Process Systems Engineering

Edited by
Efstratios N. Pistikopoulos,
Michael C. Georgiadis, and
Vivek Dua

Volume 4
Supply Chain Optimization, Part II
Volume Edited by
Lazaros G. Papageorgiou and
Michael C. Georgiadis
Related Titles

Haber, R., Bars, R., Schmitz, U. Bröckel, U., Meier, W., Wagner, G. (eds.)
Predictive Control in Process Product Design and Engineering
Engineering Best Practices
approx. 400 pages with approx. 760 pages in 2 volumes with 355 figures and
150 figures 60 tables
2007 2007
Hardcover Hardcover
ISBN: 978-3-527-31492-8 ISBN: 978-3-527-31529-1

Engell, S. (ed.) Ingham, J., Dunn, I. J., Heinzle, E.,

Logistic Optimization of Prenosil, J. E., Snape, J. B.

Chemical Production Chemical Engineering Dynamics

Processes An Introduction to Modelling and Computer
approx. 300 pages Simulation
2007
640 pages with 474 figures and 4 tables
Hardcover
2007
ISBN: 978-3-527-30830-9
Hardcover
ISBN: 978-3-527-31678-6
Dimian, A. C., Bildea, C. S.
Chemical Process Design Puigjaner, L., Heyen, G. (eds.)
Computer-Aided Case Studies Computer Aided Process and
approx. 500 pages with approx. 450 figures Product Engineering
2008 910 pages in 2 volumes with 281 figures
Hardcover and 77 tables
ISBN: 978-3-527-31403-4 2006
Hardcover
ISBN: 978-3-527-30804-0
Keil, F. J. (ed.)
Modeling of Process
Agachi, P. S., Nagy, Z. K., Cristea, M. V.,
Intensification Imre-Lucaci, A.
422 pages with 180 figures and 28 tables
2007 Model Based Control
Hardcover Case Studies in Process Engineering
ISBN: 978-3-527-31143-9
290 pages with 156 figures and 25 tables
2007
Hardcover
ISBN: 978-3-527-31545-1
Process Systems Engineering
Edited by
Efstratios N. Pistikopoulos, Michael C. Georgiadis, and
Vivek Dua

Volume 4: Supply Chain Optimization, Part II

Volume Edited by
Lazaros G. Papageorgiou and Michael C. Georgiadis
The Editors All books published by Wiley-VCH are
carefully produced. Nevertheless, authors,
Prof. Efstratios N. Pistikopoulos editors, and publisher do not warrant the
Centre for Process Systems Engineering information contained in these books,
Department of Chemical Engineering including this book, to be free of errors.
Imperial College London Readers are advised to keep in mind that
Roderic Hill Building statements, data, illustrations, procedural
South Kensington Campus details or other items may inadvertently be
London SW7 2AZ inaccurate.
United Kingdom
Library of Congress Card No.:
Dr. Michael C. Georgiadis applied for
Centre for Process Systems Engineering
Department of Chemical Engineering British Library Cataloguing-in-Publication Data
Imperial College London A catalogue record for this book is available
Roderic Hill Building from the British Library.
London SW7 2AZ
Bibliographic information published by
United Kingdom
the Deutsche Nationalbibliothek
Dr. Vivek Dua Die Deutsche Nationalbibliothek lists this
Centre for Process Systems Engineering publication in the Deutsche
Department of Chemical Engineering Nationalbibliografie; detailed bibliographic
University College London data are available in the Internet at
London WC1E 7JE <http://dnb.d-nb.de>.
United Kingdom
© 2008 WILEY-VCH Verlag GmbH & Co.
Dr. Lazaros G. Papageorgiou KGaA, Weinheim
Centre for Process Systems Engineering
All rights reserved (including those of
Department of Chemical Engineering
translation into other languages). No part of
University College London
this book may be reproduced in any form – by
London WC1E 7JE
photoprinting, microfilm, or any other
United Kingdom
means – nor transmitted or translated into a
machine language without written
permission from the publishers. Registered
names, trademarks, etc. used in this book,
even when not specifically marked as such,
are not to be considered unprotected by law.

Typesetting VTEX, Litauen

Printing Strauss GmbH, Mörlenbach
Binding Litges & Dopf buchbinderei
GmbH, Heppenheim
Cover Design Grafik-Design Schulz,
Fußgönheim

Printed in the Federal Republic of Germany

Printed on acid-free paper

ISBN: 978-3-527-31906-0
 V

Contents

Preface – Volume 4: Supply Chain Optimization XIII

List of Authors XVII

1 Supply Chains of High-Value Low-Volume Products 1

R. Sousa, N. Shah, L. G. Papageorgiou 1
1.1 Introduction 1
1.1.1 Pharmaceutical Industry 3
1.2 Literature on Supply-Chain Modeling and Optimization 4
1.2.1 Case Studies 4
1.2.2 Deterministic Studies 6
1.2.3 Case Study Part I: Problem Statement 9
1.3 Stochastic Models 11
1.3.1 Stochastic Programming with Recourse 12
1.3.2 Stochastic Dynamic Programming (for Optimal Control) 13
1.3.3 Works in Planning Optimization Under Uncertainty 13
1.3.3.1 Stochastic Programming, Probability-Based Approach 13
1.3.3.2 Stochastic Programming, Scenario-Based Approach 14
1.3.3.3 Stochastic Dynamic Programming Approach 16
1.3.4 Case Study Part II: Planning Under Uncertainty 16
1.3.4.1 Scenarios 17
1.4 Solution Algorithms 18
1.4.1 Decomposition and Hierarchical Algorithms 18
1.4.2 Hybrid Methods 20
1.4.2.1 Lagrangean Decomposition 20
1.4.2.2 Genetic Algorithms 21
1.4.3 Case Study Part III: Solution Methodology 22
1.4.3.1 Solution Algorithm Performance 22
1.4.3.2 Financial Risk 24
1.4.3.3 Comments 24
1.5 Summary 25

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
VI Contents

2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite

Systems by a Rigorous Optimization Approach 29
R. Dondo, C. A. Méndez, J. Cerdá 29
2.1 Introduction 30
2.1.1 Previous Heuristic Approaches 31
2.1.2 Exact Optimization Methods 32
2.2 The Problem Definition 34
2.3 The Problem Mathematical Formulation 35
2.3.1 The Problem Constraints 36
2.3.1.1 Assignment Constraints 36
2.3.1.2 Routing-Cost Defining Constraints 37
2.3.1.3 Arrival-Time Defining Constraints 39
2.3.1.4 Vehicle-Load Defining Constraints 40
2.3.2 The Problem Objective Function 43
2.4 Time Window-Based Variable and Constraint Elimination Rules 43
2.4.1 Preassigning Vehicles to Transportation Requests 45
2.5 Numerical Results and Discussion 46
2.6 Conclusions 56

3 A Real Time Approximate Dynamic Programming Approach:

A High Dimensional Supply Chain Application 61
N. E. Pratikakis, M. J. Realff, H. Lee 61
3.1 Introduction 61
3.1.1 Chapter Structure 63
3.2 Dynamic Programming 63
3.2.1 The Value Function 63
3.2.2 Markov Decision Processes 64
3.2.3 From Value Iteration to Asynchronous Value Iteration to Real-Time
Dynamic Programming 65
3.2.4 A Review Of Approximate Dynamic Programming Techniques 66
3.2.4.1 Minimizing the COD Concerning |S| 66
3.2.4.2 Minimizing the COD Concerning |A| 67
3.2.4.3 Minimizing the COD Concerning the Expectation Operator 68
3.2.4.4 Value Function Approximators 68
3.3 A High-Dimensional Supply Chain Case Study 69
3.3.1 Introduction 69
3.3.2 Mathematical Modelling of the Supply Chain 70
3.3.3 Sets 70
3.3.4 Material Balances at Each Tank 70
3.3.4.1 Constraints on yT ,C,(P m) (t) 71
3.3.5 Reaction and Separation Processes – The Determination of
Pru,p (t) 73
3.3.5.1 Reaction Processes 73
3.3.5.2 Separation Processes 74
 Contents VII

3.3.5.3 Hard Constraints 74

3.3.6 Decision Variables 75
3.3.7 Objective Function 75
3.3.7.1 Net Profit p(t) 75
3.3.7.2 Net Cost c(t) 75
3.4 Formulating the Problem as an MDP 76
3.4.1 State Variables 76
3.4.2 Decision Variables 76
3.4.3 Transition Function 76
3.4.4 Objective Function 77
3.4.5 Concluding Remarks 77
3.5 A Real-Time Approximate Dynamic Programming Algorithm 78
3.5.1 The Greedy RTADP Algorithm 79
3.5.2 Key Elements of Asub 80
3.5.3 On Calculating Jtπ (st ) 81
3.6 Simulation Results 82
3.6.1 Simulation Procedure 83
3.6.2 Performance Comparison 84
3.7 Conclusions 86

4 Robust Supply-Chain Operations through Rescheduling 89

A. Adhitya, R. Srinivasan, I. A. Karimi 89
4.1 Introduction 89
4.2 Refinery Supply Chain 91
4.3 Rescheduling Problem Statement 92
4.3.1 Literature Review 95
4.4 Rescheduling Methodologies 96
4.4.1 Heuristic-Based Rescheduling 96
4.4.2 Model-Based Rescheduling 98
4.5 Case Study 105
4.6 Discussions 117

5 Supply Chain Tactical Optimization in the Fruit Industry 121

G. L. Masini, A. M. Blanco, N. Petracci, J. A. Bandoni 121
5.1 Introduction 122
5.2 Literature Review 124
5.3 Fruit Industry Supply Chain Description 126
5.3.1 Farms 126
5.3.2 Packaging Plants 126
5.3.3 Cold Storage 127
5.3.4 Fruit Reception Sites 127
5.3.5 Milling Plants 127
5.3.6 Concentrated Juice Plants 127
VIII Contents

5.3.7 Cider Plants 128

5.3.8 Clients 128
5.3.9 Third Party Suppliers and Customers 128
5.3.10 Transportation 128
5.3.11 Global FISC Operations 128
5.3.12 Tactical Model Scope 129
5.4 Mathematical Programming Model 130
5.4.1 Fruit Production in Farms 130
5.4.2 Allocation of Fresh Fruit from Own Farms 131
5.4.3 Fruit Reception at PPs 132
5.4.4 Fruit Processing at PPs 132
5.4.5 Waste Fruit from PPs 132
5.4.6 Packed Fruit from PPs and TPSs 133
5.4.7 Fruit Balance at CSFs 133
5.4.8 Fruit Balance at FRSs 134
5.4.9 Fruit Balance at MPs 135
5.4.10 Raw Material Reception at CJPs 135
5.4.11 Fruit Processing at CJPs 136
5.4.12 Juice Storage at CJSs 136
5.4.13 Raw Material Reception at CPs 136
5.4.14 Fruit Processing at CPs 137
5.4.15 Packed Fruit Delivery to Clients 137
5.4.16 Concentrated Juice Delivery to Clients 137
5.4.17 Third Party Material Availability 137
5.4.18 Product Delivery Constraints 138
5.4.19 Raw Material Costs 139
5.4.20 Final Products Purchase Cost 140
5.4.21 Transportation Costs 140
5.4.22 Cold Storage Cost (Third Party Rental) 143
5.4.23 Operating Costs 143
5.4.24 Sales Income 144
5.4.25 Cold Storage Capacity Rental Income 145
5.4.26 Objective Function 145
5.4.27 Company Supply Chain Optimization Problem 145
5.5 Results and Discussion 145
5.6 Conclusions and Future Work 154

6 Short-Term Scheduling of Batch and Continuous Processes 173

M. A. Shaik, Ch. A. Floudas 173
6.1 Classification of Scheduling Formulations 173
6.1.1 Time Representation 173
6.1.2 Characteristics of Process Scheduling Problems 175
6.1.2.1 Processing Sequences 175
6.1.2.2 Processing Modes of Operation 176
 Contents IX

6.1.2.3 Intermediate Storage Policies 176

6.1.2.4 Demand Patterns 177
6.1.2.5 Resource Considerations 177
6.1.2.6 Changeovers 177
6.1.2.7 Performance Criteria 177
6.2 Short-Term Scheduling of Batch Processes 177
6.2.1 Unit-Specific Event-Based Model of Ierapetritou and Floudas [29]
(I&F) 178
6.2.2 Global Event-Based Model of Castro and coworkers [21, 22] 179
6.2.3 Global Event-Based Model of Maravelias and Grossmann [28]
(M&G) 181
6.2.4 Slot-Based Model of Sundaramoorthy and Karimi [10] (S&K) 183
6.2.5 Computational Studies 185
6.2.5.1 Example 1 185
6.2.5.2 Example 2 188
6.2.5.3 Example 3 189
6.3 Short-Term Scheduling of Continuous Processes 195
6.3.1 Unlimited Intermediate Storage 197
6.3.1.1 Allocation Constraints 197
6.3.1.2 Capacity Constraints for Processing Tasks 197
6.3.1.3 Material Balances 197
6.3.1.4 Demand Constraints 198
6.3.1.5 Duration Constraints for Processing Tasks 199
6.3.1.6 Sequencing Constraints 199
6.3.1.7 Extra Tightening Constraint 201
6.3.2 Dedicated Finite Intermediate Storage with Storage Bypassing
Allowed 201
6.3.2.1 Storage Bypassing Allowed 201
6.3.3 No Intermediate Storage 202
6.3.4 Flexible Finite Intermediate Storage with Storage Bypassing
Allowed 202
6.3.4.1 Allocation Constraints for Storage Tasks 203
6.3.4.2 Capacity Constraints for Storage Tasks 203
6.3.4.3 Material Balances 203
6.3.4.4 Duration Constraints for Storage Tasks 204
6.3.4.5 Sequencing Constraints. Different Tasks in Different Units for
Storage Tasks 204
6.3.5 Dedicated and Flexible Finite Intermediate Storage without Storage
Bypassing 205
6.3.5.1 Dedicated-Finite-Intermediate-Storage Case without Bypassing of
Storage 205
6.3.5.2 Flexible-Finite-Intermediate-Storage Case without Bypassing of
Storage 206
6.3.6 Computational Study 207
6.3.6.1 Example 4 207
X Contents

6.3.6.2 Unlimited Intermediate Storage (Case A) 210

6.3.6.3 No Intermediate Storage (Case B) 210
6.3.6.4 Flexible Finite Intermediate Storage (Case C) 211
6.4 Conclusions 214

7 Modeling and Optimization of Refinery Operations Considering

Uncertainty 219
Z. Jia, M. G. Ierapetritou 219
7.1 Introduction 219
7.2 Refinery Scheduling 221
7.3 Uncertainty Analysis 222
7.3.1 Multiobjective Robust Optimization 222
7.3.2 Parametric Mixed Integer Linear Programming (pMILP) 224
7.4 Case Studies 225
7.5 Summary and Future Work 234

8 Production and Inventory Planning for Stock Preparation in the

Tissue Paper Industry 237
J. Westerlund, M. Hästbacka, J. Kaplin, T. Westerlund 237
8.1 Introduction 237
8.2 Mathematical Programming as a Tool for Production
Optimization 239
8.3 Metsä Tissue Mänttä Mill 240
8.3.1 Process Description 241
8.3.2 Corporate Supply Planning at Metsä Tissue Corporation 244
8.4 The MISPT Tool 245
8.4.1 The MISPT Architecture 246
8.4.2 The MISPT Graphical-User Interface 248
8.4.3 The MISPT Tool for Daily Production Planning 249
8.4.3.1 Planning of Production Stops 250
8.4.3.2 Aging Profiles 250
8.4.3.3 Time Synchronization 250
8.4.4 The MISPT Tool for Strategic Planning 251
8.5 The Mixed-Time MILP Model 252
8.5.1 Mathematical Formulation 253
8.6 Illustrative Case Examples 255
8.6.1 Case Example 1 256
8.6.1.1 Sequence-Dependent Changeovers 257
8.6.1.2 MISPT Results for Case Example 1 259
8.6.2 Case Example 2 263
8.6.3 Case Example 3 265
8.7 Discussion 266
 Contents XI

9 Production Planning in Process Systems Engineering 269

C. Sung, C. T. Maravelias 269
9.1 Introduction 269
9.2 Background 271
9.2.1 Problem Statement 271
9.2.2 Classical Mathematical Programming Formulations 271
9.2.3 Software Systems 273
9.2.4 Challenges in Production Planning in the Chemical Industry 274
9.3 Integration of Production Planning and Scheduling 275
9.3.1 Scheduling Subproblems 275
9.3.1.1 Integration with Detailed Scheduling Formulations 275
9.3.1.2 Integration with Approximate Scheduling Models 275
9.3.1.3 Integration with Surrogate Models Obtained Offline 276
9.3.2 Solution Methods 276
9.3.2.1 Full-Space Methods 277
9.3.2.2 Hierarchical Methods 277
9.3.2.3 Iterative Methods 278
9.4 Applications and Extensions 278
9.5 Conclusions 279

