightly higher set point temperature to have a higher rate of heat
generation. This would be advantageous in two aspects; (i) to test the algorithm under severe
conditions to see the performance so that it would be supposed to work under mild conditions,
(ii) to be able to decrease the time required for completion of polymerization. Thus, the set point,
in all experiments, was 90.2 °C. When the reactor reached that value, desired amount of initiator
was dissolved in toluene and was charged to the reactor. Automatic measurement, control and
data collection by microcomputer was started simultaneously. Control algorithm used the current
values of the reactor and cooling jacket temperatures, and previous value of the control. The
current control signal was then calculated and sent at each sampling interval to the DC motor
driver through a DIA converter. This driver then regulated the 12V DC motor to manipulate the
cooling water flowrate passing through. In all of the experiments, we used a cooling water supply
at 70°C.
After control experiments, dilute solution viscosity of polystyrene was measured in a capillary
viscometer of Ubbelohde type to asses the molecular properties of the polymer. The average
molecular weights calculated from the intnnslC viscosity measurements by
Mark-Houwink-Sakurada equation were around 70,000 for three experiments reported here.

Results and Discussion

The experiments were exploratory in nature with the aim of determining if temperature of a batch
polymerization reactor could be effectively controlled by the proposed simple model predictive
control algorithm. At this stage of the work, only the isothermal stationary phase of the reaction
was attempted to be controlled by microcomputer.
Our preliminary results had revealed that when a cooling water at room temperature were
used, the continuing heat transfer-between the contents of the reactor and jacket even when the
cooling water flow was zero, would bring the temperature further down below the set point,
therefore limiting the time that the reactor would have been maintained at the set point. To avoid
this, we used a coolant supply of70 0c. Results from the experiments conducted thus far under
specified conditions gave temperature tracking profiles as illustrated in Figures 4 to 6.
Figure 4 and 5 indicate two attempts to bring the reactor to the desired set point and hold it
there. In both experiments, the temperature rises as the reaction starts at the time of addition of
the initiator as described in the experimental part. During the early phase of this heat-up period,
the algorithm applies no control. However, as the temperature approaches the set point, the
control effort senses that the estimated distUibance may drive the process above the set point and,
takes corrective action to compensate for its possible future effects. The closer the reactor
temperature is to the set point, the more the control is applied. Then, consequently, with a slight
overshoot, the effect of exothermic heat generation is compensated by the control action and the
rate of increase in temperature slows down, finally settling the temperature at the desired value.
 847

400T-------------------------------------------~

380

T 360
( K) r.r\... '-._._._"--./. /'. .......-A./.-._._._._ - . _ . _ . _ .
340 I 100

u (%)
320~~-~~-A~~~--~--~~---- ________~~__~
0.00 Time ( min :secl 20: 54

- Control - - Reactor Temperature

-.- Jacket Temperature ---- Set Point

Figure 4. Control Experiment, Run I, [Io =0.084 molJl, 1\:6 molJl, TCo=343 K]

Meanwhile, the control action also decreases and comes down to zero. Filter time constant (s) in
these two experiments was 100 seconds.
Figure 6 shows the closed loop behavior of the control system when a filter time constant of
200 seconds was used. In that case, the control system takes a rather sluggish action to bring the

400~--------------------------------------------__r

380

T
(Kl 360 ..........
'-._.-._._._._._._._._.-._._._._._._.
340 100

320~~----~~~--~~--~~~~------~---A~~~

0.00 Time (min:sec) 20:54

--- Control - Reador Temperature

-.- J ac ket Temperature ---.. Set Point

Figure 5. Control Experiment, Run 2, [10 =0.084 molJl, M,,= 6 molJl, TCo=343 K]
 848

temperature to the set point. The applied control effort is remarkably less compared to the case
observed in experiments shown in Figures 4 and 5. When filter time constant was small, immediate
corrective action was taken to bring the process rapidly to the set point. Large s on the other hand,
results in rather small changes in the control effort as reflected in Figure 6.

400,-----------------------------------------------,

380

,-------------------------------
T
(K)
360
_......../_.-._._.-....-._._.-.
~.-.-.-.-.-.-.--

340 100
u (%)

320~--------~---O~~~~~~~~----------~~~

0.00 20:54
Time (min:secl

- - Control - - Reactor Temperature

-.- Jacket Temperature ---- Set Point

Figure 6. Control Experiment, RWl3, [10=0.084 molll, M.,= 6 molll, TCo=343 K)

We suppose that the linearized relationship, between the overall heat transfer coefficient and
the coolant flow rate, as illustrated in Figure 2, can be assumed to hold down to a flow rate of 15
mils. It is, therefore, to be noted that, the control effort, i. e. actual operating range of the coolant
flowrate falls mostly within this linearized region (i.e. 30 to 100%) that defined by eqn. (12).
From the above discussions, we can conclude that the proposed algorithm holds promise for
further investigation and for practical applications on a larger scale, because it provides the ability
to minimize the effects of disturbances.
The suggested algorithm offers two advantages. First, it is simple to design so that it can be
easily implemented in a microcomputer without resorting to excessive computation. Second, it has
only one tuning parameter, namely E, that can be either theoretically determined for simple linear
first order systems or can be found experimentally. This is obviously an important superiority over
the conventional, three parameter PID control. No need to mention that, in addition, the inherent
advantages ofMPC, like minimization of the effect of control effort saturation, also hold.
Besides, the simple methodology we suggest here can be applied to the temperature control
of similar exothermic batch reaction systems, because we do not actually use kinetic model for
control, we rather make an online inferential measurement to sense the instantaneous reaction
conditions. Therefore, the algorithm can be expected to work even if changes in kinetics occur at
high conversions. In different polymerization conditions, however, one should take the fact into
account that controlling the reactor temperature may not be suitable to control the polymer
properties which is the ultimate goal of the control problem in polymerization systems. In the next
phase ofthls work, we would like to take the nonlinearities of the system into consideration and
use the kinetic model to calculate the control effort.
The difficulties associated with the use of low temperature coolant that we had observed
 849

during our preliminary experiments indicated the need for a heating medium computationally
coupled with this algorithm. In this case, if the temperature at the jacket side dropped sharply with
applied control effort leading to a decrease in the reactor temperature below the set point, heat
could be applied to bring the reactor back to set point. Such a medium could have also been used
for the initial heating up of the reactor to the desired reaction temperature.

Conclusion

We can conclude, from the exploratory studies reported here, that the suggested simple MPC
control algorithm can be useful in applications particularly where an easy-to-implement and
practical control algorithm is needed.
However, the more usual solution to the problem would be to live with the process
nonlinearities and use more complex algorithms such as computation of the control by a Newton
algorithm while considering the full model without any simplification. We believe that this would
lead to a more robust algorithm. In this case, one would be able to use both heating and cooling
modes by probably manipUlating the external heat input into the system together with the cooling
flowrate. Therefore, a predetermined optimal time-temperature trajectory for the whole sequence
of a batch operation would be controlled.

Acknowledgment: This work was supported by a research grant from Ankara University research
fund under contract number 91-25-00-93 which is gratefully acknowledged. We also acknowledge
fruitful discussions with Prof. C. B. Brosilow throughout this work.