10 The Supply Chain as a Dynamical System 285

K. R. Jillson, E. J. Dozal-Mejorada, B. E. Ydstie 285
10.1 Introduction 285
10.2 Literature Review 287
10.2.1 How did the Field Start? 287
10.2.2 What Needs and Areas are Being Addressed? 288
10.2.2.1 Environmental 288
10.2.2.2 Pharmaceutical 288
10.2.2.3 Semiconductors 289
10.2.2.4 Energy 289
10.2.3 What Modeling Approaches and Solutions Exist? 289
10.2.4 Back to Dynamics 290
10.3 Problem Formulation 292
10.3.1 The Beer Game 298
10.4 Inventory Control 303
10.5 Assembly, Disassembly, and Packaging 309
10.5.1 Bounds on Storage Requirements 314
10.5.2 Feedback Scheduling 319
10.6 Flow Control and Load Balancing 323
10.7 Summary and Conclusions 334

Index 339
 XIII

Preface – Volume 4: Supply Chain Optimization

Modern industrial enterprises are typically multiproduct, multipurpose, and multi-

site facilities operating in different regions and countries and dealing with a global-
wide international clientele. In such enterprise networks, the issues of global enter-
prise planning, coordination, cooperation and robust responsiveness to customer
demands at the global as well as the local level are critical for ensuring effective-
ness, competitiveness, business sustainability, and growth. In this context, it has
long been recognized that there is a need for efficient integrated approaches that
consider, in a systematic way, various levels of enterprise management, plant-wide
coordination and plant operation, in order to reduce capital and operating costs,
increase supply chain productivity and improve business responsiveness.
The supply chain concept has in recent years become one of the main approaches
to achieve enterprise efficiency. The terminology implies that a system view is taken
rather than a functional or hierarchical one. Enterprises cannot be competitive
without considering supply chain activities. This is partially due to the evolving
higher specialization in a more differentiated market. Most importantly, compe-
tition drives companies to reduced cost structures with lower inventories, more
effective transportation systems, and transparent systems able to support informa-
tion throughout the supply chain. A single company rarely controls the production
of a commodity as well as sourcing, distribution, and retail.
Many typical supply chains today have production that spans several countries
and product markets. The opportunities for supply chain improvements are large.
Costs of keeping inventory throughout the supply chain to maintain high customer
service levels are generally significant. There is a wide scope to reduce the inven-
tory while still maintaining the high service standards required. Furthermore, the
manufacturing processes can be improved so as to employ current working capital
and labor more efficiently.
It has widely been recognized that enhanced performance of supply chains ne-
cessitates: (i) appropriate design of supply chain networks and its components and
(ii) effective allocation of available resources over the network. Thus, in the last
few years, there has been a multitude of efforts focused on providing improve-
ments of supply chain management and optimization. These efforts include a wide
range of models: from commercial enterprise resource planning systems and so-

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
XIV Preface – Volume 4: Supply Chain Optimization

called advanced planning systems to academic achievements (for example, linear

and mixed-integer programming, multiagent systems).
Management of supply chains is a complex task mainly due to the large size of
the physical supply network and inherent uncertainties. In a highly competitive
environment, improved decisions are required for efficient supply chain manage-
ment at strategic, tactical, and operational levels with time horizons ranging from
several years to a few days, respectively. Depending on the level, one or more of the
following decisions are taken:
• Number, size and location of manufacturing sites,
warehouses and distribution centers.
• Network connectivity (e.g., allocation of suppliers to plants,
warehouses to markets, etc.).
• Production decisions related to plant production planning
and scheduling.
• Management of inventory levels and replenishment policies.
• Transportation decisions concerning mode of transportation
(e.g., road, rail, etc.) and also size of material shipment.
• Sustainability aspects (e.g., environmental impact
considerations, recycling policies, etc.).
Most of the above challenging research issues are addressed in Volumes 3 and 4
of this book series.
Volume 4 provides a comprehensive review of key research contributions con-
cerning tactical and operational decisions of supply chain problems for the process
industries.
Sousa, Shah, and Papageorgiou in Chapter 1 address a systematic optimization-
based supply chain planning model of large process industrial companies of prod-
ucts with a high added value per mass unit. Here, transportation costs have a re-
duced impact in the total expenditure in supply chain activities, from raw material
supply until distribution to final markets. A case study relevant to the pharmaceu-
tical industry is presented concerning production allocation to manufacturing sites
in a worldwide network. Both deterministic and stochastic cases are addressed and
compared. The large, mixed-integer linear programming (MILP) models generated
were efficiently solved using a temporal decomposition algorithm.
Dondo, Méndez, and Cerda in Chapter 7 consider an MILP-based mathematical
formulation focused on the operational level of multiple vehicle pickup and de-
livery problems with time windows commonly arising in multisite systems. The
proposed two-index model can be solved using a branch-and-cut commercial pack-
age to find the best vehicle routes and schedules for moderate-size problems. The
formulation has been generalized to also consider pure pickup and delivery nodes,
heterogeneous vehicles, multiple depots, as well as many-to-many transportation
requests. To tackle medium-size problems, several preordering and variable and
constraint elimination rules are introduced. Optimal solutions for a variety of
benchmark problems featuring different sizes in terms of customer requests and
vehicles, distinct cluster/random pickup and delivery locations and a range of time-
window width distributions are reported.
 Preface – Volume 4: Supply Chain Optimization XV

Pratikakis, Realff, and Lee in Chapter 3, present a real-time approximate dynamic

programming approach for the efficient solution of multistage supply chain deci-
sion problems under a stochastic environment. The supply chain system experi-
ences stochastic variations in demand and price of the main products which are
modeled using Markov chains. The main idea of the overall methodology is to
start with a heuristic policy derived from an MILP formulation and to gradually
construct a superior quality solution by interacting with the stochastic system via
simulation.
Disruptions could bring about adverse effects such as blockage of material, infor-
mation, and finance flows, loss of production, off-spec products, loss of efficiency,
under- or oversupply, etc. Any of these could result in operational problems, loss of
business opportunities, and financial losses. A deviation between plan and actual
realization occurs during disruptions. Depending on the magnitude of the devia-
tion, the necessary response would have to differ. Adhitya, Srinivasan, and Karimi
in Chapter 4 first describe the crude oil supply chain and clearly state the reschedul-
ing problem. Two new rescheduling methods are then presented to address supply
chain disruptions. A petroleum refinery supply chain is used to identify the salient
characteristics of the problem and review solution methods.
The group of Bandoni and coworkers in Chapter 5 propose a detailed and com-
plete tactical planning model to aid in the negotiation instance of a typical large
company that operates several nodes of the fruit industry supply chain. The pro-
posed linear programming model considers the many interactions of the real net-
work and the typical operative practices of the business. A real-world case study
from the pip fruit industry of Argentina is used to illustrate the applicability of the
proposed model.
The research area of batch and continuous process scheduling has received great
attention from both academia and industry in the past two decades. This is moti-
vated by the increasing pressure to improve efficiency and reduce costs, and by the
significant advances in relevant modeling and solution techniques and the rapidly
growing computational power. Shaik and Floudas in Chapter 6 present a com-
prehensive overview of the recent developments in scheduling of multiproduct,
multipurpose batch, and continuous processes. Different continuous-time based
scheduling models from the literature are discussed, followed by a comparative
study of these methods on some benchmark examples.
Jia and Ierapetritou in Chapter 7 address the problem of refinery scheduling
under uncertainty through a multiobjective optimization and a parametric integer
linear programming approach. Several case studies considering the optimal oper-
ations of crude oil unloading and mixing, and gasoline blending and distributions
are presented to illustrate the importance of considering uncertainty in demand in
refinery scheduling operations.
Westerlund and coworkers in Chapter 8 present an MILP-based tool for the pro-
duction and intermediate planning in the tissue manufacturing industry. The tool
is based on a mixed-time scheduling formulation, implementing tailored software
architecture as an access point to various information systems at the mill. The tool
is furthermore integrated into a user-friendly interface and is concerned with the
XVI Preface – Volume 4: Supply Chain Optimization

production and raw material cost minimization, as well as intermediate storage

utilization at the stock preparation section.
Sung and Maravelias in Chapter 9 provide a review of the methods that have
been proposed for the solution of production planning problems in the chemi-
cal industry. First, the classical formulations used in Operations Research and the
main software tools currently used in practice are discussed. Second, modeling ap-
proaches and optimization strategies for the integration of production planning
with scheduling are presented. Finally, a brief discussion of specific applications
and the integration of production planning with other models is analyzed.
In the final chapter, Ydstie and coworkers study inventory and control flows in
complex supply chains using approaches developed in the area of process con-
trol. New methods for inventory control are introduced using frequency response
analysis and an approach for feedback scheduling of assembly, disassembly, and
repackaging lines is investigated. Furthermore, a method for control is developed
which optimizes the intrinsic value of a business by managing inventory and flows
so that costs are minimized.
This collection represents a set of stand-alone works that captures recent re-
search trends in the development and application of techniques, methodologies,
algorithms, and tools for optimizing various aspects of supply chain systems. We
hope that by the end of the book, the reader will have developed a commanding
comprehension of the main aspects of integrated supply chains, the ability to criti-
cally access the key characteristics and elements related to the design and operation
of supply chains, and the capacity to implement the new technology in practice.
We are extremely grateful to the authors for their outstanding contributions and
for their patience, which have led to a final product that far exceeded our expecta-
tions.

London, June 2007

L. G. Papageorgiou
M. C. Georgiadis
 XVII

List of Authors

Arief Adhitya Eduardo J. Dozal-Mejorada

Institute of Chemical and Engineering Carnegie Mellon University
Sciences Department of Chemical Engineering
1 Pesek Road, Jurong Island 5000 Forbes Avenue
Singapore 627833 Pittsburgh, PA 15213-3890
Singapore USA

Professor José A. Bandoni Professor Christodoulos A. Floudas

Planta Piloto de Ingeniería Química Princeton University
PLAPIQUI (UNS–CONICET) Department of Chemical Engineering,
Camino La Carrindanga, Km. 7 Princeton
8000 Bahía Blanca New Jersey 08544
Argentina USA

Aníbal M. Blanco Mattias Hästbacka

Planta Piloto de Ingeniería Química Åbo Akademi University
PLAPIQUI (UNS–CONICET) Centre for Industrial Engineering and
Camino La Carrindanga, Km. 7 Management
8000 Bahía Blanca Faculty of Technology
Argentina FI-20500 ÅBO
Finland
Dr. Jaime Cerdá
INTEC (UNL–CONICET) Professor Marianthi G. Ierapetritou
Güemes 3450 The State University of New Jersey
3000 Santa Fe Department of Chemical and
Argentina Biochemical Engineering
Rutgers University
Dr. Rodolfo Dondo Piscataway, NJ 08854
INTEC (UNL–CONICET) USA
Güemes 3450
3000 Santa Fe
Argentina

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
XVIII List of Authors

Zhenya Jia Guilermo L. Masini

The State University of New Jersey Universidad Nacional del Comahue
Department of Chemical and Mechanical Engineering Department
Biochemical Engineering Engineering Faculty
Rutgers University Buenos Aires 14000
Piscataway, NJ 08854 8300 Neuquén,
USA Argentina

Kendell R. Jillson Dr. Carlos A. Méndez

Carnegie Mellon University INTEC (UNL–CONICET)
Department of Chemical Engineering Güemes 3450
5000 Forbes Avenue 3000 Santa Fe
Pittsburgh, PA 15213-3890 Argentina
USA
Dr. Lazaros G. Papageorgiou
University College London
Jarkko Kaplin Centre for Process Systems Engineering
Metsa-Tissue Corporation Department of Chemical Engineering
S-54288 Mariestad Torrington Place
Sweden London WC1 7JE
UK
Professor Iftekhar A. Karimi
Department of Chemical and Noemí C. Petracci
Biomolecular Engineering Planta Piloto de Ingeniería Química
National University of Singapore PLAPIQUI (UNS–CONICET),
4 Engineering Drive 4 Camino La Carrindanga, Km. 7
Singapore 117576 8000 Bahía Blanca
Singapore Argentina

Professor Jay H. Lee Nikolaos E. Pratikakis

Georgia Institute of Technology Georgia Institute of Technology
School of Chemical & Biomolecular School of Chemical & Biomolecular
Engineering Engineering,
311 Ferst Drive 311 Ferst Drive
N.W. Atlanta, GA 30332-0100 N.W. Atlanta, GA 30332-0100
USA USA

Professor Matthew J. Realff

Professor Christos T. Maravelias Georgia Institute of Technology
Department of Chemical and Biological School of Chemical & Biomolecular
Engineering Engineering
University of Wisconsin – Madison 311 Ferst Drive
Madison, WI 53706 N.W. Atlanta, GA 30332-0100
USA USA
 List of Authors XIX

Professor Nilay Shah Charles Sung

Imperial College London Department of Chemical and Biological
Centre for Process Systems Engineering Engineering
Department of Chemical Engineering University of Wisconsin – Madison
South Kensington Campus Madison, WI 53706
London SW7 2AZ USA
UK
Dr. Joakim Westerlund
Munawar A. Shaik Åbo Akademi University
Indian Institute of Technology Centre for Industrial Engineering and
Department of Chemical Engineering Management
Delhi Faculty of Technology
Hauz Khas FI-20500 ÅBO
New Delhi – 110 016 Finland
India
Professor Tapio Westerlund
Dr. Rui T. Sousa Åbo Akademi University
Imperial College London Centre for Industrial Engineering and
Centre for Process Systems Engineering Management
Department of Chemical Engineering Process Design Laboratory
South Kensington Campus FI-20500 Turku
London SW7 2AZ Finland
UK
Professor B. Erik Ydstie
Professor Rajagopalan Srinivasan Carnegie Mellon University
Institute of Chemical and Engineering Department of Chemical Engineering
Sciences 5000 Forbes Avenue
1 Pesek Road, Jurong Island Pittsburgh, PA 15213-3890
Singapore 627833 USA
Singapore
and
National University of Singapore
Departement of Chemical &
Biomolecular Engineering
4 Engineering Drive 4
Singapore 117576
Singapore
 1

1
Supply Chains of High-Value Low-Volume Products
Rui T. Sousa, Nilay Shah, Lazaros G. Papageorgiou

This chapter addresses supply chain modeling and planning of large process indus-
tries of products with a high added value per mass unit, where transportation costs
have a reduced impact in the total expenditure in supply chain activities, from raw
material supply until distribution to final markets. Typical examples of this kind of
industry are pharmaceuticals, agrochemicals, and chemical specialties in general.
A comprehensive review of relevant works in supply chain analysis and optimiza-
tion for the process industries as well as case studies is presented, with a special fo-
cus on supply chain management under uncertain conditions. This family of prob-
lems generally gives rise to large mixed integer linear programming models that
require the development of tailor-made solution algorithms to be solved in realistic
timescales. Traditionally, these methods have been based on hierarchical and/or
decomposition procedures aiming to reduce the number of integer variables in the
original model or its separation into several smaller subproblems faster to solve.
The chapter is illustrated with a case study for a generic large pharmaceutical
company described in generic terms.

1.1
Introduction

A supply chain, in the broad sense, may be defined as an integrated process where
several business entities, not necessarily from the same company, work together
to produce goods, services, and place them in their final customers, which in turn
may be the first participants of another supply chain [1]. This is a major area of
study in many industries, as organizations begin to appreciate the criticality of
creating an integrated relationship with their suppliers and customers. Typically,
in process industries the stages are: raw materials acquisition, primary (and sec-
ondary) manufacture, and distribution to retailers and customers. Each stage may
comprise one or more substages and products may be kept in storage facilities (e.g.,
warehouses) between stages (for a more detailed definition, see Ref. [2]).

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
2 1 Supply Chains of High-Value Low-Volume Products

Many companies, especially multinationals, possess complex trading structures

where, for tax purposes, the manufacturing plants, intellectual property, and distri-
bution centers are considered to be different entities. This brings more flexibility
to supply chain optimization as it allows the adoption of several different pricing
policies. Within these companies, we focus on the high-value low-volume process
industries, represented by fine chemical producers, pharmaceutical and agrochem-
ical companies, or, more appropriately, corporations.
Pharmaceutical and agrochemical supply chains comprise, generally, two manu-
facturing stages: (i) active ingredient (AI) production and (ii) formulation and pack-
aging. Pure AIs are very high-value products, produced in relatively low amounts
compared with the total volumes processed by the company, which makes their
transportation costs of very low significance in the total supply chain expenditure.
This allows companies with global (international) networks to maintain centralized
AI production in one or two locations worldwide and from there supplies the for-
mulation sites (internal AI customers) and any eventual external AI customer. In
the formulation stage, the AI is mixed with several excipients and other functional
materials in order to improve its action and to allow its correct administration.
This usually results in products with a low AI content and a significantly reduced
price per unit (by comparison with the pure AI), increasing the weight of trans-
portation costs. In general, each company possesses a network of formulation and
packaging sites meant to supply regional markets, although product flows between
geographic regions, despite not being common, may occur as well, as long as trade
regulations allow it. Many companies opt to outsource the formulation stage to
local companies when the regional market value/volume does not justify the in-
stallation of dedicated production facilities or transportation from a secondary site
in another region.
The product portfolio of these companies is usually very large due to all the pos-
sible formulations (excipient, concentration, mixture of several AIs) and packaging
(liquid, powder, tablets, capsules, and so on) available to process the AI, which is
translated into a high number of SKUs. Most of the manufacturing consists of
batch processes in multiproduct assets, resulting in the need for both campaign
planning at individual sites and global network coordination. For this reason, it
is also difficult, if not impossible, to separate the supply chain of individual prod-
uct families, as these always share common resources. If one is to try to apply an
enterprise wide optimization (EWO) approach [3] to one of these companies, in
general, this will result in a large-scale mixed integer linear programming (MILP)
model. However, aggregated planning is always a possibility, as pointed out by
Grossmann [3], that frequently results in an optimistic plan, which is hard to ac-
complish at the scheduling level. The situation becomes even more complex if we
take into account the fragmented nature of the orders placed by customers. These
tend to be large in number and low in volume, since in some cases customers are
not willing to keep significant stocks of the final product, making supply chain re-
sponsiveness a key issue on top of the multilevel planning integration difficulties
(Fig. 1.1).
 1.1 Introduction 3

Large portfolio Large MILP for Aggregated information on high

Complex manufacturing long term planning levels of planning
systems

Difficulty in performing scheduling according with planning Wise planning required

Lower responsiveness and capacity levels Information feedback

Fig. 1.1 Interactions and trade-offs in supply chain planning in large systems.