Notation

A : heat transfer area of reactor, cm2

Cp : mean heat capacity, cal g.' K"I
C (tk - at) : actual control applied at previous sampling period
f : initiator efficiency
fe : flowrate of cooling jacket fluid, s·'
H : control horizon
I : initiator
kci.k;, ~, lew, kid : Arrhenius rate constants for dissociation, initiation, propagation, combination termination and
disproportionation termination respectively
M : monomer
Pk : growing species oflength k units
R : initiated radical
T : temperature of reactor, K
Te : mixing cup temperature of cooling jacket fluid, K
Ts : temperature set point, K
t : time, s
tk : current time, s
U : overall heat transfer coefficient, cal cm·2 s·' K-'
u : control effort
v : reactor volume, I
[] : represents concentration
o : represents initial conditions
 850

Greek Letters

P : mean density of reactor fluid, g 1-'

Pc : density of cooling jacket fluid, g 1-'
a : a constant such that a = I corresponds to no filtering while approaching zero means ignorance of
measurement
-aH : heat of polymerization reaction, cal mol-'
at : sampling time, s
E : adjustable tuning parameter

References

I. Amrehn. H.: Computer Control in Polymerization Industry. Automatica, 13. 533?-545 (1977)
2. Billmeyer. F. w.: Textbook of Polymer Science. 2nd ed. Wiley- Interscience, New York (1971)
3. Brooks, 8.W.: Dynamic Behavior of a Continuous Flow Polymerization Reactor. Chem. Eng. Sci. 36. 589-593
(1981)
4. Cheng. C.M and Brosilow. C.8.: Model Predictive Control of Unstable Processes. AlChE 1987 Annual Meeting.
New York (1987)
5. Chen, SA and Jeng. W.F.: Minimum End Time Policies for Batchwise Radical Chain Polymerization. Chem. Eng.
Sci. 33. 735-743 (1978)
6. Garcia C. and Morari. M: Internal Model Control I. A Unifying Review and Some New Results. Ind. Eng. Chem.
Process Des. Dev. 21, 308-323 (1982)
7. Hidalgo. P.M and Brosilow. C.B.: Nonlinear Model Predictive Control of Styrene Polymerization at Unstable
Operating Points. Computers chem. Engng. 14,481-494 (1990)
8. Jutan. A and Rodrigez. E.S.: Application of Parametric Control Concepts to Decoupler Design for a Batch Reactor.
Can. 1. Chem. Eng. 65, 858-866 (1987)
9. Jutan, A and Uppal, A: Combined Feedforward-Feedback Servo Control Scheme for an Exothermic Batch
Reactor. Ind. Eng. Chem. Proc. Des. Dev. 23. 597-502 (1984)
10. Kiparissides. C. and Shah. S.L.: Self-tuning and Stable Adaptive Control of a Batch Polymerization Reactor,
Automatica, 19,225-235 (1983)
II. MacGregor, 1.F.. PenJidis, A.. Hamielec. A.E.: Control of Polymerization Reactors: A Review. Polym. Proc. Eng.
2. 179-206 (1984)
12. Marchelti. 1.L.; Mellichamp, DA; Seborg, D.E.: Predictive Control Based on Discrete Convolution Models, Ind.
Eng. Chern. Proc. Des. Dev. 22,488-495 (1983)
13. Perry. R.H. and Chilton. C.H. (editors): Chemical Engineers' Handbook, 5th ed. pp. 10-13,10-16, McGraw Hill
Inc. N.Y. (1974)
14. Ponnuswamy, S.R.; Shah, S.L. and Kiparissides, C.A: Computer Optimal Control of Batch Polymerization
Reactors, Ind. Eng. Chern. Res. 26,2229-2236 (1987)
15. Prett. D.M. and Garcia, C.E.: Fundamental Process Control. Butterworths. Boston (1988)
16. Richalet. 1.. Rault, A, Testud, 1.L. and Papon, 1.: Model Predictive Heuristic Control: Applications to Industrial
Processes. Automatica, 14,413-428 (1978)
17. Rong, G.: Microcomputer Control of a Pilot Batch Polymerization Reactor and Investigation of Fuzzy Control
Algorithms, Master Thesis, Khejiang University, China (1986)
18. Sacks. ME.; Lee. S. and Biesenberger, JA: Effect ofTemperature Variations on Molecular Weight Distributions:
-Batch Chain Addition Polymerizations, Chern. Eng. Sci. 28. 241- 257 (1973)
19. Villermaux, 1. and Blavier, L.: A New Method for Modeling Free Radical Homogenous Polymerization Reactions.
Chern. Eng. Sci. 39. 87-99 (1984)
20. Westerholt, E. von; Beard, IN.; Melsheimer. S.S.: Time- Optimal Startup Control Algorithm for Batch Processes,
Ind. Eng. Chem. Res. 30.1205-1212 (1991)
Retrofit Design and Energy Integration of Brewery
Operations

Denis 1. Mignon

Belgian National Fund for Scientific Research (F.N.R.S.), Rue dEgmont 5, B-1050 Brussels and
Process Engineering Department, Universite Catholique de Louvain, B-1348 Louvain-Ia-Neuve
BELGnJM

Abstract: A detailed simulation model of the four brewhouses of an industrial brewery has
been developed with the help of the BATCHES simulator. The use of this model has allowed a
thorough study of the utilization rates of the pieces of equipment. The study has shown that
these utilization rates are generally quite low and that some pieces of equipment are largely
oversized. Two new configurations have been proposed for the brewhouses, reducing by up
to one third the number of pieces of equipment.

On the side of the energy integration, the use of the model has shown that the combination
of the proposed modifications of the process, namely a better standardization of the brewing
plant, an adequate production planning and a deliberate limitation of the steam availability level,
permits a considerable reduction of the steam consumption peaks. From the most unfavourable
situation to the most favourable one encountered during this simulation study, the global
reduction amounts to 55%.

Keywords: BATCHES / batch processes / brewery / brewing process / energy integration /

optimization I retrofitting I simulation.

1. Using BA TeRES for the Simulation of the Brewing Process

While following the general objective of modelling and optimizing the brewing process, we
have developed a detailed simulation model of the brewhouses of a test brewery having a
capacity of two million hectolitres per year.

In its actual configuration, the "hot part" of this brewery consists of four brew houses
grouped two by two. The brewhouses of group A are equipped with classical filter-presses
 852

whereas the brewhouses of group B use the new mash filter 2001. Each group of two
brewhouses comprises only one pair of mills and one chain of wort treatment down-stream of
the boiling vessels. This configuration is represented in Fig. 1. As the four brewhouses have
not been built at the same time and do not consist of the same pieces of equipment, they also do
not work with the same load and with the same cycle time.

Realized on the basis of BATOIES [1-3], the simulation model now allows us to follow
the consumption of the raw materials, the utilization of utilities, the status of all pieces of
equipment, and so on. It can also be used to study the bottlenecks of the process or the
influence of the production planning and of the synchronization of the four brewhouses on the
parameters previously cited. Finally, it could be used to look at the way any disturbance, an
equipment failure or a lack of operators for example, will spread into the whole process, in
order do determine the time available to the operators to restore the system back to normal
operating conditions before being obliged to stop the production completely.

Fig. 1. Present configuration of thebrewhouses

More specifically, the use of the model has allowed a thorough study of the utilization rates
of the pieces of equipment of the brewhouses as well as an analysis of the influence of the
production planning on the energy consumption profiles.
 853

2. Study of the Utilization Rates of the Pieces of Equipment

The realization of a series of simulations of the functioning of the test brewery brewhouses has
allowed us to study the utilization rate of their equipment. It clearly appears, as shown in the
left part of Table I, that these utilization rates are generally quite low. This is mainly the case
for the mills, the buffer tanks, the mash coppers, the homogenization tanks upstream of the
centrifuges and, to a lesser extent, for the mashftlters.