Total costs

Inventory
carrying costs
Cost

Lost sales costs

0
0 Average inventory level 100%

Fig. 1.2 Trade-off between responsiveness (lost sale costs) and inventory costs.

Classically, responsiveness problems in the process industries have been tackled

by maintaining large stock levels in the whole supply chain (30–90% of annual
demand [1]). However, this increases the working capital costs, particularly if the
value of stored products is high (Fig. 1.2).

1.1.1
Pharmaceutical Industry

In the past 30 years the operating context of the pharmaceutical industry has
evolved and become much more challenging. The establishment of regulatory au-
thorities led to an increase in the costs and time to develop new drugs, decreasing
the productivity of the research and development (R&D) stage and shortening the
effective patent lives of new molecules. These two factors, in conjunction with the
appearance of many substitute drugs in several therapeutic areas, have led to the re-
duction of the exclusivity period of new products. Another factor having an impact
on the operation of this industry was the transition of the paying responsibilities
from individuals to governmental agencies and insurance companies, which, in
association with high demands for pharmaceuticals, due to aging populations, put
strong pressure on prices and prescription policies [4].
From the point of view of manufacturing, the global pharmaceutical industry can
be divided into five subsectors: large R&D-based multinationals, generic manufac-
turers operating in the international market, local companies based in only one
4 1 Supply Chains of High-Value Low-Volume Products

country, contract manufacturers without their own portfolio, and biotechnological

companies mainly concerned with drug discovery. The first group, the intensive
and R&D-based industries, is economically the most important and tends to have
large and complex supply chains due to the global nature of its activity. In addition,
these companies are the most vulnerable to the global financial, regulative, and
social changes so this chapter will focus on their supply chains.
The industry’s preferred mechanism to overcome the productivity crises has
been to increase investment in current business activities, primarily R&D and
sales, the two extreme ends of the supply chain. This has been implemented by or-
ganic growth or by mergers and acquisitions (M&A) to exploit economies of scale.
However, statistics show that productivity has continued to decline after a decade
of vigorous growth in investments on these areas (www.contractpharma.com/
June021.htm). There are no significant economies of scale in sales activities. The
revenues generated by a pharmaceutical company are directly proportional to
its sales, general, and administrative expenditure (S, G&A), suggesting that two
merged companies may not be more profitable than they would be separately, that
is there is no improvement on the return rates. Furthermore, despite the theoreti-
cal higher probability of successful product development with greater scale in R&D,
in reality it does not translate into improved pipeline value. Companies will only
improve their profit margins if they change the relationship between volume and
costs, which can only be achieved through productivity gains in the supply chain.

1.2
Literature on Supply-Chain Modeling and Optimization

1.2.1
Case Studies

Some examples show the potential benefits of the projects in supply chain opti-
mization, an excellent way to increase profit margins. Nowadays, this is current
practice not only in the pharmaceutical industry but also in other areas of busi-
ness. Despite the large body of work in supply chain planning using modeling
approaches, and this being a mature area of research with capabilities to deal with
real world problems [5], some of the case studies found in the literature still use
simple, though robust, approaches with very valuable results.
Laval et al. [6] present a supply chain redesign of the Image and Printing Group
(IPG) of Hewlett-Packard in Europe, Middle East, and Africa (EMEA region). To re-
duce supply chain costs and improve its operations, the IPG wanted to reduce the
number of its contract manufacturing partners in the EMEA region. Several po-
tential locations were considered in the manufacturing network redesign, without
distinguishing between existing and new ones (i.e., without considering the costs
to establish new/changing manufacturing contracts) in order to avoid the biases
toward the existing configuration and to not eliminate any good “out-of-the-box”
solutions. The customer areas and the candidate locations to establish the new
 1.2 Literature on Supply-Chain Modeling and Optimization 5

manufacturing sites were aggregated in groups to reduce the size of the problem
and make it tractable. Three different channels of delivery to final customers were
considered. The authors develop an approach combining optimization and simu-
lation. The optimization stage is used to decide which scenarios are to consider
in greater detail in the simulation stage. The optimization model helps in under-
standing the supply chain from a global perspective and proves to be a useful tool
in assessing new complex situations where past experience and empirical knowl-
edge based on it are not of great help. The simulation stage is then used to “enrich”
and analyze the suggested scenarios, incorporating expertise features absent in the
optimization model. The final recommendations promised to save the company
US$10 million in supply chain costs and maintain the existing service levels.
A restructuring of the polymers & resins (P&R) business area at Rohm and Haas
is performed by D’Alessandro and Baveja [7]. Prior to the study, there was no orga-
nized order propagation scheme in the supply chain, giving rise to time-consuming
contacts between customers and manufacturing members of the supply chain and
causing scheduling plans to be constantly reviewed and plants to permanently oper-
ated on an uncertain environment. Strategic customer focus was also absent from
the management policies. An analysis of the customer database revealed that the
company served more than 4000 customers with the same level of service, even
though sales and customer service organizations recognized that some customers
were more important than others. In the beginning of the study, products were
divided into two groups according to their sales volume and demand variance,
which were found to correlate (high demand → low variance, low demand → high
variance). The manufacturing process was then reorganized so the high-demand
products were produced under a make-to-stock policy (i.e., stocks were kept within
prespecified margins) rather that in reaction to individual orders as used to be the
case. This allowed the company to plan 82% of its production volume in a stable
environment, with a smaller number of orders to process and fulfill. This group of
products also had priority in production planning, and the corresponding strategic
high-volume customers where guaranteed a faster delivery time which the com-
pany was now able to achieve. The other class, of low-volume high-variance prod-
ucts, started to be produced in a make-to-order scheme and longer lead times for
these products where arranged with the respective, less important, customers. This
enabled reductions in the stocks of raw materials and final products while fitting
the production plan around the plan of the first group of products. The assessment
of the lead times for the second group of products was made with a simulation
model. The changes saved millions of dollars for the company and transformed
the business into a leaner unit.
Jones et al. [8] developed a model for production planning under uncertainty of
the corn seed supply chain at Syngenta. The problem is characterized by the need
of establishing the production plan of all the hybrid corn seeds for the season with-
out knowing the actual demand levels as well as the production yields of the cul-
tures, dependent on climatic conditions. A second production season is available
sixth month after the first one, in the Southern Hemisphere, when the demand un-
certainties and production yield of the first season are already known. The uncer-
6 1 Supply Chains of High-Value Low-Volume Products

tain conditions are handled through a dynamic programming framework, which is

then reformulated in a stochastic programming approach with a discrete number
of scenarios. The planning of each product is independent of all the others as it is
considered that there is no limitation on the field area for the seed production (no
resource constraints). A test with data from the year 2000 showed that the model
would have saved the company US$5 million. Today Syngenta uses the model to
plan 80% of the total seed-corn sales volume.
Other supply chain studies use more formal approaches, supported almost ex-
clusively by modeling. Arntzen et al. [43] describe the restructuring of the supply
chain at Digital Equipment Corporation with savings of over US$100 million. They
developed a large MILP that incorporates a global, multiproduct bill of materials
(BOM) for supply chains with arbitrary structure and a comprehensive model of
integrated global manufacturing and distribution decisions.
Kallrath [44] reports on a project in BASF where a multisite, multiproduct, mul-
tiperiod production/distribution network planning model was developed, aiming
to determine the production schedule in order to meet a given demand. Besides
financial savings, this program allowed the client to handle the complexity of the
planning problem in a comprehensive way.
Camm et al. [45] describe a project related to P&G’s supply chain in North Amer-
ica. The main objective of the study was to streamline the work processes to elim-
inate nonvalue-added costs and duplication. The study involved hundreds of sup-
pliers, over 50 product lines, 60 plant locations, 10 distribution centers, and hun-
dreds of customer zones. It allowed the company to save US$200 million before
taxes.

1.2.2
Deterministic Studies

Supply chain management in the process industries has long been used as a tool
to define production and distribution policies as well as product allocation to man-
ufacturing resources. Planning problems may range from short- to mid-term pro-
duction and sales planning to long-term strategy delineation, contemplating in-
vestments such as capacity expansion and disinvestments concerning plant clo-
sures [5]. Models may be steady state or multiperiod, in which case time is generally
described by a discrete, indexed, set of time periods whose durations add up to the
total horizon. Multiperiod models can accommodate properties (parameters) and
decisions (variables) that change throughout time as well as inventory carryover
between time periods.
Supply chain planning problems are described by:
• sets of locations defining the nodes of the network,
comprising existing or potential suppliers, manufacturing
plants, storage units, and customer areas;
• product portfolios, comprising raw materials, intermediates,
and final products;
• a specific time horizon (in the case of multiperiod models);
 1.2 Literature on Supply-Chain Modeling and Optimization 7

• specific production recipes;

• sets of available distribution channels between sources and
destinations, more or less preestablished, in accordance with
the degree of flexibility allowed in the network;
• a financial structure with associated pricing policies and cost
structure; and
• in some cases, an available budget for investments may also
be defined.

Cohen and Lee [46] describe the modeling of a supply chain composed of raw
material vendors, primary and secondary plants (each one with inventories of
raw materials and finished products), distribution centers, warehouses, and cus-
tomer areas. The goal was to predict the impact of alternative manufacturing and
supply chain strategies on the performance at three levels: cost of products, ser-
vice level provided to the customers, and responsiveness and flexibility of produc-
tion/distribution systems. Later, Cohen and Moon [47] used supply chain optimiza-
tion to analyze the impact of scale, complexity (the operating costs are a function of
the utilization rates and number of products being processed in each facility), and
weight of each cost factor (e.g., production, transportation, and allocation costs) on
the optimal design and utilization patterns of the supply chain systems.
Wilkinson et al. [9] developed a long-term planning model for an international
supply chain with 3 manufacturing plants in 3 different countries and 15 market
areas with respective warehouses. The authors highlight the challenges posed by
integrated optimization on a global scale (e.g., large systems versus detailed plans
and large time horizons versus small timescales for the manufacturing operations)
where a trade-off between accuracy and model size and tractability is crucial. To
tackle the problem, an aggregated formulation is used to calculate the equipment
utilization at each manufacturing location and the overall supplies from these to
final destinations (customers). At the detailed level, each plant is planned in order
to reach the targets set by the upper level aggregated model. The aggregated for-
mulation is derived directly from the Resource Task Network (RTN) of the system
by aggregating the constraints of the detailed formulation of RTN over portions of
the planning horizon (aggregated time periods (ATPs)) (Wilkinson et al. [48]). This
produces a relaxation of the original problem in which the resource levels are only
determined at the boundaries of each ATP. One problem is left open, the integra-
tion of the two planning levels. A small deviation is indeed observed between the
production levels predicted by the upper level and the maximum feasible at the
detailed scheduling level, although the value was not significant. Other systems or
sets of data may require a refined approach at the top level to guarantee the feasi-
bility of the proposed solutions at the lower level.
McDonald and Karimi [10, 42] establish a deterministic mid-term planning prob-
lem for global supply chains, addressing the allocation of products to manufactur-
ing resources in multipurpose sites, often in different continents, accommodating
continuous, semicontinuous and batch processes. No capacity expansions are con-
8 1 Supply Chains of High-Value Low-Volume Products

sidered. The work aims to be a generic MILP formulation for multisite planning
over a mid-range time horizon of 1–3 years.
Bok et al. [11] tackle a short-term planning problem of a flexible manufactur-
ing network whose assets are distributed over several sites and may be dedicated
or multipurpose. The objective is the maximization of profits over the entire time
horizon punctuated by a set of orders and where costs and prices are allowed to
vary. The most interesting aspect of the work is the bi-level decomposition algo-
rithm used to solve the model and detailed in Section 1.1.
Timpe and Kallrath [12] develop a model for supply chain planning with two
levels of time resolution heterogeneously distributed. The unusual time descrip-
tion aims to reduce the variable space of the model while keeping a significant
level of detail in modeling the manufacturing operations. The higher level (lower
resolution) accommodates commercial events such as sales and inventory fluctua-
tions, while the second level (higher resolution) contains production events and is
embedded in the first one. Changeover operations between the possible modes in
which the manufacturing equipment may operate and source tracking of products
along the supply chain are also contemplated at the expense of a higher number
of variables. External purchase of final products is allowed to compensate capacity
limitations in the production stage.
Jayaraman and Pirkul [13] develop a capacitated plant location problem (CPLP)
type model for planning and coordination of production and distribution facilities
for multiple commodities, comprising raw materials suppliers, production sites,
warehouses, and customer areas. The authors follow a holistic approach to the
supply chain, resulting in a deterministic, steady-state, multiechelon problem. The
model must determine several decisions such as locations of the manufacturing
plants and distribution centers (warehouses), material flows between echelons, and
the distribution network for raw materials and goods in the process.
Jackson and Grossmann [14] built a multiperiod optimization model for the plan-
ning and coordination of production, transportation, and sales for a network of geo-
graphically distributed multiplant facilities supplying several markets. This model
is the top level of a global supply chain optimization problem that includes, on
the lower levels, site planning (production planning campaigns to meet customer
orders and inventory targets) and plant process optimization (process-operating
conditions). The model contains nonlinearities, introduced by the constraints de-
scribing each site in detail. The outputs of the global planning problem are the
decisions concerning how much of each product each site will produce, how much
of the demand will be fulfilled, and how the distribution network to final customers
is organized.
Several authors addressed the issue of supply chain optimization and long-term
process planning in the pharmaceutical industry. Rotstein et al. [15] started a se-
ries of papers dedicated to the specific problem of supply chain optimization in the
pharmaceutical industry. Given the candidates a product portfolio of a company,
they build a stochastic, scenario-based model that must choose which products
should be developed together with the consideration of production issues and ca-
pacity investment plans. Two years later, Papageorgiou et al. [16] published a paper
 1.2 Literature on Supply-Chain Modeling and Optimization 9

based on that of Rotstein et al., where the production stage is formulated with a
high degree of detail and the trading structure of the company is included. Their
model is deterministic and with up to eight possible products in the company’s
portfolio. Levis and Papageorgiou [17] extended this work to account for uncertain
demand forecasts, dependent on the results of the clinical trials for each product.
In the last years, some works have arisen in supply chain management with a
particular focus on the choice of suppliers. Yan et al. [18] use BOM logical con-
straints to define the relationship between suppliers and producers. The authors
use logical cuts to favor particular solutions that are known to be better based on
experience and, at the same time, provide a natural way of reducing the model size
(since the number of constraints is reduced) and accelerate the solution process.
Park et al. [19] present an extension of two generic supply chain planning mod-
els to include raw material purchase and final product sale contracts. These mod-
els are then solved by means of disjunctive programming (Grossman and Lee
[49]).
Temporal multilevel supply chain planning with integrated financial decisions
optimization may be found in Ref. [20].

1.2.3
Case Study Part I: Problem Statement

In this example, we aim to optimize the supply chain of a large pharmaceutical

company, whose structure is introduced in Figs. 1.3 and 1.4. The supply chain
components are primary sites (AI manufacturers) and respective storage facilities,
secondary sites and respective warehouses and final product market areas (Fig. 1.4).
Distribution networks within each market area are out of the scope of this example.

Europe
Sites
Markets &
Products

S. Amer. Asia
Sites Sites
Markets & Markets &
Products Primary Products
Sites &
Products

N. Amer. Afr. M. East

Sites Sites
Markets & Markets &
Products Products

Fig. 1.3 Supply chain structure of the enterprise.

10 1 Supply Chains of High-Value Low-Volume Products

t=n

Inventory
t=n 13 63 81
1 ... 62 ...
3 6 5

inventory
1 4,8 2 7 ... ... ... ...
2 9 1 10 13 Market

Sec. Products
3 ...
4 62 ...
5 ... Secondary
AI’s

Sites 63
6 t = n+m
64
7

Inventory
... 13 ... 81
t = n+m
8 63 ... 62
3 6 10 80

inventory
9 ... ... ...
8 2 7,4 81
10 9 5 1 Market

(a) (b)

Fig. 1.4 (a) Internal organization of the primary sites. The

allocation of products to sites may change through the different
time periods. (b) Each secondary geographic area only
produces and consumes a subset of secondary products from
the company’s portfolio. As in primary sites, the allocation of
products to sites may change in time.

Each primary site may supply the AI to any of the secondary sites and be lo-
cated in any place around the world. There are five geographic areas for secondary
sites and markets: Europe, Asia, Middle East and Africa, North and South America.
Since the transportation costs are very significant at this end of the supply chain,
material flows between two different geographic areas are not allowed.
The problem statement is summarized below:
• the multiperiod demand profile of the company’s portfolio;
• production, transportation, and allocation specific costs;
• tax rates; and
• the company’s supply chain structure, with primary and
secondary locations available.