Utilization rate (%) of the equipment

in the case of configuration
Name of the piece of present present modified #1 modified #2
equipment group A groupB
Malt mill 39.3 5U.5 79.1 tll.7
Adjuncts mill 9.8 13.4 20.8 21.5
Malt flour hopper #1 100.0 100.0 100.0 100.0
Malt flour hopper #2 100.0 100.0 100.0 100.0
Adjuncts flour hopper #1 100.0 100.0 100.0 100.0
Adjuncts flour hopper #2 100.0 100.0 --- ---
Mash mixer # 1 64.5 68.1 56.0 58.5
Mash mixer #2 64.5 69.0 56.2 58.6
Mash copper #1 30.5 32.7 55.0 56.8
Mash copper #2 30.4 34.2 --- ---
Mash ftlter # 1 42.0 46.6 81.3 85.2
Mash filter #2 42.0 47.6 --- ---
Buffer tank #1 18.2 4.7 --- ---
Buffer tank #2 18.4 4.8 --- ---
Boiling copper #1 69.2 91.8 78.5 81.1
Boiling copper #2 69.2 91.9 79.7 81.5

Hopstar 20.3 12.4 19.4 10.0

Homogenization tank 47.8 39.3 67.1 31.4
Centrifuges 51.8 46.7 72.5 38.3
Cooler 51.8 46.7 72.5 38.3

Table 1. Equipment utilization statistics: comparison of the modified

configurations to the present configuration

This study has also shown that some pieces of equipment are largely oversized. This is the
case for the buffer tanks located between the mash filters and the boiling coppers, as well as for
the homogenization tanks.

Taking into account all these observations, we have proposed two new configurations for
the brewhouses, quite different from those in operation. In order to be as rational as possible,
 854

these configurations consist of four identical brewhouses and should be operated with a
constant load value. Such a restriction could be considered as an important limitation of the
operating flexibility of the industrial plant. However, it represents a totally acceptable
constraint for large dedicated breweries like the one that has been dealt with in our study. As
illustrated by Fig. 2 and summarized in Table 2, the suggested configurations represent a
notable reduction of the number of pieces of equipment required. This reduction amounts to
one third in the most favourable case. As can be easily noticed, the only difference between
both modified configurations is the presence of one chain of wort treatment downstream of the
boiling vessels in the case of the flI'St one, whereas the second one comprises two such chains,
which results in a larger operating flexibility.

!lIS SI! t6 IQUIPIa! IS'U! or IIIJII'IID

c:owICIIIU'IDI U OIILJ'. D t-. CUI ar aD'1ID
a.tCDIUIOI 11. w. ...s IICII!II""
IOILIW: CCftIIS .US IIRCDCiI ICIl1' ~ 11

_....
--
Fig. 2. Modified configurations of !he brewhouses

A series of simulations have been performed on the basis of these two modified
configurations. They have proven their feasibility, provided that a strict and quite rigid control
of the production planning be enforced in the case of the first one, whereas the second one has
the same operating flexibility as the present configuration. At the present time, two new
breweries whose lay-outs are based on this second modified configuration are being put into
operation in Belgium. The main difference between the configuration that we have proposed
and the one adopted for these new plants is that they are composed of three brewing groups in
parallel, instead of two. The other differences are just design details.
 855

The right part of Table 1 shows the utilization rates of the equipment of both modified
configurations. It illustrates very well the obtainable gain in comparison with the basic confi-
guration Oeft part of the table). The absence of figure in the right part of the table means that
these particular pieces of equipment are no longer included in the proposed new configurations.

Type of equipment Present Modified configurations

:
configuration #1 #2
malt & adjuncts mills
mash coppers ~ ~
adjuncts flour hoppers 4 2 2
mash filters 4 2 2
buffer tanks 4 0 0
(before boiling)
wort treatment chains 2 1 2
(after boiling)
TOTAL NUMBER OF 42 25 30
PIECES OF EQUIPMENT

Table 2. Number of pieces of equipment of the brew houses

3. Energy Integration

The simulations that have been performed have allowed the determination of the time profiles of
the process utilities consumption and, in particular, of the steam consumption. It has been
possible to show that these profiles exhibit some peak demands of very short duration which
greatly exceed the mean demand. Figure 3 shows an example of such a profile.

10000

aooo
~
<!!-
~ I
,l,llll
6000
Ie
1!i
'"0Z 4000 ~
\,) ~II'I"'I 11U,IUI

~ 2000

'"
1000 2000 3000 4000 SOO
TIME (min)

Fig. 3. Steam consumption profile for the realization of 15 brews in each of the four
brewhouses. Cycle time: 145 min. Offset between brewing groups: 0 min.
 856

3.1. Standardization of the brewhouses

The standardization of the brewhouses that can be reached by adopting one of the modified
configurations gives rise to a reduction of some 20% of the maximum peak value of the steam
demand. Indeed, by using the four brewhouses with constant values for the load and for the
cycle time, one avoids having an offset between the brewhouses that varies with time, thus
avoiding the steam consumption profUe exhibiting a quasi unforeseeable behaviour.

3.2. Influence of the production planning

It also appears that the production planning exerts an important influence on the steam demand
profile. Indeed, given the design of the brewhouses (one single wort treatment downstream of
the boiling operation per group of two brewhouses), the two brewhouses of a same group are
necessarily synchronized. On the other hand, in the case of the present configuration and of the
second modified configuration, there is no synchronization constraint between the two brewing
groups. In such a context, it was opportune to study the influence of the offset between these
two groups on the various parameters of the steam consumption. Table 3 presents the results
of the simulation runs for different values of the offset between the two brewing groups. It
shows that an adequate offset of the two brewing groups starting times gives rise to a
considerable reduction, once again about 20%, of the maximum level of the steam peak
demands.

Ly~letime Steam availability Offset between the Highest steam

[min] level brewing groups consumption level
[leW] [min] [leW]
145 infinite 0 9500
145 infinite 10 8756
145 infinite 20 7583
145 infinite 30 8223
145 infinite 40 8273
145 infinite 50 7547
145 infinite 60 9500
145 infinite 70 8419

Table 3. Influence of the offset between brewing groups on the peak

values of the steam consumption profile

However, the approach used so far is not satisfactory. Indeed, if we wanted to study the
steam consumption profile for all the possible values of the offset, we would have to make as
 857

many simulation runs as there are such values. Even on a fast computing platform. such a
study would be extremely time-consuming. On the other hand. if we do not consider all the
possible offset values. we may miss the real optimal value which. it is quite obvious and was
proved by the simulation study. can not be predicted by any means.

For this reason. we have developed a systematic approach of the problem which still rests
on the use of the simulation model. but not so heavily. This approach can be split up into the
following stages:

a) A simulation run is used to generate the steam consumption proflle for 50 consecutive
brews in the same brewhouse. These brews are processed with such a cycle time that their
respective consumption profiles do not interfere with each other and do not overlap. Figure
4.a presents a part (10 brews) of the resulting proflle. We then use these data to feed
computer routines that are independent from the simulator. These routines. written in C
programming language, execute the next four steps (b to e) of the method.

b) On the basis of the 50 individual proflles generated by the simulation run, a mean profi Ie is
computed. If we call P(t) the profile for n consecutive brews generated with a cycle time
value p, the mean profile M(t) is given by the formula

n
L P[t + (i-1).p]
M(t) =.:..i_=~1_~_ _ "V t E [O,p[ (1)
n

whose application requires that the suitable values of P(t) be computed by interpolation
since P(t) is defined in a discrete way. The mean profile for one individual brew that has
been computed by applying formula 1 is shown at Fig. 4.b.

c) The mean profile is then used to compute the resulting profile S(t) of consecutive
overlapping brews in one brewhouse, as defined by the formula

S(t) = M(t) + M(t-n (2)

where T is the brewing cycle time. The periodic part of S(t) is represented at Fig. 4.c.

d) We can then generate the steam consumption profile F(t) of one brewing group consisting
of two brewhouses with an offset D, following the formula

F(t) = S(t) + S(t-D) (3)

 858

~
~
.000

2500
~--.-----,.--~--.,.-----,
..•••• _ •• _- _... . •• _.+. . . _ . ~

~
l 0 0 0 r - -_ _ r----,.--~--...,

n!

:. . . . . 2 . .
~ 2000 ~ 2000

VJ 1500 +ll'H-Hl++R+1~IfH-Ifll~HfI+ffH-t/'H

8 ~
... 1000

~ o~~~~~wu~~~~~~
o 1000 2000 lQOO 100 200 100 400
TIME (min) TIME (min)

~
~ 4000t--_ _...,...,_ _ _---.._ _ _- ,
~ 4000

i8 : :
Z z
1000

~
~
u
r-1
" ttt-t--t-;--t--t--ft''Tti
8 1000 Inn 1000 t----;---I--t-jf--ti
::£

e 11111
100 200
t .00
TIME (min) TIME (min)

Fig. 4. a) Steam consumption profile of consecutive. non-overlapping brews in one

brewhouse. ProfIle generated by a simulation run with BATCHES.
b) Computed mean steam consumption profIle of one isolated brew.
c) Computed periodic part of the steam consumption profIle of consecutive.
overlapping brews in one brewhouse. Cycle time =290 min.
<0 Computed periodic part of the steam consumption profIle of a brewing group
consisting of two brewhouses operating in parallel. Offset = 145 min.

The periodic part of F(t) is represented at Fig. 4.d in the particular case where D = T I 2
which means that the two brewhouses are perfectly synchronized. Moreover. this would be
the normal operating way for a configuration such as the one studied here ..

e) The global steam consumption profiles of the brewing plant consisting of two (or three)
groups operating in parallel with an offset X (or with offsets X and Y ) are given by

R2.X (t) = F(t) + F(t-X) II= R2 (t,X) (4)

II
R3.x.Y (t) = F(t) + F(t-X) + F(t-Y) = R3 (t,X,Y) (4')

and the maximum consumption peak, as a function of the offset(s) between the brewing
groups, is defined by the formulae

max [R2 (t,X)] ~ MAX (X) (5)

t
max [R3 (t,X,Y)] ~ MAX (X,Y) (5')
t
 859

~800)"T""'-....,....---r--"--"'T""---'T--r--"'T'""-.....,..--r-"'"']
~

~
~ 700) 4--4-4__L-4-~-4----4----4----4_--_44---~~--~-4
E
~
~
8 ~4---~----~--~4M~~-4~~--~~--~--~----+--4
~
'><"
-<
'" 500)4----4----4----4----4---~----4_--_4~--~----~~
o 15 30 45 60 75 90 lOS 120 135

OFFSET BETWEEN GROUPS 1 AND 1 (min)

GROUPS I & 2 (min) GROUPS 1 & 3 (min)

Fig. 5. Influence of the offset(s) between brewing groups on the maximum

consumption peak. a) Two brewing groups in parallel.
b) Three brewing groups in parallel

Figure 5 represents the function MAX (X) and the response surface MAX (X,Y) whereas
Table 4 summarizes the most important information of these two diagrams, namely the
highest and the lowest values of the maximum consumption peak, as well as the
 860

corresponding offsets between the brewing groups. As can be also seen, the difference
between the situations with the most unfavourable offset and with the most favourable one
amounts to 35% for two brewing groups, and reaches 41 % in the case of three groups.

minimum maximum % difference

maximum
consumption peak 5182 7969 35
2 BREWING [leW]
GROUPS offset between
groups I and 2 72.5 0 --
[min]
maxunum
consumption peak 7007 11953 41
[leW]
3 BREWING onset between
GROUPS
groups 1 and 2 48 0 --
[min]
offset between
groups 1 and 3 96 0 --
[min]

Table 4. Highest and lowest values of the maximum peak of the global steam
consumption profile of a brewing plant composed of two or three
brewing groups, and corresponding offsets between these groups.

f) In order to check the validity of the optimum offset computed above in the case of two
brewing groups (72.5 min), we called upon the simulator once again. In this instance the
simulated highest consumption peak reaches 8345 kW, a value that is much higher than the
one predicted. Obviously, this difference is due to some parameters of the simulation
model whose stochastic nature could not be taken into account in our method. However, as
will be shown in the next paragraph, the information that we have obtained through its
application will prove very useful since it will be possible to come very near to the
computed optimum.

3.3. Deliberate limitation of the steam availability level

Actually, a statistical analysis of the steam consumption in the most favourable case computed
by the method described above (second modified configuration; cycle time: 145 min; offset:
72.5 min; computed steam peak demand: 5182 kW; real steam peak demand: 8345 kW)
reveals, as illustrated by Fig. 6.b, that only 1.7% of the total steam consumption consists of
 861

demand peaks higher than 6000 kW, and only 3.7% of peaks higher than 5500 kW. As a
consequence, we have studied the behaviour of the process when the steam availability level is
voluntarily limited to 6000 kW, or even to 5500 kW, all other parameters being the same as in
the fonner simulation runs.

Figures 6.c and 6.d show the steam consumption profile and its statistical analysis in terms
of power classes for simulation runs with the most severe of these two constraints. The
analysis of the results of these runs shows that a more regular steam consumption profile has
been obtained, as could be expected, but also that the additional waiting times generated by the

~ ~-·-~··-~·--···-·+-+-'·~·--·+·+I~·+·+ri

145O l,-:-t1fH~
~
TIME (min) LEVEL OF STEAM CONSUMPTION (kW)

Ei.1..M

Ii"
6 ~I--.'--.-!-.--.-.+--'.--'.+-.-- ... +--.-.-.i

~
!
~
~
TIME (min)
LEVEL OF STEAM CONSVMPnON (kW)

Fig. 6. Influence of the availability level of steam on its consumption profile. Cycle
time: 145 min. Offset: 72.5 min. (6.a and 6.b: consumption profile and
statistical analysis for infinite availability level; 6.c and 6.d: consumption
profile and statistical analysis for availability level limited to 5500 kW)

new constraint are almost negligible. They can be completely neglected in the case of a steam
availability limited to 6000 kW whereas, in the case of a limitation to 5500 kW, they are of the
same order of magnitude as the stochastic variations of the operation times of the tasks they
 862

affect. Even in this last case, they are acceptable since they do not cumulate, thus lengthening
the overall cycle time only in a small extent and leaving unaffected the qUality of the end
product. These supplementary waiting times are summarized in Table 5, together with the
values of the maximum demand peaks obtained during these last simulation runs.

Steam Highest steani MEAN ADDmONAL WAlTING TIME AT

availability consumption [min]
levellkWl levellkWl Mash mixing Mash cooking Filtration Wort boiling
infinite 8345 -- -- - --
6000 5986 0.1 0.0 0.8 1.2
5500 5296 6.6 4.8 3.3 3.7

Table 5. Values of the maximum peak of the global steam consumption prome
of a brewing plant composed of two brewing groups, with a limited
steam availability, and corresponding mean additional waiting times
generated by this conslrainL (Cycle time: 145 min; Offset: 72.5 min)

Globally, the combination of the proposed modifications of the process, namely a better
standardization of the brewing plant, an adequate production planning and a voluntary limi-
tation of the steam availability level, allows us to reach a considerable reduction of the steam
consumption peaks. From the most unfavourable situation to the most favourable one encoun-
tered during our simulation study, the global reduction amounts to 55%.

As the size of the utilities production plants is directly proportional to the value of the
highest demand peak which has to be satisfied by these plants, it is essential to try to minimize
this value as early as possible during the design stage of a new industrial plant in order to
reduce its investment and operating costs. The use of a simulation model such as the one that
has been developed with BATCHES has proved very useful in following this objective.