The model aims to optimize decisions that supply chain managers usually face:
• allocation of the primary products (AIs) production to
primary manufacturing sites and how to manage the
available resources during the whole time horizon;
• primary production and inventory profiles;
• primary product flows to each secondary geographic region;
• allocation of secondary product manufacture to secondary
sites within each geographic area;
• secondary production and inventory profiles; and
• secondary product distribution network and flows from
secondary plants to markets.
 1.3 Stochastic Models 11

Each secondary geographic area produces and consumes some product fami-
lies, but not all, from the company’s portfolio. Secondary products with the same
formulation but different package, label and so on are considered to belong to dif-
ferent families. It was decided that a single sourcing policy would be followed, that
is each product (both primary and secondary) will be produced in one and only
one site during each time period. Allocation transfer (i.e., movement of production
allocations between sites) is allowed between different time periods although the
number of exchanges is limited. For secondary products, this transfer can only take
place between sites in the same geographic area.
The location of the production and distribution sites is an extremely important
issue for large multinational enterprises, as the profit after taxes may change sig-
nificantly due to the different taxes charged in different countries. So, the objective
is set as the maximization of the enterprises net profit value (NPV).The terms of
the objective function are listed below. A cost for unfulfilled demand has also been
considered.

Revenues Costs

Secondary products sales Raw material costs

10 AIs Production costs
10 primary sites Inventory handling costs
5 secondary geographic areas Unfulfilled demand costs
100 secondary product families Tax costs
70 secondary sites
10 market areas
12 time periods

The problem described above generates a model containing 84,096 binary vari-
ables.

1.3
Stochastic Models

From the modeling point of view, deterministic models are much easier to handle,
as they are simpler and have a smaller number of variables than their stochastic
counterparts. However, a decision maker who disregards the effects of uncertainty
in the operation parameters in planning problems is, on the one hand, adopting a
position potentially overoptimistic and, on the other hand, may be missing some
strategic opportunities over competitors. Numerous works in supply chain design
and operation prove that optimizing a deterministic case for the average values of
any uncertain parameters does not necessarily provide the solution that on aver-
age performs best over the entire range of possible combinations of outcomes of
the uncertain events. This is particularly true when performing long-term process
planning. Uncertainty factors in parameters such as demand forecasts and tech-
12 1 Supply Chains of High-Value Low-Volume Products

nological aspects have to be taken into account in order to produce robust plans
and/or designs.
There are simulation and optimization-based approaches for planning problems.
In this chapter, we shall focus on the latter. Sahinidis [21] and Guillén et al. [22] pro-
vide good summaries of the different optimization-based approaches available for
modeling and handle uncertainty. In this section, we provide a brief review of some
areas of work in supply chain design and planning under uncertainty. Sahinidis
identifies three main types of approaches: stochastic programming, dynamic pro-
gramming (for optimal control), and fuzzy mathematical programming. The first
two types are probabilistic in nature and have been widely used in supply chain
design and planning.

1.3.1
Stochastic Programming with Recourse

The first reference to this method goes back to Dantzig [50]. In stochastic program-
ming, the uncertain parameters are represented as random variables with a given
probability distribution. The deterministic version of the model is modified to re-
flect this probabilistic nature of some parameters. Integration over the probability
space of all parameters is then performed in order to obtain the expected value of
the objective function. Several methods are available to perform the integration
over the probability space as detailed later in this section.
Traditionally, the strategy has been the two-stage approach, with the separation
of decision variables in two different types:
• First stage or here-and-now decisions that have to be made
early in the time horizon, and so are independent of the
outcomes of the uncertain events.
• Second stage or wait-and-see decisions that are made in later
periods in the time horizon, as uncertainty unfolds.

The second stage is defined as an inner, recourse problem of the first-stage or outer
problem. An extension of this approach is multistage stochastic programming,
where uncertainty is considered to unveil gradually along the time horizon. The
beginning of each stage is then signaled by the occurrence of an uncertain event,
and its variables are linked to the variables of previous stages.
Often, the probability distributions of uncertain parameters are discrete, making
them suitable for a scenario approach, where each scenario will correspond to a
particular combination of the possible outcomes of all uncertain parameters. In the
case of continuous probability distributions in the uncertain parameters, specific
strategies (e.g., Monte Carlo) may be used to generate a discrete set of scenarios
that representatively covers the probability space. Under the scenario approach,
the problem is reformulated as an extended, large deterministic model. The model
size will grow exponentially with the number of scenarios, which in turn cannot be
too low if a representative coverage of the probability space is desired.
 1.3 Stochastic Models 13

For continuous distributions, a distribution-based approach is sometimes used,

where the probability functions are explicitly used in the problem formulation. The
integration is then performed using several possible techniques, including analyt-
ical resolution.
Stochastic programming has been one of the most popular approaches for mid-
to long-term planning optimization under uncertainty problems and one with a
large number of published works.

1.3.2
Stochastic Dynamic Programming (for Optimal Control)

The philosophy behind this approach states that independently of how we reach a
certain state somewhere in the time horizon, the decision process has to be optimal
for the tailing problem and made according to a decision rule or policy (stochastic
optimal control or Markov decision processes). Dynamic programming creates a
framework to solve these problems; first, the tail subproblems for the last stage
are solved, whose time span only covers a small portion of the total time horizon
(the last stage). In the following steps, more stages are gradually added to the tail
problem (or subproblems) in a backward temporal sequence, and at each step the
resolution process uses the solutions of the previous, shorter length, tail problems.
Dynamic programming suffers the “curse” of dimensionality, as the number of
tail problems grows exponentially in the number of state and control variables so
efficient algorithms to tackle individual subproblems are needed.
Comparing the two probabilistic approaches, stochastic programming and sto-
chastic dynamic programming, application to planning under uncertainty prob-
lems, Cheng et al. [23] stress the different focus and limitations of each strategy.
Stochastic programming deals with numerical procedures while optimal control is
mainly concerned with analytical solutions. The first approach becomes imprac-
ticable when the number of scenarios is too large while the second one is only
tractable for small state spaces.

1.3.3
Works in Planning Optimization Under Uncertainty

1.3.3.1 Stochastic Programming, Probability-Based Approach

Applequist et al. [24] tackle a supply chain design and management optimization
problem with risk measure for investments in supply chain using a financial ba-
sis, allowing the results to be readily comparable with other alternative financial
investments. Authors use a distribution-based approach to deal with uncertainty.
Polytope integration over the probability space provides the expected value of the
objective function, which comprises a trade-off between risk and return on invest-
ments, and its variance under uncertain demands.
Gupta and Maranas [25] develop a single-period, single-site planning model,
based on the mid-term planning model of Mcdonald and Karimi [10]. A two-stage
approach is used to determine a production plan (outer problem) and recursive
14 1 Supply Chains of High-Value Low-Volume Products

supply chain variables (sales and final inventories, inner problem), subject to un-
certain demand forecasts. Given the closed form of the inner problem through
LP duality, the probability space is integrated analytically allowing the redefinition
of the original model as a mixed integer nonlinear programming, exclusively in
terms of the first stage variables. This, in turn, is solved using a modified version
of the outer approximation algorithm (Duran and Grossmann, 1986 [51]). Gupta
and Maranas [26] extend the work is to the multisite, multiperiod case and to in-
clude inventory carryover between time periods and backlogs in orders satisfaction.
Ierapetritou and Pistikopoulos [27] tackle a plant design and operation under
uncertainty in two stages using a distribution-based approach. The authors use a
quadrature method to calculate the integral.

1.3.3.2 Stochastic Programming, Scenario-Based Approach

Liu and Sahinidis [28] develop a two-stage stochastic program for process planning
under uncertainty. They consider some or all of the parameters that determine the
economics of the production plan as being random; more specifically, they address
the case where the forecasts for prices, demands, and availabilities come in a finite
number of possible scenarios, each of which has an associated probability. A two-
stage, scenario generation approach is adopted in this work. This is a steady-state
model, whose formulation includes investment and capacity expansion constraints
(scenario independent), mass balances, purchases, and sales limitations (scenario
dependent). The objective is the maximization of the expected NPV. Since the re-
sulting model may assume a large number of scenarios, a decomposition algorithm
was developed to tackle it. In this algorithm, the original MILP model is solved
iteratively through a sequence of MILP master problems (upper bound) and LP
subproblems (lower bound).
Ahmed and Sahinidis [29] take the capacity expansion problem addressed by
them earlier [30] and extend it to deal with uncertain input parameters such as
product demands, variable and fixed costs of capacity acquisition, and costs to al-
locate capacity to products. This is a multistage approach where each stage cor-
responds to a single time interval of the horizon resolution. The scenario tree is
“shaped” using sets of nonanticipativity constraints that state that two scenarios
belonging to the same node in the scenario tree in the same time period must have
the same value for the decision variables. A heuristic procedure is developed to
solve the large MILP.
Iyer and Grossmann (1998) [35] extend the work by Liu and Sahinidis [28] to
the multiperiod case. The inputs are a set of available chemical processes, an es-
tablished network, and demand forecasts affected by uncertainty, leading to an
MILP, multiperiod planning model with multiple scenarios for each time period.
The model estimates decision variables addressing: capacity expansion of existing
processes and selection of new ones, their capacity expansion policy and produc-
tion, sales and raw materials purchase profiles in each time period and scenario.
The objective is the maximization of the expected profits.
Tsiakis et al. [31] address a strategic problem of planning for multiechelon sup-
ply chains with rigid structures. The model considers several products with both
 1.3 Stochastic Models 15

Fig. 1.5 Tree generation with potential clinical trial outcomes [32].

deterministic and stochastic demand forecasts, to determine warehouse (W) and

distribution center (DC) location from a set of possible candidates as well as the
production rates. Transportation costs are a function of the material flow. Uncer-
tainty is modeled through a two-stage scenario-based approach. Since the aim is
the design of a single structure for the supply chain, binary decision variables con-
cerning the location of W and DC are scenario independent (first stage decisions),
that is the model seeks a robust supply design over all possible scenarios.
Levis and Papageorgiou [17] introduce the uncertainty factor in the model de-
scribed in Papageorgiou et al. [16]. Uncertainty comes from demand forecasts,
which are dependent on the results of the clinical trials for each product. The au-
thors choose a scenario-based approach to solve the stochastic model and develop
a hierarchical algorithm to deal with the large resultant MILP.
Gatica et al. [32] address the issue of product development and uncertainty in
clinical tests outcomes, but from a different perspective compared to previous
works. They consider several pharmaceutical drugs in different stages of the devel-
opment process. Each trial has four possible degrees of success, which correspond
to four different demand forecasts for the product being tested. This multistage,
stochastic, multiperiod problem gives rise to a multiscenario MILP model. Each
time a product finishes its clinical tests, a new stage starts and each scenario of
the previous time period splits into four new ones. The objective function is the
expected NPV with a performance measure that takes appropriate account of risk
and potential returns. The authors test two different solution methods. The most
successful one consists of scenario aggregation in clusters to reduce the size of the
model. Usually, this aggregation will just occur in the last stages of the scenario
tree. The aggregated model is solved and then, the strategic variables calculated in
this step are fixed and the model is gradually disaggregated from the first to the
last stages, with intermediate optimizations until the whole desegregation process
is complete (Fig. 1.5).
16 1 Supply Chains of High-Value Low-Volume Products

Balasubramanian and Grossmann [33] address a multistage, multiperiod, batch

plant planning problem under demand uncertainty. At each stage m, the model
considers scheduling decisions for period m and sales decisions for period m −
1 to be deterministic, while sales of period m are defined recursively in the next
stage. The resulting large MILP is solved through a shrinking horizon strategy,
where an approximate two-stage model of the original problem is used. In each
step, stage m being optimized is considered to address “first-stage” decisions while
the remaining time periods are aggregated in a single second stage. In the next
step, stage m decisions are fixed, the time frame moves one stage and the whole
procedure is repeated.
The line of research in stochastic programming applied to this class of prob-
lems comes from Guillén et al. [22, 34]. Guillén et al. [34] develop a two-stage,
scenario-based model for supply chain design and retrofit under uncertain demand
forecasts. The first stage decisions refer to the supply chain design (locations and
capacities) while second stage is concerned with supply chain operation. The objec-
tive is to establish the multiobjective Pareto curve displaying the trade-off between
profitability and customer service level. A measure to limit downside financial risk
is included. Guillén et al. [22] extend the work to multistage stochastic program-
ming and develop a hybrid solution method, combining genetic algorithms (GA)
and mathematical programming. Details of the solution method are found in Sec-
tion 1.4.

1.3.3.3 Stochastic Dynamic Programming Approach

Cheng et al. [23] define a capacity planning and inventory control problem as a
multiobjective Markov decision addressing expected net profits and downside risk.
Uncertainty is introduced via demand forecasts. The authors develop a dynamic
programming framework to propagate the set of Pareto optimal solutions backward
in time (as described in Section 1.3.2). By using a simulation-based optimization
strategy to calculate the tailing problems, the dimensionality factor that usually
limits dynamic programming usage is overcome.

1.3.4
Case Study Part II: Planning Under Uncertainty

There are special cases where the combinations of all possible realizations for the
uncertain factors result in a relatively small number of scenarios that arise in a
natural way. This is the case when the uncertain parameters have a discrete prob-
ability distribution. In the pharmaceutical industry, this may be represented by
drugs for respiratory conditions whose demand levels are linked with the average
atmospheric conditions over a given period, or, alternatively, by the introduction of
new drugs, still in some stage of development and whose success is still unknown.
In this example, we develop the stochastic version of the problem described in
Section 1.2.3, assuming that the uncertainty comes from several families of drugs,
still in development stage, whose success of the clinical trials is still unknown.
 1.3 Stochastic Models 17

t1 to t4 t5 to t8 t9 to t12

1,2,3,4 1,2,3,4
J=1
t1 to t4 t5 to t8 t9 to t12
1 5,6,7,8 5,6,7,8
1,2
J =2 1 to 8 1 to 8
2

J=3 1 to 8 1 to 8
3
3,4
4 1,5
1ry sites 1,2,5,6
5 2,6
5,6 J=4
6 3,7
3,4,7,8
7 4,8
7,8
8 J=5 1 to 8 1 to 8

Fig. 1.6 Scenario tree for the two echelon supply chain problem
of Chapter 3. The scenario multiplication for the primary sites is
only due to the uncertainty in demands of the secondary
products. The double lines on the secondary areas refer to
scenario generation due uniquely to multiple scenarios on the
primary area.

Figure 1.6 presents the scenario tree for the problem. The uncertain parame-
ters are the demand forecasts of secondary products in several geographic regions
(zones 1 and 4 in Fig. 1.6). This will induce scenario duplication in these areas,
which will also generate new scenarios for the primary site planning problems.
These in turn will affect the remaining secondary areas, even if these latter ones
have deterministic demands. It is clear from the figure that the number of sce-
narios will grow very steeply with the number of uncertain parameters (up) in the
secondary areas and the possible realizations for each of them. On each stage (st),
considering that each uncertain parameter may assume two values, the number of
nodes is
N(st) = N(st − 1) × 2up(st) .
The model formulation is the same as in the deterministic case except that the
variables have an extra scenario index, k. The objective function is the weighted
NPV average over all scenarios k. Let Probk be the probability of each scenario. Let
Zk be the NPV for scenario k:


eNPV = probk Zk ∀k ∈ K.
k

1.3.4.1 Scenarios
Two families of products based on the same AI, still in development stage, are
planned to enter the markets in geographic area Europe (J1) in time period T4 (be-
ginning of the second stage). Two possible levels of success are assumed for these
products, high and low, with demand forecast levels of 150% and 50% of the deter-
ministic case, respectively. In the deterministic case, these product families account
for 35% of the AI consumption (in value) of geographic area Europe.
18 1 Supply Chains of High-Value Low-Volume Products
Table 1.1 Success probability for the clinical trials.

T4 T8

J1 High: 60% –
Low: 40% –
J4 High: 50% High: 30%
Low: 50% Low: 70%

Table 1.2 Scenario configuration and probability.

Scenario configuration J1.T4 J4.T4 J4.T8 Probabilities

Sc 1 H H H 0.09
Sc 2 H H L 0.21
Sc 3 H L H 0.09
Sc 4 H L L 0.21
Sc 5 L H H 0.06
Sc 6 L H L 0.14
Sc 7 L L H 0.06
Sc 8 L L L 0.14

For geographic area North America (J4), the company has one new family of prod-
ucts planned to start in time period T4 (20% of AI sales in North America, in value)
and a second one planned to start in T8 (15% of AI sales in North America, in value).
As for Europe, we assume two possible success levels, high and low, with demand
levels of 150% and 50% of the determinist case, respectively. Table 1.1 shows the
probability of each outcome of the development stages.
These leads to the configuration and probability for each scenario as shown in
Table 1.2.
The resulting MILP has more than 100,000 binary variables. Commercial solvers
using standard settings are unlikely to solve such a large integer problem in a real-
istic time. A tailored solution method is required.