References-

1. Batch Process Technologies:BATCHES Reference Manual. Batch Process Technologies Inc.,

West-Lafayette, USA 1989-1990
2. Batch Process Technologies: BATCHES Use(s Manual. BalCh Process Technologies Inc.,
West-Lafayette, USA 1989-1990
3. Clark, S.M. and Kuriyan, K.: BATCHES - Simulation software for managing semicontinoous and batch
processes. Batch Process Technologies Inc., West-Lafayette, USA April 1989 (unpublished)
 List of Participants

NATO Countries

Belgium
Christine Bernot, Chemical Eng. Dep., 0. of Massachusetts, Amherst, MA USA, Presently at Solvay,
Belgium.
Denis J. Mignon, Unite des Procedes - Universite Catholique de Louvain 1, Voie Minckelers B-1348
Louvain-Ia-neuve, Belgium
Canada
John F. MacGregor, McMaster University, Chemical Eng. Dept., Hamilton, Ontario L8S 4L7, Canada

Germany
Henner Schmitt, LSGC ENSIC BP 45154001, Nancy Cedex, France
Greece
Christos Georgakis, Lehigh 0., Chemical Engineering Department., Bethlehem, PA 18015, USA
Savoula Papageorgaki, Purdue University, 50 S. Meridith Av. Apt. 1, Pasadena, CA 91106, USA
Lazarus Papageorgio, University of London (Imperial College), Dept. of ChE & Chern. Tech., Prince
Consort Road, London, SW7 2BY, England
Athanasios G. Tsirukis, California lnst. of Technology, Chemical Eng. 210-41, Pasadena, CA 91125, USA
Vasilius Voudouris, Dept. Chern. Eng., Carnegie Mellon Univ., Pittsburgh, PA 15213, USA

Italy
Sandro Macchietto, University of London (Imperial College), Dept. of ChE & Chern. Tech., Prince Consort
Road, London, SW7 2BY, England
France
Michel Lucet, Associate Director, Rhone-Poulenc Industrialisation, Centre DE Decines 24, Avenue Jean
Jaures, 69151 Decines Charpieu Cedex, B.p.16, France
Norway
Kristian M Lien, Universitetet I Trondheim, Norges Tekniske Hogskole, Institutt for Kjemiteknikk, N-7034
Trondheim, Norway
Eva Sorensen, Dept. of Chemical Eng., University of Trondheim-NTH, Sem Salandsvei 4, N-7034
Trondheim, Norway
Portugal
Ana Cristina Santos Amaro, I.S.C.A.C. (Polytechnic of Coimbra), Rua Luis De Camoes, 3000 Coimbra,
Portugal
JoseAlmiro Castro, Universidade De Coimbra, Departamento de Engenharia Quimica, Largo Marques de
Pombal, 3000 Coimbra, Portugal
M Lucelinda Alcantara da Cunha, Dcpartamcnto de Cicncia dos Materials, FCTIUNL Quinta Da Torre,
2825 Montc De Caparica, Portugal
 864

Paulo Falcao Ferreira, Dept. Chern. Eng., Univ. of Coimbra, Rua Cidade de Salamanca 93, 3000 Coimbra,
Portugal
Jorge MS.S. Martills, Escola Superior De Tecnologia De Viseu, Estr. Circunvalao, 3500 Viseu, Portugal
Alirio Rodriguez, Universidad de Porto, Departamento de Engenharia Quimica, Faculdade de Engenharia,
Rua dos Bragas, 4099 Porto Codex, Portugal

Spain
Estollia Espulla, Universitat Politecnica De Katalunya, Departament d'engineria Quimica, ETSEIB,
Diagonal, 647, E-08028 Barcelona, Spain
Raimoll Grav, Universitat Politecnica De Katalunya, Departament d'engineria Quimica ETSEIB, Diagonal,
647, E-08028 Barcelona, Spain
Luis Puigjaller, Univcrsitat Politecnica De Katalunya, Departament d'engineria Quimica, ETSEIB, Diagonal,
647, E-08028 Barcelona, Spain

Turkey
Ugur Akmall, Bogazi~i V, Chemical Engineering Dep., 80815 Bebek, Istanbul, Tiirkiye
Rldvall Berber, V of Ankara, Kimya Miihendisligi B6liimii, Tandogan Ankara 06100,Tiirkiye
Esell Bolat, ytldlz Universitesi, Kimya Miihendisligi B6liimii, ~i~1i Istanbul, Tiirkiye
Vildall Di/1l;ba~, Bogazi~i V, Chemical Engineering Dep., 80815 Bebek, Istanbul, Tiirkiye
Turker GUrkall, METU Chern. Eng. Dept., 06531 Ankara, Tiirkiye
Oller Hortapu, Bogazi~i V, Chemical Engineering Dep., 80815 Bebek, Istanbul, Tiirkiye
Ersall Kalafatoiflu, Marmara Ara~tIrma Merkezi, Kimya Miihendisligi Ar~tIrma B6liimii, PK 21, Gebze
Kocaeli, Tiirkiye
Ali Karadllmall, V of Ankara, Kimya Miihendisligi B6liimii, Tandogan Ankara 06100,Tiirkiye
Derya Kibar, Petkim Ar~lIrma Merkezi, Kocacli, Tiirkiye
Bilgill Klsakurek, Chemical Eng. Dep., METU., 06531 Ankara, Tiirkiye
hsell Ollsall, Bogazi~i V, Chemical Engineering Dep., 80815 Bebek, Istanbul, Tiirkiye
Nurall Drs, Marmara Ar~tIrma Merkezi, Kimya Miihendisligi Ar~tIrma B6liimii, PK 21, Gebze Kocaeli,
Tiirkiye
Callall Ozgell, Chem. Eng. Dept., METU 06531 Ankara, Tiirkiye
Calldall Tamerler, Bogazi~i V, Chemical Engineering Dep., 80815 Bebek, Istanbul, Tiirkiye

Ertall Ta~kJII, Chemical Eng. Dep., METV, 06531 Ankara, Tiirkiye

Metill Turkay, Chemical Eng. Dep., METU., 06531 Ankara, Tiirkiye

United Kingdom
Kamal Kllriyall, University of London (Imperial College), Dept. of ChE & Chem. Tech., Prince Consort
Road, London, SW7 2BY, England
Phillip Law, ETH Technisch-Chemisches Lab., Systems group, CH-8092 Zurich, Switzerland
Jack W. POlltOIl, Chemical Engineering Department,University of Edinburgh, Edinburgh, Scotland
David W. T. Rippill, ETH, Technisch-Chemisches Lab., Systems group CH-SOn, Zurich, Switzerland
 865

Rodger.W.n. Sargent, University of London (Imperial College), Dept. Qf ChE & Chern. Tech., Prince
Consort Road, London, SW7 2BY, England
Nilay Shah, University of London (Imperial College) Dept. of ChE & Chern. Tech., Prince Consort Road,
London, SW7 2BY, England
Zhang Xueya, University of London (Imperial College), Dept. of ChE & Chern. Tech. Prince Consort Road,
London, SW7 2BY, England
United States of America
Mukul Agarwal, ETH, Technisch-Chemisches Lab. Systems group, CH-8092 Zurich, Switzerland
Yaman Arkun, Georgia Institute of Tech., School of Chemical Eng., Atlanta, GA 30332-0100, USA
Ali {:mar, lllinois Institute ofTechnology, Chemical Engineering Department, lOW. 33rd St., Perlstein Hall,
Rm. 105, Chicago, IL 60616, USA
Luis Garcia-Rubio, Chemical Engineering Dep., U. of South Florida, Tampa F133620, USA
Alicia Garcia, Civil Engineering Department, University of South Florida, Tarnpa FI 33620, USA
Ignacio E. Grossmann, Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh,
Pennsylvania 15213-3890, USA
Muzaffer Kapanogtu, Industrial Eng. Dep., University of South Florida, Tampa FI 33620, USA
Girish Joglekar, Batch Proeess Technologies, 1291 E. Cumberland Avenue, West Lafayette, IN 47906, USA
Joseph F. Pekny, Purdue University, School of Chemical Eng., West Lafayette, IN 47907, USA
Gintaras V. Reklaitis, Purdue U., School of Chemical Eng., West Lafayette, IN 47907, USA
Aydm K. Sunol, Chemical Engineering Dep., University of South Florida, Tarnpa Fl 33620 USA
Venkat Venkatasubramanian, Purdue U., School of Chemical Eng. West Lafayette, IN 47907, USA
Arthur W. Westerberg, Carnegie-Mellon U., Chemical Eng. Dep. Pittsburgh, PA 15213, USA
Mike Zentner, Purdue University, School of Chemical Eng., West Lafayette, IN 47907, USA