1.4
Solution Algorithms

1.4.1
Decomposition and Hierarchical Algorithms

In order to keep the model size within reasonable limits, it cannot be too detailed
in its description of the supply chain. Nevertheless, even with this approach, it is
necessary to use decomposition algorithms.
 1.4 Solution Algorithms 19

Sahinidis and coworkers develop some solution algorithms based on hierarchical

and decomposition procedures. Liu and Sahinidis [28] adopt a bi-level decomposi-
tion strategy, with master and recourse problems. The objective of the master prob-
lem is to search for (or to propose) a set of optimal capacity expansion decisions.
These ones are then fixed in the subproblem, which in turn is responsible for the
calculation of the operational variables. The process is repeated iteratively until the
two solutions converge. The subproblem is formulated only with the mass balance
constraints and mass flow limitations, and the NPV does not include the produc-
tion costs. The master problem includes all the capacity expansion constraints and
a Benders cut. The calculation of the Benders cut, provided by the solution of the
subproblem, is estimated (but not necessarily exactly determined) through a Monte
Carlo sampling procedure. The method does not require the computation of the op-
timal solution of the MILP master problem (as long as the upper and lower bonds
do not cross), which significantly reduces the total CPU time.
Ahmed and Sahinidis [29] construct the solution procedure in three stages: in
the first one, the integral constraints are relaxed and the LP is solved. In the second
step, the nonanticipativity constraints are relaxed, originating as many indepen-
dent subproblems (SPs) as a number of scenarios and the relaxed binary variables
are perturbed, as described by Ahmed and Sahinidis [30], to provide an integer
set of the binary decisions. In the third step, nonanticipativity constraints are re-
inforced and the binary variables relating to the capacity expansion decisions are
set to their maximum values within all scenarios belonging to the same scenario
bundle (group of scenarios that have not differentiated yet among themselves at a
given stage). This ensures a suboptimal, good feasible solution.
To solve the large MILP of the original model (P), Iyer and Grossmann [35] use
a bi-level (hierarchical) decomposition. In the first step, a design problem (DP) is
solved, where the binary time dependent, investment decision variables are sub-
stituted by another set, time invariant, which only concerns the choice of the
processes to develop. DP has fewer binary variables than P, so is easier to solve.
Some of the P model constraints are relaxed or removed in DP, which means that
its solution will be an upper bound to P (the objective is a maximization). In the
second level model (operation model, OP), only the processes chosen to be devel-
oped will be subjected to investment. OP is similar to P but with extra constraints
and fewer binary variables to estimate since DP directly sets some of them to 0. OP
is faster to solve than P and provides a lower bound to the original problem. It can
occur that DP delivers an infeasible set of processes to develop to OP. In this case
an integer cut is added to DP and the whole process is repeated again. The solution
of OP provides integer cuts to DP and the problem is solved iteratively until the
upper and lower bonds converge.
Bok et al. [11] use a bi-level decomposition method inspired by Iyer and Gross-
mann [35]. An aggregated model, with a lower number of binary variables, is solved
to determine the deliveries to final customers but without providing the detailed
production plan. Since this model is a relaxation (RP) of the original problem, it
provides an upper bound to the optimal solution. The delivery plan is then fixed in
a second model of the subproblem defined by the solution of the first model. If SP
20 1 Supply Chains of High-Value Low-Volume Products

is infeasible, an integer cut is added to RP to eliminate the current solution from

the set of possible solutions of the first stage and the process is repeated. If SP is
feasible, its optimal solution provides a set of integer cuts to add to RP. Since SP
is the original problem constrained by a particular order delivery plan, its solution
will be a lower bound to the profit value of the optimal solution. The iterations
proceed until the difference between the upper and lower bounds falls within a
prespecified tolerance.
Levis and Papageorgiou [17] use a hierarchical algorithm to tackle the large MILP
stochastic model. In the first step, the authors solve an aggregated model, compu-
tationally less expensive, although detailed enough to make the “here-and-now”
decisions in their stochastic problem. In the second step, the values of the corre-
sponding variables are fixed and the detailed model is solved. In the first step, an
aggregated formulation is solved, where the manufacturing “suites” are not indi-
vidualized but aggregate. This model is computationally less expensive to solve due
to the lower number of constraints and variables; however, it is detailed enough to
make the “here-and-now” decisions. In the second step, these values are fixed and
the detailed model is solved in order to estimate the “wait-and-see” variables.

1.4.2
Hybrid Methods

Hybrid solution methods combine heuristic methods such as GA and Lagrangean

decomposition with mathematical programming.

1.4.2.1 Lagrangean Decomposition

Some authors solve the large models resulting from supply chain optimization
problems through Lagrangean decomposition. Gupta and Maranas [36] formulate
an extension of the “economic-lot-sizing” problem, characterized by determination
of the production levels of multiple products, in multiple sites, with deterministic
demands and multiple time periods. They use a decomposition scheme based on
Lagrangean relaxation to decompose the product portfolio into several, indepen-
dent, product groups. Jayaraman and Pirkul [13] relax three blocks of constraints
concerning assignment of customers to warehouses, raw materials availability, and
material flow balance. This allows them to decompose the original problem into
three different sets of SPs. Maravelias and Grossmann [52] introduce a good ex-
ample of a model composed of two (or more) independent submodels with one
linking constraint. Making use of this feature, the authors duplicate the linking
constraint variables and build a Lagrangean decomposition scheme, obtaining two
independent submodels. Jackson and Grossmann [14] adopt two Lagrangean de-
composition methods, spatial and temporal decomposition, to tackle the problem.
In the first approach, the mass flow variables (between echelons) are duplicated and
the corresponding equality constraint is relaxed through the inclusion of penalty
factors (Lagrangean multipliers) in the objective function. This generates indepen-
dent NLP and LP problems for each site and market, respectively, whose solution
allows obtaining an upper bound to the original problem. In the temporal decom-
 1.4 Solution Algorithms 21

position, the inventory variables are doubled and a similar procedure to the one
described above is performed. This decomposes the main problem into several in-
dependent SPs, one per each time period. In both cases, the authors follow the
regular algorithm of Lagrangean decomposition to reach the optimal solution of
the original problem (as described in Reeves [53]). The numerical examples show
that temporal decomposition works significantly better than spatial decomposition.
According to the authors, this is due to several factors: first, there are fewer tem-
poral interconnection variables and second, since the entire production model is
solved for each time period, all the solutions provided are feasible, which is not
necessarily the case with the spatial decomposition, where the mass balances are
disrupted.

1.4.2.2 Genetic Algorithms

Guillén et al. [22] combine the solution strategy GA and mathematical program-
ming algorithms in two levels. At the upper level (master problem), GA are imple-
mented to manage the DP and generate potential supply chain structures. At the
inner level (slave problem), a two-stage stochastic scheduling model is computed
within a rolling horizon strategy [33] to yield scheduling decisions at all nodes of
the scenario tree and provide an estimate of the expected profit.
Zhou et al. [37] solve an allocation problem where costumers are assigned to
warehouses. They use a multicriterion objective function where the performance
indicators are the shipping cost and the transit time between warehouses and their
customers. With regards to the success in GA application in previous allocation
problems, the authors adopt this approach to tackle the problem. Each chromo-
some is an array with dimension equal to the number of customers (genes). Each
position of the vector can assume an integer value that ranges between one and the
number of warehouses (alleles). This way of representing the solutions ensures
that each customer will be assigned to one and only one warehouse.
Vergara et al. [38] develop an evolutionary algorithm to find the production se-
quence at each supplier for multicomponents and a synchronized delivery cycle
that minimizes the costs of a simple multistage supply chain. The authors com-
pare the performance of the evolutionary algorithm with an enumeration method
for the synchronization of supply chains, and conclude that the latter one soon
becomes intractable as the problem grows. Chan et al. [39] study the issue of col-
laboration in supply chain management and propose the constitution of a Central
Coordinating System that considers collaboration barriers such as information and
benefit sharing and joint decision making. The SC is composed of manufacturing
sites, warehouses, and final customers. The authors use a hybrid method where
solutions are generated and optimized through a GA while the fitness values are
estimated via an analytical hierarchy process (AHP). Jang et al. [40] build a supply
management system with four modules: (i) a supply network design optimization
module, (ii) a production/distribution-planning module (from raw materials sup-
pliers to costumers), (iii) a model management module, and (iv) a data manage-
ment module. The second module generates the production and distribution plan
for the supply network, synthesized on the first module, over the planning hori-
22 1 Supply Chains of High-Value Low-Volume Products

zon. An integrated model, consisting of three types of multistage, multiproduct,

and multiperiod planning submodels, represents the module. This problem is dif-
ficult to solve due to the large number of integer variables, thus a GA approach
is adopted. Integer values are used for both integer and continuous variables. The
results obtained by the authors concerning the comparison of GA versus mathe-
matical programming are not very conclusive, but there is a trend for the GA to
become faster than the other approach as the dimension of the model increases.
Jaramillo et al. [41] evaluate the performance of GA as an alternative procedure to
reach optimal or near optimal solutions of several generic location problems: un-
capacitated and capacitated fixed charge problem, the maximum covering model,
and medianoid and centroid models. The results obtained in this study are not
very conclusive; the GA performance relative to other heuristic methods is depen-
dent on the kind of problem being solved. Particularly for capacitated models, the
method performs very poorly.

1.4.3
Case Study Part III: Solution Methodology

None of the full variable space models resulting from the problem statement in
Parts I and II is solvable in 250,000 s of CPU time. A decomposition method is
therefore developed to tackle both the deterministic and stochastic problems intro-
duced in Parts I and II. The method has two steps. In the first step, the constraints
containing the inventory balances are relaxed and a variable substitution is per-
formed in the allocation constraints in order to decompose the problem into multi-
ple subproblems, one per time period of the time horizon. A set of binary variables
is calculated for each time period. The subproblems are solved in chronological se-
quence. For both models, a new set of parameters transfers information between
the submodels in order to guarantee that the number of allocation changes be-
tween consecutive time periods is not higher than the maximum value established
by upper management. In addition to this, for the stochastic model, a set of “long-
range information propagation” constraints ensures the integrity of the scenario
tree structure. In the second step, the original model is run with the binary vari-
ables fixed at the values calculated in the previous stage to determine the optimal
value of the continuous variables.

1.4.3.1 Solution Algorithm Performance

As mentioned in Section 1.3, in the case of the stochastic model, the temporal
decomposition algorithm (Fig. 1.7) was applied in conjunction with the nonantici-
pativity constraints approach and following the inverse chronological sequence. An
uncertain event in a given time period will not affect the decisions taking place in
the subsequent periods; however, these belong to the “wait-and-see” variable group,
so they may be reoptimized in later stages of the project. On the other hand, the
first time period(s) decisions, the “here-and-now” group, will be affected by all the
events taking place at later periods, which is in accordance with the desired goal
 1.4 Solution Algorithms 23

Start

t* = 1
(t* = T)

Transfer information Solve stage 1

t*-1 t* submodel for t = t*
(t*+1 t*)

t* = t+1
(t* = t-1)

n
t* > T
(t* < 1)

Fix stage 1 binary

variables

Solve stage 2 model

End

Fig. 1.7 Temporal decomposition algorithm.

Table 1.3 Temporal decomposition applied to the stochastic model.

Z opt Gap (%) CPU (s) ETa (s)

Deterministic (Part I) 1,002,610 5.8 126 –

Deterministic (Part II) (scenario average) 1,206,440 4.9 – –
First step (allocation1) 956,742 – 892 547
Second step (allocation2) 1,235,890 6 720 113

a ET, execution (generation + preprocessing) time.

of the project, to determine a robust set of decisions that performs well under all
scenarios.
Table 1.3 shows the results of the application of temporal decomposition to the
deterministic and stochastic models.
24 1 Supply Chains of High-Value Low-Volume Products

1.5E+06

1.4E+06

1.3E+06

1.2E+06

1.1E+06
NPV (m.u.)

1.0E+06

9.0E+05

8.0E+05

7.0E+05

6.0E+05

5.0E+05
K1 K2 K3 K4 K5 K6 K7 K8
scenario

s tochastic deterministic

Fig. 1.8 NPVs obtained with different sets of first stage

variables, obtained with the stochastic model and the
deterministic model for the mean scenario.

The decomposition method allows good solutions to be obtained, within less

than a 6% optimality gap, for both models. The deterministic model took 126 s to
solve, with a negligible execution time. The large stochastic model was solved in
less than 40 min of CPU time accounting for both execution and solution tasks.
Note that the decomposition method allowed the solution of a stochastic problem
whose deterministic version model had over 84,000 binary variables.

1.4.3.2 Financial Risk

Figure 1.8 shows the NPV obtained for each scenario fixing the first stage variables
obtained with the stochastic model (temporal decomposition) and the deterministic
case for the mean scenario.
One important output of stochastic modeling is the assessment of financial risk
incurred by the supply chain, that is the trade-off between the minimum guar-
anteed profits and the risk exposure of the assets. Figure 1.9 shows the scenario
histograms, where the scenarios are sorted according to their corresponding NPV,
which in turn is plotted against the accumulated probability over all scenarios with
lower objective function values. This provides a measure of the minimum proba-
bility of reaching a lower profit level at any point in the curve.

1.4.3.3 Comments
In this chapter, we formulated a stochastic version of the pharmaceutical supply
chain allocation problem. The supply chain has five semiindependent geographic
areas, linked by the AI production process creating an unusual scenario tree struc-
ture. This is a (temporal) multistage stochastic planning problem with binary vari-
ables to be determined in all the stages and with a strong link between them along
 1.5 Summary 25
1

0.9

0.8

0.7

0.6
prob

0.5

0.4

0.3

0.2

0.1

0
9.0E+05 1.0E+06 1.1E+06 1.2E+06 1.3E+06 1.4E+06 1.5E+06
NPV (m .u.)

s tochas tic determinis tic

Fig. 1.9 Scenario histogram for first stage decisions obtained

with the stochastic model and the deterministic model for the
mean scenario.

the time periods as the number of changes in production allocations between con-
secutive time periods is limited. This was the basis to create an adapted version for
the stochastic case of the temporal decomposition algorithm through the introduc-
tion of “long-term allocation control” constraints.
The weighted average over all scenarios of the NPV obtained with the determin-
istic model differs by only 2% of the eNPV obtained with the stochastic model. In
part, this is explained by some similarity between the production sites in region
Europe (where the difference between scenarios is larger) and some slack capacity
in the primary sites, reducing the weight of the allocation decisions in the levels of
demand fulfillment. Nevertheless, with the stochastic approach, the risks of having
profits lower than 980,000 units are reduced by ∼8%.

1.5
Summary

Mathematical models for supply chain management and planning provide valu-
able tools to support the decision-making process and to improve the competitive-
ness of chemical process companies. Supply chain studies have evolved from deter-
ministic studies, where all the information is assumed to be accurately known to
approaches where uncertainty (e.g., demand forecasts) is also taken into account,
allowing more robust solutions and reduced risk in supply chain planning.
Usually, supply chain planning problems give rise to large MILPs that are diffi-
cult to solve by direct application of commercial solvers. Hence, an extensive body
of literature is available addressing several solution methods either based in pure
mathematical programming techniques or hybrid methods combining mathemat-
26 1 Supply Chains of High-Value Low-Volume Products

ical programming and heuristic procedures. So far, none of the methods described
has been distinguished as a standard procedure, nevertheless, decomposition and
hierarchical algorithms have consistently provided good results.
In this chapter, we presented a case study relevant to the pharmaceutical indus-
try concerning production allocation to manufacturing sites in a worldwide net-
work. Both deterministic and stochastic cases were addressed and compared. The
introduction of uncertainty in the model parameters allows making early stage al-
location decisions that on average perform better for the possible realizations of
the uncertain events, providing a more robust plan (when compared with the pure
deterministic approach). The large MILP models generated were efficiently solved
using a temporal decomposition algorithm.

Acknowledgments

Author Rui T. Sousa acknowledges financial support from the Portuguese agency,
Fundação para a Ciência e a Tecnologia, through fellowship POCTI, SFRH/BD/
12177/2003.