Non NATO Countries

Czechoslovakia
Miroslav Hofmeister, Technisch-Chemisches Lab., Systems group, ETH, CH-8092 Zurich, Switzerland
Hungary
Bela Csukas, Dept. of Chern. Eng. Cybernetics, U. ofVeszprem, H-8201 POB 158, Hungary
Gyula Kortve/yessy, Research Manager, SZEVIKI R&D Institute, Stahly Str. 13, Budapest, H-I085 Hungary
Laszlo Halasz, Technisch-Chemisches Lab., Systems group, ETH-Zentrum, CH-8092 Zurich, Switzerland
Japan
Shinji Hasebe, Kyoto University, Dep. of Chemical Engineering, Process Systems Engineering Group,
SaJ..:yo-ku, Kyoto, 606-01, Japan
Sweden
Dag E. Ravemark, Teehnisch-Chemisches Laboratorium, Systems group, ETH 8092 Zurich, Switzerland
Index

A* 533-538 distributed computing 393, 412-415

algebraic/differential equations 38-41, dynamic simulation 174, 175, 189-212
331-359,361,365,366,379,386 energy (heat) integration 20, 27, 28,851,
algorithm design 393, 407-409 854-862
algorithms adversary 20, 37 estimation 295-307
ACORN-D 574-590 fault detection 20, 29-32, 35, 242, 251-257,
artificial intelligence 517-529, 595-600, 631,634,635,654-658
615-629,631,706,709,710,716-722, flexibility 41, 48,86-89,495-515
726-730 flexibility index 497, 515
baker's yeast plant 821,823-829 flexible scheduling 20, 24, 25, 40, 41, 64-69
batch design I, 10-15,20,23-28,41,74,75, frames 554-560, 716, 725, 726, 728, 733
86-108,114-119,151-155,160-167,451, fuzzy membership 645-647
478,479,482-488,495,496,510-515,595, GAMS 114, 116, 133-143, 145
608-613,750-752,762,768,776 Gantt charts 40, 47, 62, 390, 392, 706, 730,
batch distillation 174-215,274-280, 292 732, 737-743
batch monitoring 20, 28-32, 34, 242-257 Gantt-kit 706-743
- empirical multivariate statistical models generalized dual 426-431
246-257 generalized reduced gradient search 417,
model based 242-246 432-435
batch plant operation I, II, 12, 15-18,20, 24, genetic programming 595, 600-613
25,28-35,41,50-57,61-76 granularity 406
batch polymerization 838-849 heuristic approach 814-818
batch reactors 242-246,259, 265, 266, hierarchical parallelism 410, 411
274-277, 278 high pressure chromatography 216, 231,233,
batch size 8-10, 660, 684-685 234,240
BATCHES 38-40, 368, 376-392, 851, 852, high index problems 331, 346-349
862 identification 295, 297
Bender's decomposition 417, 446-449, index reduction 331, 348-359
466-471,484-486,488,501,514 inference 530, 539-547
blackboard 562, 571-576 input-output 309-329
branch and bound 451,457-462, 706, 716-721 integration 754-757
brewery 851-862 interactive scheduling 706-743
campaign 660, 682-685 interior point algorithms 417, 444-446
chromatography 216, 227, 229, 231, 233, 234, intermediate storage 86, 96, 106, 114, 121,
240 124,679-681
clustering 309, 312, 313,642-649 interval analysis 779-807
computer integrated manufacturing 20, 29, 32, intraparticle convection 216, 231-240
40,41,49,50,52-54,58-61,75 just in time 24, 49, 50, 58, 59
computer architecture 402-406 knowledge based system architecture 561-574
constrained derivatives 417, 422-425 knowledge based systems 530-594, 706, 716,
continuous/discrete simulation 20, 38-40, 361, 717,719-721
370-375,376-392 Kuhn-Tucker multipliers 425, 426
control 47, 20, 30-32, 35,174,210-212,259, lagrange multipliers 417, 423-425
268,269, 838-849 learning (supervised/unsupervised) 631,
- strategies 274, 275, 284-292 639-640, 658
controlability 274, 275, 280-284, 286-292 leather industry 808, 812-814
cycle time 9, 660, 683, 684 linear programming 417,437-449
data base management 759, 760 linear mathematical model 279-283
data base intelligent 525-527 logic programming 542-547
design under uncertainty 495-514 machine learning 595, 614-629
DlCOPT ++ 114, 116, 133-143, 145,453,470, manufacturing environment 21-25
478,485 materials handling 20, 26, 28
 868

mathematical model 216-241, 277-279 principle component analysis 654, 655

mathematical programming 27,36-38,41, production planning 1,16-18,706,727-730
114-149,151-172 ,451-494, 498, 499,512, production management 750, 765-767
513,779-807,808-814 production systems 548-554
mixed integer linear programming 451-465, qualitative simulation 523, 524
473-480,499,500,808-814 radial basis functions 309-313, 319, 320, 325
mixed integer non-linear programming 27, reactive scheduling 20, 29-34, 41,699-703
36-38,41,114-116,117-147,151-167, reactive batch distillation 207-209, 274-280,
169-172,451-455,465-471,484-486,488, 291
498,512,513,598 recipe 1,4,660,666
model mismatch 295, 297-299 reformulation techniques 451, 471-478
model predictive control 838-849 regularity 331, 344
modeling 174, 182-188,761-767 representation 530, 538-561, 706, 719-723
- environments 525-527, 760, 761 resource 151-154, 160-167
module based scheduling algorithm 64-67 - constrained scheduling 20, 33, 693-698
monte carlo simulation 821-823, 831, 832 - constraints 660, 693-698, 774-807
multiperiod optimization 201-205 retrofit design 20, 27, 28, 86, \03-108,
multiplant coordination 20, 33, 34 151-154,160-167,851,853-855
multiproduct plant 1,12-14,15,86-89, rule based programming 522, 523, 539-542,
92-\08,114-147,453,480,481,483-486, 711,714,717
495,496,510-513,515,660,671-672, SC-net 595, 617-629
680-686, 808-819 scheduling 1,15-18,20,29-34,36,40,41,
mUltipurpose plant I, 15-17,20,26,27,49, 44,61-69,86,88,89,98-108,451,454,
70-76,98-108,151-154,160-167,453-455, 478-484,486-488, 595-600, 606-608,
660,671,673,686-688,693-694,750, 660-704,706-749,750-752,762,762,
772-776 779-807,808,821,834-836
multiway principle component analysis 243, search 530-538
247-257 semi-batch reactors 242-245, 249-257
neural networks 309-313, 319, 595, 616, 617, simulated moving bed 216, 228-231
631-659 simulated annealing 88
nonconvex optimization 451, 468-477, simulation 174, 189-190, 192-197
486-488 single product 88, 91, 92
nonlinear dynamic models 309-329 sorption processes 216-241
nonlinear autoregressive models 309, 310, special purpose algorithms 399, 419
313-317,327-9 speed-up end efficiency 400
numerical solutions 331, 332, 334, 349-359 standardization of equipment 44-46
object oriented programming 514, 525, 560, state estimation 242-246, 259, 267, 268
561 statistical design 495
operating strategy 174, 175, 178-212,660, statistical process control 242, 246
664,665,671-674 stochastic flexibility 497,5\0-513,515
operating policy 280-282 storage 660, 679
optimal control 174, 191-212 successive quadratic programming 39, 417,
optimization 48, 20, 36-38,174,97-111, 435-437
114-118,259,270,271,417-449,495-515 task 667-668
ordinary differential equations 331-359 - networks 20, 25, 26, 478-480
parallel computing 393-416 - structure synthesis 140-147
parametric pumping 216, 217, 214-217 tendency modeling 259, 262-272
pattern search 417, 431-2 textile industry 808, 815-819
pharmaceuticals 78, 80-82 time advance mechanisms 361, 364-368
pipeless plant 49, 70-75 time discretization 779, 782-793
planning 595-600, 660-704 time interval 779, 789-793
plant layout 20,27,28 uncertainty 20, 23, 24, 32
prediction 301, 302, 305 waste water equalization 821, 823-825,
preliminary design 20, 26, 27, 114-147 833-836
pressure swing adsorption 231-233, 235-240 worst case analysis 495, 498
NATO ASI Series F
Including Special Programmes on Sensory Systems for Robotic Control (ROB) and on
Advanced Educational Technology (AET)
Vol. 137: Technology-Based Learning Environments. Psychological and Educational Foundations.
Edited by S. Vosniadou, E. De Corte and H. Mandl. X, 302 pages. 1994. (AET)
Vol. 138: Exploiting Mental Imagery with Computers in Mathematics Education. Edited by
R. Sutherland and J. Mason. VIII, 326 pages. 1995. (AET)
Vol. 139: Proof and Computation. Edited by H. Schwichtenberg. VII, 470 pages. 1995.
Vol. 140: Automating Instructional Design: Computer-Based Development and Delivery Tools. Edited
by R. D. Tennyson and A. E. Barron. IX, 618 pages. 1995. (AET)
Vol. 141: Organizational Learning and Technological Change. Edited by C. Zucchermaglio, S.
Bagnara and S. U. Stucky. X, 368 pages. 1995. (AET)
Vol. 142: Dialogue and Instruction. Modeling Interaction in Intelligent Tutoring Systems. Edited by
R.-J. Beun, M. Baker and M. Reiner. IX, 368 pages. 1995. (AET)
Vol. 143: Batch Processing Systems Engineering. Fundamentals of Chemical Engineering. Edited by
G. V. Reklaitis, A. K. Sunol, D. W. T. Rippin, and O. Hortac;:su. XIV, 868 pages. 1996.
Vol. 144: The Biology and Technology of Intelligent Autonomous Agents. Edited by Luc Steels. VIII,
517 pages. 1995.
Vol. 145: Advanced Educational Technology: Research Issues and Future Potential. Edited by T. T.
Liao. VIII, 219 pages. 1996. (AET)
Vol. 146: Computers and Exploratory Leaming. Edited by A. A. diSessa, C. Hoyles and R. Noss.
VIII, 482 pages. 1995. (AET)
Vol. 147: Speech Recognition and Coding. New Advances and Trends. Edited by A. J. Rubio Ayuso
and J. M. L6pez Soler. XI, 505 pages. 1995.
Vol. 148: Knowledge Acquisition, Organization, and Use in Biology. Edited by K. M. Fisher and M. R.
Kibby. X, 246 pages. 1996. (AET)
Vol. 149: Emergent Computing Methods in Engineering Design. Applications of Genetic Algorithms
and Neural Networks. Edited by D.E. Grierson and P. Hajela. VIII, 350 pages. 1996.
Vol. 150: Speech reading by Humans and Machines. Edited by D. G. Stork and M. E. Hennecke. XV,
686 pages. 1996.
Vol. 151: Computational and Conversational Discourse. Burning Issues - An Interdisciplinary
Account. Edited by E. H. Hovy and D. R. Scott. XII, 202 pages. 1996.
Vol. 152: Deductive Program Design. Edited by M. Broy. IX, 467 pages. 1996.
Vol. 153: Identification, Adaptation, Learning. Edited by S. Bittanti and G. Picci. XIV, 553 pages. 1996.
Vol. 154: Reliability and Maintenance of Complex Systems. Edited by S. Ozekici. XI, 589 pages. 1996.
Vol. 156: Microcomputer-Based Labs: Educational Research and Standards. Edited by R.F. Tinker.
XIV, 398 pages. 1996.
NATO ASI Series F
Including Special Programmes on Sensory Systems for Robotic Control (ROB) and on
Advanced Educational Technology (AET)
Vol. 114: Intelligent Systems: Safety, Reliability and Maintainability Issues. Edited by O. Kaynak,
G. Honderd and E. Grant. XI, 340 pages. 1993.
Vol. 115: Learning Electricity and Electronics with Advanced Educational Technology. Edited by
M. Caillot. VII, 329 pages. 1993. (AET)
Vol. 116: Control Technology in Elementary Education. Edited by B. Denis. IX, 311 pages. 1993. (AET)
Vol. 117: Intelligent Learning Environments: The Case of Geometry. Edited by J.-M. Laborde. VIII, 267
pages. 1996. (AET)
Vol. 118: Program Design Calculi. Edited by M. Broy. VIII, 409 pages. 1993.
Vol. 119: Automating Instructional Design, Development, and Delivery. Edited by. R. D. Tennyson.
VIII, 266 pages. 1994. (AET)
Vol. 120: Reliability and Safety Assessment of Dynamic Process Systems. Edited by T. Aldemir,
N. O. Siu, A. Mosleh, P. C. Cacciabue and B. G. G6ktepe. X, 242 pages. 1994.
Vol. 121: Learning from Computers: Mathematics Education and Technology. Edited by C. Keitel and
K. Ruthven. XIII, 332 pages. 1993. (AET)
Vol. 122: Simulation-Based Experiential Learning. Edited by D. M. Towne, T. de Jong and H. Spada.
XIV, 274 pages. 1993. (AET)
Vol. 123: User-Centred Requirements for Software Engineering Environments. Edited by D. J.
Gilmore, R. L. Winder and F. Detienne. VII, 377 pages. 1994.
Vol. 124: Fundamentals in Handwriting Recognition. Edited by S. Impedovo. IX, 496 pages. 1994.
Vol. 125: Student Modelling: The Key to Individualized Knowledge-Based Instruction. Edited by J. E.
Greer and G. I. McCalla. X, 383 pages. 1994. (AET)
Vol. 126: Shape in Picture. Mathematical Description of Shape in Grey-level Images. Edited by
Y.-L. 0, A. Toet, D. Foster, H. J. A. M. Heijmans and P. Meer. XI, 676 pages. 1994.
Vol. 127: Real Time Computing. Edited by W. A. Halang and A. D. Stoyenko. XXII, 762 pages. 1994.