References

1 Shah, N., Comput. Chem. Eng. 29(6) 10 McDonald, C. M., Karimi, I. A.,
(2005), pp. 1225–1235. Ind. Eng. Chem. Res. 36(7) (1997), pp.
2 Papageorgiou, L. G., Supply Chain 2691–2700.
Management and Optimisation. In: 11 Bok, J. K., Grossmann, I. E., Park, S.,
Computer-Aided Process and Product Ind. Eng. Chem. Res. 39(5) (2000), pp.
Design, Vol. 2, Puigjaner, L., Heyen 1279–1290.
G. (eds.). Wiley-VCH, Wienhiem, 12 Timpe, C. H., Kallrath, J., Europ. J.
2006, pp. 621–641. Oper. Res. 126(2) (2000), pp. 422–435.
3 Grossmann, I., AIChE J. 51(7) (2005), 13 Jayaraman, V., Pirkul, H., Europ. J.
pp. 1846–1857. Oper. Res. 133(2) (2001), pp. 394–408.
4 Shah, N., Comput. Chem. Eng. 28(6/7) 14 Jackson, J. R., Grossmann, I. E., Ind.
(2004), pp. 929–941. Eng. Chem. Res. 42(13) (2003), pp.
5 Kallrath, J., Or Spectrum 24(3) 3045–3055.
(2002), pp. 219–250. 15 Rotstein, G. E., Papageorgiou, L. G.,
Shah, N., Murphy, D. C., Mustafa,
6 Laval, C., Feyhl, M., Kakouros, S.,
R., Comput. Chem. Eng. 23 (1999), pp.
Interfaces 35(3) (2005), pp. 238–246.
S883–S886.
7 D’Alessandro, A., Baveja, A., Inter-
16 Papageorgiou, L. G., Rotstein, G. E.,
faces 30(6) (2000), pp. 1–16.
Shah, N., Ind. Eng. Chem. Res. 40(1)
8 Jones, P.C., Kegler, G., Lowe, T. J., (2001), pp. 275–286.
Traub, R. D., Interfaces 33(1) (2003), 17 Levis, A. A., Papageorgiou, L. G.,
pp. 80–90. Comput. Chem. Eng. 28(5) (2004), pp.
9 Wilkinson, S. J., Cortier, A., Shah, 707–725.
N., Pantelides, C. C., Comput. Chem. 18 Yan, H., Yu, Z. X., Cheng, T. C. E.,
Eng. 20 (1996), pp. S1275–S1280. Comput. Oper. Res. 30(14) (2003), pp.
2135–2155.
 References 27

19 Park, M., Park, S., Mele, F. D., 37 Zhou, G. G., Min, H., Gen, M., Int. J.
Grossmann, I. E., Ind. Eng. Chem. Prod. Econ. 86(1) (2003), pp. 35–45.
Res. 45(14) (2006), pp. 5013–5026. 38 Vergara, F. E., Khouja, M.,
20 Guillén, G., Badell, M., Espuna, A., Michalewicz, Z., Comp. Ind. Engin.
Puigjaner, L., Comput. Chem. Eng. 43(3) (2002), pp. 407–421.
30(3) (2006), pp. 421–436.
39 Chan, F. T. S, Chung, S. H., Wad-
21 Sahinidis, N. V., Comput. Chem. Eng. hwa, S., Int. J. Prod. Res. 42(1) (2004),
28(6/7) (2004), pp. 971–983. pp. 1–19.
22 Guillén, G., Mele, F. D., Espuna, A.,
40 Jang, Y. J., Jang, S. Y., Chang, B. M.,
Puigjaner, L., Ind. Eng. Chem. Res.
Park, J., Comp. Ind. Eng. 43(1/2)
45(22) (2006), pp. 7566–7581.
(2002), pp. 263–281.
23 Cheng, L. F., Subrahmanian, E.,
Westerberg, A. W., Ind. Eng. Chem. 41 Jaramillo, J. H., Bhadury, J., Batta,
Res. 43(9) (2004), pp. 2192–2208. R., Comp. Oper. Res. 29(6) (2002), pp.
24 Applequist, G. E., Pekny, J. F., 761–779.
Reklaitis, G. V., Comput. Chem. Eng. 42 Karimi, I. A., McDonald, C. M.,
24(9/10) (2000), pp. 2211–2222. Ind. Eng. Chem. Res. 36(7) (1997), pp.
25 Gupta, A., Maranas, C. D., Ind. Eng. 2701–2714.
Chem. Res. 39(10) (2000), pp. 3799– 43 Arntzen, B. C., Brown, G. G., Har-
3813. rison, T. P., Trafton, L. L., Interfaces
26 Gupta, A., Maranas, C. D., Comput. 25(1) (1995), pp. 69–93.
Chem. Eng. 27(8/9) (2003), pp. 1219– 44 Kallrath, J., Or. Spectrum 24(3)
1227. (2002), pp. 219–250.
27 Ierapetritou, M. G., Pistikopou-
45 Camm, J. D., Chorman, T. E., Dill,
los, E. N., Ind. Eng. Chem. Res. 35(3)
F. A., Evans, J. R., Sweeney, D. J., We-
(1996), pp. 772–787.
gryn, G. W., Interfaces 27(1) (1997),
28 Liu, M. L., Sahinidis, N. V., Ind. Eng.
pp. 128–142.
Chem. Res. 35(11) (1996), pp. 4154–
4165. 46 Cohen, M. A., Lee, H. L., Operations
29 Ahmed, S., Sahinidis, N. V., Oper. Research 36(2) (1988), pp. 216–228.
Res. 51(3) (2003), pp. 461–471. 47 Cohen, M. A., Moon, S., Europ. J.
30 Ahmed, S., Sahinidis, N. V., Pis- Oper. Res. 50(3) (1991), pp. 266–279.
tikopoulos, E. N., Comput. Chem. 48 Wilkinson, S. J., Shah, N., Pan-
Eng. 23(11/12) (2000), pp. 1589–1604. telides, C. C., Computers & Chemical
31 Tsiakis, P., Shah, N., Pantelides, Engineering 19 (1995), pp. S583–S588.
C. C., Ind. Eng. Chem. Res. 40(16) 49 Lee, S., Grossmann, I. E., Computers
(2001), pp. 3585–3604. & Chemical Engineering 27(11) (2003),
32 Gatica, G., Papageorgiou, L. G., pp. 1557–1575.
Shah, N., Chem. Eng. Res. Des. 81(A6)
50 Dantzig, G. B., Management Sciences
(2003), pp. 665–678.
1(3–4) (1955), pp. 197–206.
33 Balasubramanian, J., Grossmann,
I. E., Ind. Eng. Chem. Res. 43(14) 51 Duran, M. A., Grossmann, I. E.,
(2004), pp. 3695–3713. Mathematical Programming 36(3)
34 Guillén, G., Mele, E., Bagajew- (1986), pp. 307–339.
icz, M. J., Espuna, A., Puigjaner, 52 Maravelias, C. T., Grossmann, I. E.,
L., Chem. Eng. Sci. 60(6) (2005), pp. Industrial & Engineering Chemistry
1535–1553. 40(26) (2001), pp. 6147–6164.
35 Iyer, R. R., Grossmann, I. E., Ind. 53 Reeves, C. R., Genetic Algorithms and
Eng. Chem. Res. 37(2) (1998), pp. 474– Lagrangean Decomposition. Modem
481, Heuristics Techniques for Combinatorial
36 Gupta, A., Maranas, C. D., Ind. Eng. Problems, Reeves, C. R. (ed.). Black-
Chem. Res. 38(5) (1999), pp. 1937– well Scientific Publications, 1995.
1947.
 29

2
Solving Multiple Vehicle Pickup and Delivery Problems in
Multisite Systems by a Rigorous Optimization Approach
Rodolfo Dondo, Carlos A. Méndez, Jaime Cerdá

Transportation plays a central role in every supply chain (SC) because products are
rarely produced and consumed in the same location. In the general case, the sup-
ply chain comprises a set of facilities (production plants, suppliers, warehouses),
manufacturing, and/or storing the products, which need to be efficiently delivered
to a set of consumer nodes (retailers, stores, customers) in order to fulfill certain
requests. This work presents a novel MILP-based mathematical formulation fo-
cused on the operational level of multiple vehicle pickup and delivery problems
with time windows (MVPDPTW) commonly arising in multisite systems. Since
it is a NP-hard problem, most of the current solution approaches are of heuris-
tic type, thus providing good but not necessarily optimal solutions. The proposed
two-index model can be solved using a branch-and-cut commercial package to find
the best vehicle routes and schedules for moderate-size MVPDPTW problems. The
formulation has been generalized to also consider pure pickup and delivery nodes,
heterogeneous vehicles, multiple depots as well as many-to-many transportation
requests. The definition of allocation variables assigning requests rather nodes
to vehicles allows to drastically diminishing the number of 0−1 variables. Using
the notion of compatible requests, exact pruning rules aim to eliminate unneces-
sary variables and constraints and speed-up the MILP solution procedures have
been derived. Optimal solutions for a variety of benchmark problems featuring dif-
ferent sizes in terms of customer requests and vehicles, distinct cluster/random
pickup and delivery locations, and a range of time window width distributions are
reported. A MVPDPTW problem instance with many-to-many transportation re-
quests was also successfully tackled. From the results, it follows that medium-size
MVPDPTW problem instances comprising up to 25 customer requests, 50 pickup
and delivery locations, and 14 vehicles can be solved to optimality in quite reason-
able CPU times. Hybrid solution methods combining the proposed exact model
with heuristic-type procedures can be used to provide effective solutions for large-
scale problems.

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
30 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

2.1
Introduction

The pickup and delivery problem (PDP) has attracted a great attention from re-
searchers and practitioners in the last 20 years. It is a combinatorial optimization
problem aimed at satisfying a set of customer requests involving pickup and deliv-
ery tasks by means of a vehicle fleet at minimum total cost. Each customer request
specifies the size of the load to be transported, the locations where the goods are to
be collected (the origins), and the locations to which it has to be delivered (the des-
tinations). Each demand has to be fulfilled by a single vehicle transporting goods
from origins to destinations without any transshipment at other locations. To fulfill
the assigned transportation requests, every vehicle departs from the base and visits
a number of customers along the selected route. At each stop location, the vehicle
can either pick up or deliver an amount of load but not both. Loading and unload-
ing times are incurred at every stop. Moreover, each vehicle starts and ends its jour-
ney at the same depot. Vehicle data generally include the transportation capacity,
the alternative start/end locations (depots) for the journey, and the subset of trans-
portation requests that can be allocated to every unit. If the pickup and/or delivery
location has a time interval within which the service must begin, then the problem
is known as the pickup and delivery problem with time windows (PDPTW). Since
real-world pickup and delivery problems include some time restrictions, most of
the papers have been focused on the PDPTW. In particular, the PDPTW variant
involving transportation requests with a single origin and a single destination, and
a vehicle fleet departing and returning to a unique central depot was the most stud-
ied. The VRPTW is a particular case of the PDPTW in which the customer requests
all have either their origins (pickup version) or their destinations (delivery version)
at the central depot.
Three classes of PDPTW have usually been tackled. One of them is the so-called
single-vehicle pickup and delivery problem with time windows (1-PDPTW) where
the pickup and delivery services are done through a single vehicle. If there are
multiple vehicles available for the pickup and delivery services, then the problem is
known as the multi-vehicle pickup and delivery problem (MVPDPTW). Most of the
contributions have been devoted to these two PDPTW classes. In many practical
situations, however, the cargo must be collected from multiple nodes and trans-
ported to a single delivery location or vice versa. Furthermore, the problem may
involve open tours, a heterogeneous vehicle fleet, multiple depots, and each vehi-
cle may have its own starting/ending base. In addition, pure pickup nodes and/or
delivery nodes demanding just pickup or delivery services but not both can si-
multaneously be considered. A PDPTW problem with all these features is called
the general pickup and delivery problem with multiple vehicle and time windows
(GPDPTW) that so far has received a rather limited attention. Problems 1-PDPTW,
MVPDPTW or GPDPTW, all assume that the set of customer requests is known
beforehand and remains unchanged while the pickup and delivery services are be-
ing performed. They can be regarded as distinct variants of the static PDPTW. In
 2.1 Introduction 31

real-world problems, new customer requests become available in real-time and are
immediately eligible for consideration. As a consequence, the set of routes has to
be optimized at some point in time to include the new transportations services. It is
the dynamic PDPTW. In any case, the problem goal is to construct a set of optimal
routes and schedules for the vehicle fleet in order to fulfill all customer requests
at minimum value of the problem goal. Three different types of costs are usually
considered. First, the total vehicle fixed costs, including acquisition and mainte-
nance expenses, aimed at minimizing the number of used vehicles. Second, the
distance-based and/or the time-based transportation cost accounting for the fuel
consumption, vehicle maintenance, and driver wages. Third, the customer incon-
venience originated by pickups or deliveries performed either sooner or later than
the desired service time windows. The customer inconvenience is usually regarded
as a linear function of the time window constraint violations. Often, the authors
selected a weighted combination of the total distance and the total time required
to service all the customers jointly with the total customer’s inconvenience as the
problem objective to minimize.
Different practical problems can be modeled as a pickup and delivery problem.
Among them, the VRP with backhauls (VRPB), the dial-a-ride problem (DARP),
handicapped person transportation problems (HTP), and the courier company
pickup and delivery problems (CCPDP). The VRPB includes both a set of cus-
tomers to whom products are to be delivered (pure delivery nodes) and a set of
nodes whose goods need to be transported back to the distribution center. More-
over, all deliveries have to be made before all the pickups. Problems DARP and
HTP are concerned with the transportation of people while CCPDP deals with
transporting messages and parcels. A great deal of the papers on PDPTW is on
dial-a-ride problems. In contrast, fewer contributions have been focused on the
pickup and delivery of packages and goods.

2.1.1
Previous Heuristic Approaches

Surveys on the pickup and delivery problem can be found in Bodin et al. [1], Savels-
bergh and Sol [2], Fisher [3], Desrosiers et al. [4], and Desaulniers et al. [5]. Two types
of solution methodologies have been proposed: heuristic techniques and exact op-
timization approaches. Heuristic techniques can be classified into four groups:
construction, insertion, improvement, and metaheuristic methods. The best con-
struction heuristic methods are decomposition techniques based on the idea of
dividing the problem into three phases: clustering, routing, and scheduling. Sex-
ton and Bodin [6] applied Benders decomposition to a mixed 0−1 nonlinear pro-
gramming formulation of the single-vehicle many-to-many DAR problem. Dumas
et al. [7] applied a decomposition approach to the MVPDPTW based on the no-
tion of mini-clustering. A mini-cluster is a segment of a route along which the
demands from a small group of clients can be satisfied by a vehicle starting and
ending such a route segment with the same load or empty. Then, a mini-cluster
can be treated as an aggregate transportation request entirely satisfied by a sin-
32 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

gle vehicle. The optimization algorithm is then applied to a more compact set of
transportation requests, thus lowering the number of rows in the set partition-
ing matrix. Mini-clusters are constructed from known pickup and delivery plans
by cutting them into pieces such that each piece starts and ends with an empty
vehicle. On the other hand, insertion heuristics develop a set of routes by insert-
ing one request at a time into a given route. Two types of insertion procedures
have been proposed: sequential [8] and parallel [9] procedures inserting one cus-
tomer at a time into a single route or into one of several open routes. In turn,
improvement heuristic procedures start from a set of feasible routes and applied
some kind of local search algorithm to improve the current solution. A local search
algorithm examines all neighbor solutions to find the best one, with the notion
of neighbor depending on the procedure. Cyclic transfers [10] improved the solu-
tion by moving a number of requests amongst routes. Variable depth search tech-
niques [11] applied several arc-exchange mechanisms to improve the current so-
lution. More recently, metaheuristic methods, including simulated annealing [11],
constrained-direct search [12], tabu search [13, 14], threshold algorithms, neural
networks [15], and genetic algorithms have also been proposed. Li and Lim [16,
17] developed a tabu-embedded simulated annealing algorithm for the general
mutiple-vehicle PDPTW that restarts a search procedure from the current best so-
lution after several nonimproving search iterations. In addition, test cases were
generated from all of the Solomon’s benchmark instances for VRPTW. It is one
of the first heuristic approaches to solve large multiple-vehicle PDPTW problem
instances with various distribution properties. Tam and Kwan [18] generated a sys-
tematic scheme to adapt the Large Neighborhood Search (NLS) to efficiently solve
PDPTWs. The NLS is an iterative process of relaxation and re-optimization to con-
tinually improve on the existing routing plan until the convergence to a local min-
imum or resource exhaustion occurs. Their results compare favorably with those
obtained with Li and Lim’s metaheuristic search algorithm. Bent and Van Henten-
ryck [19] developed a two-stage hybrid algorithm for MVPDPTW. The first stage
uses a simple simulated annealing algorithm to decrease the number of routes,
while the second stage used Large Neighborhood Search (NLS) to decrease the
travel cost. New best solutions were found on problems with 50, 100, and 300 re-
quests.

2.1.2
Exact Optimization Methods

The development of exact optimization methods started in the early 1980s. One of
the earliest contributions was due to Psaraftis [20] who developed a dynamic pro-
gramming algorithm for solving the static single-vehicle many-to-many DAR prob-
lem with time windows. Generally speaking, the exact approaches can be grouped
into two classes: branch-and-price and branch-and-cut methods. Branch-and-price
techniques apply a branch-and-bound scheme in which lower bounds are com-
puted by column generation. In branch-and-cut methods, valid inequalities or cuts
are incorporated in the formulation at each node of the branch-and-bound tree to
 2.1 Introduction 33

tight the problem LP relaxation problems which are solved by the simplex algo-
rithm. [21] presented a dynamic programming approach to the large-scale single-
vehicle DAR problem with time windows, which proved to be very effective when
time windows are tight and vehicle capacities are small. Because of its intrinsic
complexity, a limited number of exact methods for the MVPDPTW has been pub-
lished. Dumas et al. [22] presented a technique based on the Dantzig-Wolfe de-
composition/column generation scheme and a pricing subproblem consisting on
a shortest path problem with capacity, coupling, precedence, and time window con-
straints. The approach optimally solves 1-PDPTW problems with up to 55 paired
requests and multiple vehicle PDPTW with a small number of paired requests per
vehicle. It works well only under restrictive capacity or time window constraints.
Savelsbergh and Sol [23] introduced a general model for the multiple-vehicle many-
to-many PDP that can handle all practical constraints. The optimization approaches
mentioned above are based on the notion of direct predecessor since they all de-
scribe routes in terms of a set of arcs aij . Savelsbergh and Sol [23] proposed an-
other branch-and-price approach for the PDPTW with some interesting features.
It uses construction and improvement heuristics to solve the pricing subproblem,
a primal heuristic at each node of the search tree to compute upper bounds, and a
clever column handling procedure to keep the column generation master problem
as small as possible. Lou and Dessouky [24] developed an MILP optimization-based
framework and a branch-and-cut algorithm for solving the MVPDP with time win-
dows. The approach does not depend on elimination techniques for reducing the
solution space due to violations in time window and capacity constraints. More-
over, it permits to solve problem instances of up to 5 vehicles and 17 customer
requests on problems without clusters within a stopping criterion of three CPU
hours on a Sun Fire 4800 Server. Recently, Cordeau [25] developed a branch-and-
cut algorithm for the dial-and-ride problem based on a three-index formulation.
The algorithm was able to solve problem instances with four vehicles and 32 re-
quests.
In this work, a new mixed-integer linear programming formulation (MILP) for
the multiple vehicle multiple-depot PDP with time windows is presented. The best
solution is found by applying a standard branch-and-cut tree search. The proposed
optimization approach is capable of handling transportation requests with multiple
origins and/or destinations, heterogeneous vehicles, and multiple depots. More-
over, it can still be applied even if pure pickup and/or delivery nodes are also con-
sidered. The MILP model includes all types of constraints usually arising in pickup
and delivery problems with time windows, i.e., pairing, preceding, time –windows,
and capacity constraints. Node sequencing variables are based on the notion of
general predecessor rather than direct predecessor since the routes are described
by specifying all the predecessors of every node on the tour. The one with no pre-
decessors is the first visited. Vehicle-request assignment and node sequencing de-
cisions are separately managed to get a two-index formulation with a significant
saving of 0–1 variables. Based on the notion of incompatible requests, effective
variable-and-constraint elimination rules have been developed to generate a more
compact solution space still containing the optimal solution. Different types of ob-
34 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

jective functions like vehicle fixed costs, distance-based traveling cost, time-based
traveling cost, customer inconvenience or a weighted combination of them can be
alternatively used. Though the many-to-many variant of the MVPDPTW with pure
pickup/delivery nodes can be handled by the proposed formulation, most of the
examples solved in the paper involve single pickup/single delivery requests. Only
one deals with many-to-many requests and a pair of depots.