Vol. 128: Computer Supported Collaborative Learning. Edited by C. O'Malley. X, 303 pages. 1994.
(AET)
Vol. 129: Human-Machine Communication for Educational Systems Design. Edited by M. D.
Brouwer-Janse and T. L. Harrington. X, 342 pages. 1994. (AET)
Vol. 130: Advances in Object-Oriented Database Systems. Edited by A. Dogac, M. T. Ozsu, A. Biliris
and T. Sellis. XI, 515 pages. 1994.
Vol. 131: Constraint Programming. Edited by B. Mayoh, E. Tyugu and J. Penjam. VII, 452 pages.
1994.
Vol. 132: Mathematical Modelling Courses for Engineering Education. Edited by Y. Ersoy and A. O.
Moscardini. X, 246 pages. 1994. (AET)
Vol. 133: Collaborative Dialogue Technologies in Distance Learning. Edited by M. F. Verdejo and
S. A. Cerri. XIV, 296 pages. 1994. (AET)
Vol. 134: Computer Integrated Production Systems and Organizations. The Human-Centred
Approach. Edited by F. Schmid, S. Evans, A. W. S. Ainger and R. J. Grieve. X, 347 pages. 1994.
Vol. 135: Technology Education in School and Industry. Emerging Didactics for Human Resource
Development. Edited by D. Blandow and M. J. Dyrenfurth. XI, 367 pages. 1994. (AET)
Vol. 136: From Statistics to Neural Networks. Theory and Pattern Recognition Applications. Edited
by V. Cherkassky, J. H. Friedman and H. Wechsler. XII, 394 pages. 1994.
NATO ASI Series F
Including Special Programmes on Sensory Systems for Robotic Control (ROB) and on
Advanced Educational Technology (AET)
Vol. 92: Hypermedia Courseware: Structures of Communication and Intelligent Help. Edited by
A. Oliveira. X, 241 pages. 1992. (AE7J
Vol. 93: Interactive Multimedia Learning Environments. Human Faciors and Technical Considerations
on Design Issues. Edited by M. Giardina. VIII, 254 pages. 1992. (AE7J
Vol. 94: Logic and Algebra of Specification. Edited by F. L. Bauer, W. Brauer, and H. Schwichtenberg.
VII, 442 pages. 1993.
Vol. 95: Comprehensive Systems Design: A New Educational Technology. Edited by C. M. Reigeluth,
B. H. Banathy, and J. R. Olson. IX, 437 pages. 1993. (AE7J
Vol. 96: New Directions in Educational Technology. Edited by E. Scanlon and T. O'Shea. VIII, 251
pages. 1992. (AE7J
Vol. 97: Advanced Models of Cognition for Medical Training and Practice. Edited by D. A. Evans and
V. L. Patel. XI, 372 pages. 1992. (AE7J
Vol. 98: Medical Images: Formation, Handling and Evaluation. Edited by A. E. Todd-Pokropek and
M. A. Viergever. IX, 700 pages. 1992.
Vol. 99: Multisensor Fusion for Computer Vision. Edited by J. K. Aggarwal. XI, 456 pages. 1993. (ROB)
Vol. 100: Communication from an Artificial Intelligence Perspective. Theoretical and Applied Issues.
Edited by A. Ortony, J. Slack and O. Stock. XII, 260 pages. 1992.
Vol. 101: Recent Developments in Decision Support Systems. Edited by C. W. Holsapple and A. B.
Whinston. XI, 618 pages. 1993.
Vol. 102: Robots and Biological Systems: Towards a New Bionics? Edited by P. Dario, G. Sandini and
P. Aebischer. XII, 786 pages. 1993.
Vol. 103: Parallel Computing on Distributed Memory Multiprocessors. Edited by F. Ozguner and
F. Enpl. VIII, 332 pages. 1993.
Vol. 104: Instructional Models in Computer-Based Learning Environments. Edited by S. Dijkstra,
H. P. M. Krammer and J. J. G. van Merrienboer. X, 510 pages. 1993. (AE7J
Vol. 105: Designing Environments for Constructive Learning. Edited by T. M. Duffy, J. Lowyck and
D. H. Jonassen. VIII, 374 pages. 1993. (AE7J
Vol. 106: Software for Parallel Computation. Edited by J. S. Kowalik and L. Grandinetti. IX, 363 pages.
1993.
Vol. 107: Advanced Educational Technologies for Mathematics and Science. Edited by D. L.
Ferguson. XII, 749 pages. 1993. (AE7J
Vol. 108: Concurrent Engineering: Tools and Technologies for Mechanical System Design. Edited by
E. J. Haug. XIII, 998 pages. 1993.
Vol. 109: Advanced Educational Technology in Technology Education. Edited by A. Gordon,
M. Hacker and M. de Vries. VIII, 253 pages. 1993. (AE7J
Vol. 110: Verification and Validation of Complex Systems: Human Factors Issues. Edited by J. A.
Wise, V. D. Hopkin and P. Stager. XIII, 704 pages. 1993.
Vol. 111: Cognitive Models and Intelligent Environments for Learning Programming. Edited by
E. Lemut, B. du Boulay and G. Dettori. VIII, 305 pages. 1993. (AE7J
Vol. 112: Item Banking: Interactive Testing and Self-Assessment. Edited by D. A. Leclercq and J. E.
Bruno. VIII, 261 pages. 1993. (AE7J
Vol. 113: Interactive Learning Technology for the Deaf. Edited by B. A. G. Elsendoorn and F. Coninx.
XIII, 285 pages. 1993. (AE7J
NATO ASI Series F
Including Special Programmes on Sensory Systems for Robotic Control (ROB) and on
Advanced Educational Technology (AET)
Vol. 70: Numerical Linear Algebra, Digital Signal Processing and Parallel Algorithms. Edited by
G. H. Golub and P. Van Dooren. XIII, 729 pages. 1991.
Vol. 71: Expert Systems and Robotics. Edited by T. Jordanides and B.Torby. XII, 744 pages. 1991.
Vol. 72: High-Capacity Local and Metropolitan Area Networks. Architecture and Performance Issues.
Edited by G. Pujolle. X, 536 pages. 1991.
Vol. 73: Automation and Systems Issues in Air Traffic Control. Edited by J. A. Wise, V. D. Hopkin and
M. L. Smith. XIX, 594 pages. 1991.
Vol. 74: Picture Archiving and Communication Systems (PACS) in Medicine. Edited by H. K. Huang,
O. Ratib, A. R. Bakker and G. Witte. XI, 438 pages. 1991.
Vol. 75: Speech Recognition and Understanding. Recent Advances, Trends and Applications. Edited
by P. Laface and Renato De Mori. XI, 559 pages. 1991.
Vol. 76: Multimedia Interface Design in Education. Edited by A. D. N. Edwards and S. Holland. XIV,
216 pages. 1992. (AET)
Vol. 77: Computer Algorithms for Solving Linear Algebraic Equations. The State of the Art. Edited by
E. Spedicato. VIII, 352 pages. 1991.
Vol. 78: Integrating Advanced Technology into Technology Education. Edited by M. Hacker,
A. Gordon and M. de Vries. VIII, 185 pages. 1991. (AET)
Vol. 79: Logic, Algebra, and Computation. Edited by F. L. Bauer. VII, 485 pages. 1991.
Vol. 80: Intelligent Tutoring Systems for Foreign Language Leaming. Edited by M. L. Swartz and
M. Yazdani. IX, 347 pages. 1992. (AET)
Vol. 81: Cognitive Tools for Learning. Edited by P. A. M. Kommers, D. H. Jonassen, and J. T. Mayes.
X, 278 pages. 1992. (AET)
Vol. 82: Combinatorial Optimization. New Frontiers in Theory and Practice. Edited by M. AkgOI, H. W.
Hamacher, and S. TOfekg. XI, 334 pages. 1992.
Vol. 83: Active Perception and Robot Vision. Edited by A. K. Sood and H. Wechsler. IX, 756 pages.
1992.
Vol. 84: Computer-Based Learning Environments and Problem Solving. Edited by E. De Corte, M. C.
Linn, H. Mandl, and L. Verschaffel. XVI, 488 pages. 1992. (AET)
Vol. 85: Adaptive Learning Environments. Foundations and Frontiers. Edited by M. Jones and P. H.
Winne. VIII, 408 pages. 1992. (AET)
Vol. 86: Intelligent Learning Environments and Knowledge Acquisition in Physics. Edited by
A. Tiberghien and H. Mandl. VIII, 285 pages. 1992. (AET)
Vol. 87: Cognitive Modelling and Interactive Environments. With demo diskettes (Apple and IBM
compatible). Edited by F. L. Engel, D. G. Bouwhuis, T. Bosser, and G. d'Ydewalle. IX, 311 pages.
1992. (AET)
Vol. 88: Programming and Mathematical Method. Edited by M. Broy. VIII, 428 pages. 1992.
Vol. 89: Mathematical Problem Solving and New Information Technologies. Edited by J. P. Ponte,
J. F. Matos, J. M. Matos, and D. Fernandes. XV, 346 pages. 1992. (AET)
Vol. 90: Collaborative Learning Through ComputerConferencing. Edited by A. R. Kaye. X, 260 pages.
1992. (AET)
Vol. 91: New Directions for Intelligent Tutoring Systems. Edited by E. Costa. X, 296 pages. 1992.
(AET)
 Springer
and the
•
environ
A.S.,in ... we firmly believe that an
in'~llatior..l science publisher has a
spe.:jal obligation to the environment,
L__!!!!!!PUlco"pc)ra.,e policies consistently
reflect this conviction.
We also expect our business partners -
paper mills, printers, packaging
manufacturers. etc. - to commit
themselves to using materials and
production processes that do not harm
the environment. The paper in this
book is made from low- or no-chlorine
pulp and is acid free, in conformance
with international standards for paper
permanency_

Springer