2.2
The Problem Definition

The multiple vehicle pickup and delivery problem with time windows consists on
finding the optimal routes for a vehicle fleet in order to fulfill a set of customer
requests r ∈ R at minimum total cost while satisfying all problem constraints.
In the simplest case, each request r is defined by a pickup location i, a delivery
location j , and a load wr to be delivered from one to the other site. In the general
case, each customer request r involves a set of pickup locations i ∈ Ir from each
a load (wl)ir is to be collected, and a set of delivery locations j ∈ Jr to which an



amount (wd)j r is to be delivered in such a way that: i∈Ir wlir = j ∈Jr wdj r .
Pure pickup (i ∈ I PP ) and pure delivery (j ∈ JPD ) nodes can also be serviced. If so,
some vehicles will leave their depots with an initial cargo qvo and/or return to their
f
bases with a final cargo qv . Therefore,
   
wli,PP = qvf ; wdj,PD = qvo ,
i∈IPP v∈V j ∈JPD v∈V

where (wl)i,PP is the load to be moved from pure pickup node i ∈ I PP to the depot,
whilst (wd)j,PD stands for the cargo to be delivered from the depot to pure delivery
node j ∈ JPD .
Time windows [ai , bi ] are given for pickup and delivery tasks within which they
must be accomplished. Vehicles depart and return to the same depot but the one
assigned to a particular vehicle is a problem decision. Vehicle capacities and de-
pot locations are also problem data. While the vehicles perform the pickup and
delivery tasks, numerous constraints are to be satisfied. They are the following: (a)
every used vehicle has its assigned depot (depot assignment constraint); (b) each
customer request r ∈ R must be serviced by a single vehicle visiting a pickup node
first (vehicle assignment, pairing, and precedence constraints); (c) the capacity of a
vehicle can never be exceeded after visiting a pickup node (capacity constraints at
pickup nodes); (d) a vehicle must transport enough load to meet customer demand
when stopping at a delivery node (capacity constraints at delivery nodes); (e) each
used vehicle should return to its base (final destination constraints); (f) the total
time/distance traveled by a vehicle from the starting depot to a particular node lo-
cation must be greater than the one required to reach a preceding node on the tour
(time-based sequencing constraints); (g) the service at each node must be started
within the specified time window (time window constraints). The problem goal is
to minimize the total cost of providing pickup or delivery service to every node.
 2.3 The Problem Mathematical Formulation 35

Three types of costs are considered in the objective function: fixed costs for using
vehicles, variable traveling costs, and penalty costs due to both time window and
maximum routing time constraints violations. The objective function is a weighted
combination of such fixed and operating costs.

2.3
The Problem Mathematical Formulation

Consider a route–network represented by a graph G{I, J, P, R, A}, where I =

{i1 , i2 , . . . , in } denotes the set of pickup nodes, J = {j1 , j2 , . . . , jn } is the set of de-
livery nodes, P = {p1 , p2 , . . . , pl } represents the set of depots, R = {r1 , r2 , . . . , rs }
stands for the transportation requests each one featuring a set of pickup and deliv-
ery nodes Nr ={Ir ∪ Jr }, and A = {aii  /(i, i  ) ∈ I ∪ J ∪ P } defining the set of mini-
mum cost arcs among nodes. The set I may comprise pure pickup nodes (I PP ⊂ I )
while J could include pure delivery nodes (J PD ⊂ J ). There is a vehicle-dependent
distance-based traveling cost matrix C = {cii  }v , a vehicle-dependent travel time
matrix
= {tii  }v and a route-distance matrix D = {dii  } all associated to the set
A. Let us assume that transportation requests involving a single pickup node and a
single delivery node are first considered. For the pickup node i ∈ Ir of a request r,
there is a given load wlir to be collected within a time window [ai , bi ], and subse-
quently delivered to node j ∈ Jr (wdir = wlj r = wr ) within a time window [aj , bj ].
To fulfill pickup and delivery tasks, a set of vehicles V = {v1 , v2 , . . . , vm } is avail-
able. The service time at a pickup/delivery node k ∈ {I ∪ J } by vehicle v is denoted
by stvk . The set {I ∪ J } stands for the whole set of pickup and delivery nodes whilst
Nr denotes the pickup and delivery nodes associated to request r ∈ R. The solution
to the problem will provide a finite sequence of routes {pp , . . . , i, i  , . . . , pp }, with
ai,i  ∈ A, such that: (i) each vehicle starts and ends the trip at the same warehouse
p; (ii) each pickup or delivery site k ∈ {I ∪ J } is assigned to exactly one route;
(iii) the actual load carried out by a vehicle should neither exceed its capacity qv
nor become negative; (iv) a pickup or delivery node k ∈ {I ∪ J } should be serviced
within its time window [ak , bk ]; (v) the duration of the trip for any vehicle v should
be shorter than the maximum allowed routing time tvmax .
Three types of decision variables are included in the mathematical model: depot-
vehicle assignment variables (Xpv ), vehicle-request allocation variables (Yrv ), and
node sequencing variables (Sii  ). If pure pickup/delivery nodes are considered, then
the variables Yiv (i ∈ I PP ) and Yj v (j ∈ J PD ) are also to be defined. Binary variable
Xpv has value one if depot p ∈ P is assigned to vehicle v ∈ V . If request r ∈ R is
serviced by vehicle v, then the 0–1 variable Yrv becomes equal to one. In turn, the
sequencing variable Sii  is equal to one whenever pickup/delivery nodes (i, i  ) ∈
{I ∪ J } are on the same route and node i is visited earlier.
The proposed MILP formulation also includes six important nonnegative con-
tinuous variables: Ci , CVv , Ti , TVv , Li , and Ui . Variable Ci is the distance-based
transportation cost from the starting depot to node i ∈ Ir along the route traveled
by the vehicle assigned to request r. The travel time to go from the starting depot
36 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

to node i ∈ I is given by Ti . In turn, CVv is the overall routing cost for the tour
assigned to vehicle v. TVv is the total time required by vehicle v to complete the
tour. The overall load collected by vehicle v while going from the starting depot
up to node i ∈ I , including the initial cargo qvo , is given by Li . In contrast, Ui is
the overall cargo delivered by the assigned vehicle from the starting depot to node
i ∈ I . Additional continuous variables are also defined to denote the time window
upper/lower limit constraint violations [ai , bi ] and the maximum routing time
constraint violation Tv .

2.3.1
The Problem Constraints

2.3.1.1 Assignment Constraints

• Allocation of vehicles to depots: If used, a vehicle v ∈ V must
be assigned to a single depot p ∈ P from which it starts the
tour and to which it will return after accomplishing all the
pickup/delivery tasks.


Xvp
1, ∀v ∈ V . (2.1)
p∈P

• Assignment of vehicles to transportation requests: Each

transportation request r ∈ R must be fulfilled by just a single
vehicle v ∈ V .


Yrv = 1, ∀r ∈ R. (2.2)
v∈V

If pure pickup and/or delivery nodes are also considered,

then the following constraints (2.2’) must be taken into
account:

 
Yiv = 1, ∀i ∈ I PP Yj v = 1, ∀j ∈ J PD . (2.2’)
v∈V v∈V

• Vehicle availability condition: An unused vehicle v ∈ V has no

designated depot (Xvp = 0). If Xvp = 0, then the vehicle v is
not available to fulfill any customer request r ∈ R.
Consequently, none of the variables Yrv , Yiv or Yj v can be
equal to one as prescribed by Eq. (2.3). The parameter MX is
an upper bound on the number of transportation requests
allocated to any vehicle v.

   
Yrv + Yiv + Yj v
MX Xvp , ∀v ∈ V . (2.3)
r∈R i∈I PP j ∈J PD p∈P
 2.3 The Problem Mathematical Formulation 37

2.3.1.2 Routing-Cost Defining Constraints

• Minimum routing cost from the vehicle base to node i: If pickup
node i ∈ Ir is serviced by vehicle v (Yrv = 1) housed in depot
p (Xpv = 1), then the traveling cost from depot p to node
v . The
i (Ci ) must always be greater than or equal to cpi
v
parameter cpi represents the least travel cost from depot p to
v if pickup
node i. The value of Ci will be exactly equal to cpi
node i is the first one visited by the assigned vehicle v and
depot p is the selected base for v.

v
Ci  cpi (Xvp + Yrv − 1), ∀i ∈ Ir , r ∈ R, v ∈ V , p ∈ P . (2.4)

If vehicle v starts the trip empty, then a pickup node i ∈ Ir

for some request r will first be visited. In some cases,
however, the PDP problem involves pure delivery nodes and
therefore, some vehicles will leave their bases transporting a
finite load and may visit a pure delivery node j ∈ J PD first. If
pure delivery nodes are considered, then the additional
constraints (2.4’) are to be included in the problem
formulation.

v
Cj  cpj (Xvp + Yj v − 1), ∀j ∈ J PD , v ∈ V , p ∈ P . (2.4’)

Similar constraints for pure pickup nodes can be derived

from Eq. (2.4’) by simply replacing the variable Yj v by Yiv ,
and the set J PD by I PP .
• Distance-cost based sequencing constraints: Let (cii  )v stand for
the least travel cost from node i ∈ Nr to node i  ∈ Nr 
(i  = i) on vehicle v, where r, r  ∈ R and Nr includes
pickup/delivery nodes related to request r. If both nodes
(i, i  ) are on the same tour (Yrv = Yr  v = 1, for some vehicle
v) and node i is earlier visited (Sii  = 1), then the travel cost
from the base to node i  (Ci  ) must always be greater than Ci
by at least (cii  )v . If node i is visited later (Sii  = 0), the
reverse statement holds. Such conditions are enforced by
constraints (2.5a) and (2.5b) which become redundant
whenever nodes (i, i  ) ∈ N are serviced by different vehicles
(Yrv + Yr  v < 2), for any v. By definition, MC is an upper
bound on the travel cost from the depot to any node i ∈ N. It
is important to remark that just single sequencing variable
Sii  (i < i  ) is to be defined for every pair of nodes (i, i  ).

Ci   Ci + civi  − MC (1 − Si i  ) − MC (2 − Yrv − Yr  v ),
∀i ∈ Nr , i  ∈ Nr  , i < i  , r, r  ∈ R, v ∈ V . (2.5a)
38 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Ci  Ci  + civ i − MC Si i  − MC (2 − Yrv − Yr  v ),
∀i ∈ Nr , i  ∈ Nr  , i < i  , r, r  ∈ R, v ∈ V . (2.5b)
r
If case r = and every customer request involves a single
pickup node i ∈ Ir and a single delivery node j ∈ Jr , then
pickup node i should be visited earlier, and Sij = 1 assuming
i < j . Therefore, the distance-based sequencing constraints
(2.5a, 2.5b) reduce to the following pairing condition:

Cj  Ci + civj − MC (1 − Yrv ), ∀r ∈ R, v ∈ V . (2.5’)

If i  is a pure pick-node, then the constraints (2.5a) and (2.5b)
take the following form:

Ci   Ci + ciiv  − MC (1 − Sii  ) − MC (2 − Yrv − Yi  v ). (2.3)

Ci  Ci  + civi  − MC Si i  − MC (2 − Yrv − Yi  v ),
∀i ∈ Nr , i  ∈ I PP , r ∈ R, v ∈ V . (2.4)
When one of the nodes i or i  is a pure delivery node or both
are pure pickup/delivery nodes, then proper changes on
constraints (2.5a) and (2.5b) should be made. The variable
Yrv should be replaced by Yj v (or Yiv ) and the sets
{Nr , r ∈ R} by J PD (or I PP ).
• Overall routing cost for the tour assigned to vehicle v: The overall
traveling cost incurred by vehicle v (CVv ) to satisfy the
assigned requests r ∈ Rv must always be greater than the
traveling expenses from the origin-depot p to any delivery
node j ∈ Jr on the tour (i.e., Cj ) by at least the amount cjp v

standing for the travel cost from delivery node j to depot p.

Indeed, the last node visited by vehicle v is the one defining
the largest value of CVv and therefore, the constraint (2.6) for
such a delivery node and the visiting vehicle v is just binding
at the optimum. If delivery nodes j, j  ∈ J are both on the
tour and j  is the last visited, then constraint (2.6) for node j
will become redundant because the travel cost cjp v , by

definition, is smaller than or at most equal to the traveling

expenses from j to p through at least another node j  , i.e.,
v
cv + cv .
cjp jj  j p

CVv  Cj + cjv p Xv p − MC (1 − Yrv ),
p∈P
∀j ∈ Jr , r ∈ R, v ∈ V . (2.6)
If the problem includes pure pickup nodes, then some
vehicles will return to their bases transporting some loads
and a pickup node may then be the last visited on the tour.
 2.3 The Problem Mathematical Formulation 39

Similar constraints for pure pickup/delivery nodes can be

derived from (2.6) by simply replacing the variable Yrv by Yiv
(or Yj v ) and the set R by I PP (or J PD ).

2.3.1.3 Arrival-Time Defining Constraints

• Earliest visiting time for node i: The assigned vehicle v will
never arrive at node i ∈ I before time tpi v , where t v is the
pi
least travel time from depot p to node i. Constraint (2.7)
assumes that vehicle v is ready at t = 0. Otherwise, the ready
time for vehicle v should be added to tpi v .

v
Ti  tpi (Xvp + Yrv − 1), ∀i ∈ Ir , r ∈ R, v ∈ V , p ∈ P . (2.7)
If the problem includes pure delivery nodes, then some
vehicles can leave the depot transporting a finite load and a
delivery node may first be visited. A similar set of constraints
for pure pickup/delivery nodes can be derived from (2.7) by
simply replacing the variable Yrv by Yiv (or Yj v ) and the set R
by I PP (or J PD ).
• Time-based sequencing constraints: Let us assume that nodes
i ∈ Nr and i  ∈ Nr  (i  = i) are both serviced by the same
vehicle v. If node i is visited before (Sii  = 1), then the arrival
time at node i  (Ti  ) should be greater than Ti by at least the
sum of both the service time (sti ) at node i, and the traveling
time (tii  )v from i to i  . If not (Sii  = 0), the reverse
statement holds. If one of the nodes is not on the tour, then
Yrv + Yr  v < 2 for any vehicle v and therefore, constraints
(2.8a) and (2.8b) both become redundant. MT is an upper
bound on the duration of any tour.

Ti   Ti + sti + tiiv  − MT (1 − Sii  ) − MT (2 − Yrv − Yr  v ). (2.8a)

Ti  Ti  + sti  + tiv i − MT Si i  − MT (2 − Yrv − Yr  v ),

∀i ∈ Nr , i  ∈ Nr  , r, r  ∈ R, v ∈ V : i < i  . (2.8b)
In case r = r  and every customer request involves a single
pickup node i ∈ Ir and a single delivery node j ∈ Jr , then
node i should be visited before node j and Sij = 1.
Moreover, the time-based sequencing constraints (2.8a) and
(2.8b) reduce to the following pairing condition:

Tj  Ti + tijv − MT (1 − Yrv ), ∀r ∈ R, v ∈ V . (2.8c)

i
When either i or in constraints (2.8a) and (2.8b) is a pure
pickup/delivery node, the expression of such constraints
should be properly modified. The variable Yrv (and/or Yr  v )
40 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

should be replaced by Yiv (and/or Yj v ) and the sets Nr , r ∈ R

(or Nr  , r  ∈ R) by I PP (and/or J PD ).
• Overall routing time along the tour assigned to vehicle v: The
total time required by vehicle v to complete the assigned
transportation requests is found by adding both the service
time sti and the travel time tip v along the edge (i, p) to the

arrival time at the delivery node j ∈ J last visited. Since the

last delivery node visited by vehicle v is not known
beforehand, then Eq. (2.9) is written for every node j ∈ J .

TVv  Tj + stj + Xvp tjv p − MT (1 − Yrv ),
p∈P
∀j ∈ Jr , r ∈ R, v ∈ V . (2.9)
When the set I PP
is not an empty set, the last visited node
may be a pure pickup node. Similar constraints for pure
pickup/delivery nodes can be derived from (2.9) by simply
replacing the variable Yrv by Yiv (and/or Yj v ) and the set R
by I PP (or J PD ).
• Time window constraint-violation due to early service at node i:

ai  ai − Ti , ∀i ∈ N. (2.10)
• Time window constraint-violation due to late service at node i:

bi  Ti − bi , ∀i ∈ N. (2.11)
• Routing-time constraint-violation for vehicle v:

Tv  TVv − tvmax , ∀v ∈ V . (2.12)

Variables ai , bi and Tv are nonnegative.

2.3.1.4 Vehicle-Load Defining Constraints

• Capacity constraint on the load transported by vehicle v after
visiting node i: Equation (2.13a) states that the cargo
transported by vehicle v ∈ V just after servicing a pickup
node i ∈ Ir (Yrv = 1) must not exceed the vehicle capacity qv .
Such a load can be computed as the difference between
nonnegative variables Li and Ui . Variable Li stands for the
overall load collected by vehicle v along the route after
servicing node i, including the initial cargo, whilst Ui
represents the overall cargo unloaded from vehicle v from
the starting depot up to node i. The vth-vehicle capacity is
given by qv and Q = maxv (qv ) is an upper bound on the
value of qv . In turn, Eq. (2.13b) specifies that the cargo
 2.3 The Problem Mathematical Formulation 41

transported by the vehicle v ∈ V after servicing a delivery

node j ∈ Jr (Yrv = 1) must be nonnegative.

Li − Ui
qv Yrv + Q(1 − Yrv ), ∀i ∈ Ir , r ∈ R, v ∈ V . (2.13a)

Lj − Uj  0, ∀j ∈ Jr , r ∈ R. (2.13b)

When the PDPTW problem also involves pickup/delivery

nodes, then additional constraints (2.13a, 2.13b) for pure
pickup/delivery nodes are incorporated in the mathematical
model. They are obtained by simply replacing the variable
Yrv by Yiv (and/or Yj v ) and the set R by I PP (or J PD ).
• Load-based sequencing constraints: Let us assume that nodes
(i, i  ) ∈ N are linked to requests r ∈ R and r  ∈ R,
respectively, both serviced by vehicle v ∈ V (Yrv + Yr  v = 2).
If node i is visited before node i  (Sii  = 1), then the total
cargo loaded on vehicle v from the starting depot to node
i  (Li  ) should be greater than Li by at least (wl)i  r  . Such a
condition is enforced by Eq. (2.14a). The parameter (wl)i  r  is
zero if i  is a delivery node. Conversely, if node i  ∈ Ir  is
visited earlier, then Eq. (2.14b) holds and Li must be greater
than Li  by at least (wl)ir . Similarly, Eq. (2.15a) states that the
total load delivered by vehicle v from the starting depot up to
node i  ∈ Nr  (Ui  ) should be greater than Ui by at least
(wd)i  r  if both nodes are serviced by v and node i is visited
earlier. The parameter (wd)ir is zero if i is a pickup node. If
node i ∈ N is visited later (Sii  = 0), the reverse condition
holds and Eq. (2.15b) may become active. If one of the
requests r, r  ∈ R or both are not serviced by vehicle v, then
Yrv + Yr  v < 2, and all constraints (2.14) and (2.15) become
redundant. By definition ML is an upper bound on the total
load collected or delivered by a vehicle along the
assigned tour.

Li   Li + (wl)i  r  − ML (1 − Si i  ) − ML (2 − Yrv − Yr  v ) ,
i ∈ Nr , i  ∈ Nr  (i < i  ), r, r  ∈ R, v ∈ V . (2.14a)

Li  Li  + (wl)ir − ML Si i  − ML (2 − Yrv − Yr  v ) ,
i ∈ Nr , i  ∈ Nr (i < i  ), r, r  ∈ R, v ∈ V . (2.14b)

Ui   Ui + (wd)i  r  − ML (1 − Si i  ) − ML (2 − Yrv − Yr  v ) ,
∀i ∈ Nr , i  ∈ Nr  (i < i  ), r, r  ∈ R, v ∈ V . (2.15a)
42 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Ui
Ui  + (wd)ir + ML Sii  + ML (2 − Yrv − Yr  v ) ,

i ∈ Nr , i  ∈ Nr  (i < i  ), r, r  ∈ R, v ∈ V . (2.15b)
i
When i and/or or both in constraints (2.14a) and (2.14b)
and (2.15a) and (2.15b) stand for pure pickup/delivery nodes,
the expressions of such constraints should be properly
modified. The variable Yrv (and/or Yr  v ) should be replaced
by Yiv (and/or Yj v ), the sets Nr , r ∈ R (or Nr  , r  ∈ R) by I PP
(and/or J PD ), and the loads {(wl)ir or (wd)j r } by {(wl)i,PP or
(wd)j,PD }.
• Upper bounds on the values of Li and Ui :
The total cargo loaded on a vehicle v after visiting a node
i (Li ) can never be greater than the total load collected by
vehicle v along the tour plus the initial cargo with which it
leaves the assigned base.
 
Li − qvo Yrv − (wl)kr  Yr  v
ML (1 − Yrv ),
r  ∈R k∈Ir 
∀i ∈ Ir , r ∈ R, v ∈ V . (2.16)
Only a single constraint (2.16), the one related to the vehicle
assigned to node i will be binding at the optimum. Similarly,
the total cargo delivered by vehicle v after visiting a node
i (Ui ) can never be greater than the total amount of goods
that it delivers along the whole tour.
 
Ui − (wd)kr  Yr  v
ML (1 − Yrv ),
r  ∈R k∈Jr 
∀i ∈ Ir , r ∈ R, v ∈ V . (2.17)
When node i is a pure pickup/delivery node, the expressions
of constraints (2.16) and (2.17) should be properly modified.
For pure pickup nodes, Eqs. (2.16) and (2.17) will take the
following expressions:
  
Li − qvo Yiv − (wl)kr  Yr  v − (wl)i  Yi  v
ML (1 − Yiv ),
r  ∈R k∈Ir  i  ∈I PP
PP
∀i ∈ I ,v ∈ V (2.16’)

  
Ui − (wd)kr  Yr  v − (wd)j Yj v
ML (1 − Yiv ),
r  ∈R k∈Jr  j ∈J PD

∀i ∈ I PP , v ∈ V . (2.17’)
For pure delivery nodes, the set I PP in constraints (2.16’) and
(2.17’) should be replaced by J PD while the variable Yiv must
be substituted by Yrv and the set I PP by the set Nr for a
customer request r.
 2.4 Time Window-Based Variable and Constraint Elimination Rules 43

• Minimum cargo loaded on/delivered by a vehicle v after visiting

any node i: Equation (2.18a) states that the total cargo loaded
on the vehicle up to any node i ∈ Nr must at least be equal to
the initial cargo plus (wl)ir . The parameter (wl)ir takes a
positive value just for pickup nodes. Symmetrically, Eq.
(2.18b) states that the total cargo unloaded from the vehicle
up to node i ∈ Nr must at least be equal to (wd)ir . The
parameter (wd)ir takes a positive value just for delivery
nodes.
 
Li  (wl)ir + qv0 Yrv , ∀i ∈ Nr , r ∈ R, v ∈ V . (2.18a)
Ui  (wd)ir , ∀i ∈ Nr , r ∈ R. (2.18b)
For pure pickup/delivery nodes, the variable Yrv should be
replaced by Yiv (or Yj v ), the sets {Nr , r ∈ R} by I PP (or J PD ),
and the load {(wl)ir or (wd)j r } by {(wl)i,PP or (wd)j,PD }.

2.3.2
The Problem Objective Function

The problem goal is to minimize a weighted combination of vehicle fixed costs, dis-
tance and time-based routing costs as well as some measure of the total customer
inconvenience (time window constraint violations).
 
  
Minz = cfv Xpv + CVv + ρv Tv  + ρi (ai + bi ). (2.19)
v∈V p∈P i∈N

Assignment constraints (2.1)–(2.3) together with the distance-based sequencing

constraints (2.4)–(2.6) all define an alternative formulation to the traditional m-TSP.
Timing constraints are considered by the model through time-based sequencing
time constraints (2.7)–(2.9) and time window constraint (2.10)–(2.12), whilst capac-
ity constraints are handled through Eqs. (2.13)–(2.18). Timing window constraints
can be treated as hard constraints by driving the variables ai , bi , and tv all
to zero in constraints (2.10)–(2.12) and removing the associated penalty cost terms
from the objective function. The proposed mathematical model can account for
heterogeneous fleets, multiple depots, and even open tours not starting and end-
ing at the same depot.

2.4
Time Window-Based Variable and Constraint Elimination Rules

In order to improve the computational efficiency of the MILP branch-and-branch

solution procedure, exact elimination rules based on time window specifications
were proposed to remove a significant number of sequencing variables and con-
straints from the mathematical model. In this way, a more compact solution space
44 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

can be generated. The narrower the time windows, the larger the impact of the
elimination rules on the problem size. To maximize the effect of the elimination
rules, time windows are shrinked before applying them through the so-called TW-
contraction rules. They were derived in a straightforward manner from similar
rules proposed by Cyrus [26] and Desrochers et al. [27] for the VRPTW.

aknew = max akold , min (a + st + tk ), min (ai − stk − tki ) (2.20)
ak ∈A aki ∈A

bknew = min bkold , max (b + st + tk ), max (bi − stk − tki ) . (2.21)
ak ∈A aki ∈A

Two types of exact elimination rules are defined. Rules of type I are based on the
concept of incompatible transportation requests. In turn, rules of type II rely on
the fact that some pickup/delivery nodes associated with compatible requests must
be serviced in a certain order when visited by the same vehicle.

Incompatible Pickup/Delivery Nodes

Two nodes (i, k) ∈ N that cannot be assigned to the same vehicle are called incom-
patible. Otherwise, a time window constraint will be violated. The incompatibility
conditions for nodes (i, k) are given by:

ai + sti + ti,k > bk ,

ak + stk + ti,k > bi , (2.22)
where (ai , bi ) and (ak , bk ) stand for the time windows of nodes i and k, respectively.

Incompatible Transportation Requests

Two transportation requests r : (Ir , Jr ) and r  : (Ir  , Jr  ) that cannot be assigned
to the same vehicle are said to be incompatible. If they are serviced by the same
vehicle, some time window constraints will be violated. Requests r, r  ∈ R are said
to be incompatible if at least a pair of nodes (i, i  ) such that i ∈ Nr and i  ∈ Nr 
is incompatible according to definition (2.22). For a pair of incompatible requests
(r, r  ) ∈ IR, the following condition holds:
Yrv + Yr  v
1, ∀v ∈ V , (r, r  ) ∈ IR. (2.23)
If requests r and r  cannot be fulfilled by the same vehicle, then Eqs. (2.5a), (2.5b),
(2.8a), (2.8b), (2.14a), (2.14b), and (2.15a), (2.15b) for the pair (r, r’) ∈ IR and their
related nodes i ∈ Nr and i  ∈ Nr  together with the related sequencing variables Sii 
can all be deleted from the problem formulation. When a single-vehicle PDPTW
problem includes n pairs of incompatible requests involving a single pickup node
and a single delivery node, then 4n sequencing variables, and (8n+8n+2n+2n) =
20n sequencing constraints can be eliminated.

Preordered Pickup/Delivery Nodes of Compatible Requests

A pair of nodes (i, k) ∈ N is said to be a preordered one if they should be visited in
a certain order by the same vehicle if time window constraints are to be satisfied.
 2.4 Time Window-Based Variable and Constraint Elimination Rules 45

For instance, node i should be serviced before node k by a particular vehicle if the
following conditions hold:

ai + sti + ti,k
bk , (2.24a)

ak + stk + ti,k > bi . (2.24b)
Let us assume that the pickup/delivery nodes (i, k) are related to a pair of compat-
ible requests (r, r  ), i.e., i ∈ Nr and k ∈ Nr  , and satisfy conditions (2.24a), (2.24b).
Then, nodes (i, k) can be serviced by the same vehicle only if node i is visited first,
i.e., Sik = 1. Therefore, Eqs. (2.5a) and (2.5b) and (2.8a) and (2.8b) reduce them-
selves to simpler constraints (2.25) and (2.26), respectively.
v
Ck  Ci + cik − MC (2 − Yrv − Yr  v ),
i ∈ Nr , k ∈ Nr  (i < k), r, r  ∈ R, v ∈ V , (2.25)

Tk  Ti + stvk + tik
v
− MT (2 − Yrv − Yr  v ),
i ∈ Nr , k ∈ Nr  (i < k), r, r  ∈ R, v ∈ V . (2.26)
In addition, Eqs. (2.14a), (2.14b) and (2.15a), (2.15b) turn into constraints (2.27)
and (2.28):

Lk  Li + (wl)kr  − ML (2 − Yrv − Yr  v ),
i ∈ Nr , k ∈ Nr  (i < k), (r, r  ) ∈ R, v ∈ V , (2.27)

Uk
Ui + (wd)kr  + ML (2 − Yrv − Yr  v ),
i ∈ Nr , k ∈ Nr  (i < k), (r, r  ) ∈ R, v ∈ V . (2.28)
This rule can be viewed as the generalization of the one developed for Langevin
et al. [28] for the TSPTW. When a single-vehicle PDPTW problem includes n pairs
of preordered nodes, nSik -sequencing variable, and at most 6n constraints can be
eliminated.

2.4.1
Preassigning Vehicles to Transportation Requests

Elimination rules introduced in the previous section focus on reducing the number
of sequencing variables and constraints. The use of exact vehicle-request preassign-
ment rules can additionally reduce the number of assignment variables. Note that
in routing problems with identical vehicles, there is a large number of equivalent
solutions that differ only by the vehicle name. By swapping the assignment of two
vehicles, we get two similar solutions yielding the same transportation cost but fea-
turing different values for the assignment variables. A large number of alternative
optima greatly deteriorate the efficiency of the MILP branch-and-cut procedure. For
an homogeneous vehicle fleet, just one of the transportation requests r1 ∈ R can
arbitrarily be assigned to a particular vehicle unit v1 (Yr1,v1 = 1). By definition,
46 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Yr1,v1 + Yr,v1
1 for any transportation request r ∈ R(r = r1 ) that is incompatible

with r1 . If IRr1 stands for the set of requests that are incompatible with r1 , then
Yr,v1 = 0 for r ∈ IRr 1 ⊂ R. In this way, such assignment variables can be removed
from the problem formulation. From the set of r1 th-incompatible transportation re-
quests IRr1 , we can choose an arbitrary one r2 ∈ IRr1 to assign it to vehicle v2 = v1
(Yr2,v2 = 1). Therefore, Yr,v2 = 0 for any request r ∈ R1 such that (r, r2 ) are in-
compatible, i.e., for any r ∈ IRr2 ⊂ IRr1 . This assignment variable elimination
process can be repeated until the set IRrn just include compatible transportation
requests.
Let us assume that the vehicle fleet comprises m vehicles. Each time a re-
quest is preassigned to a vehicle, m assignment variables and 12(m − 1) con-
straints are saved. The preassignment strategy can be extended to the heteroge-
neous vehicle fleet case. Let us consider that there are 1, 2 . . . , k (k < m) types
of vehicles. In such a case, one can restrict the number of choices for a trans-
portation request r1 arbitrarily selected to a particular vehicle unit of each type
(Yr1,v1 + Yr1,v2 + · · · + Yr1,vk = 1). This restricts the set of vehicles that can accom-
plish the request r to a subset of V involving just one vehicle of each type. Conse-
quently, there are (m − k) vehicles that are forbidden to service node i. This implies
that (m − k) assignment variables and 12(m − k) constraints are dropped from the
problem formulation. This multiple vehicle-request preassignment process can be
repeated until the set IRrn just include compatible transportation requests. Obvi-
ously, preassignment rules are less effective for the heterogeneous fleet case than
for the homogeneous fleet case.

2.5
Numerical Results and Discussion

We tested the proposed formulation on some PDPTW examples taken from Li and
Lim [16]. Such examples were generated from the VRPTW Solomon benchmark
problems [29] by randomly pairing up the customer locations within routes in so-
lutions obtained through a heuristic approach for the VRPTW [14]. Benchmark
problems have been grouped into three different categories: LC, LR, and LRC. The
data set for every category comprises several problems with the same geographical
distribution, a central depot, similar vehicle capacities but different time windows.
Problems of class LC have clustered customers whose time windows have been
generated based on known solutions. LR-class problems have customer locations
uniformly randomly generated over a square. In turn, LRC-class problems have a
combination of clustered and randomly generated customers and time windows.
Problems of each class are further classified into two types “1” and “2.” Type-1 prob-
lems have narrow time windows and small vehicle capacity, while type-2 problems
have wider time windows and larger vehicle capacity. Therefore, solutions to type-
2 problems feature fewer tours and longer scheduling horizons. In this way, six
classes of data sets were generated: LC1, LC2, LR1, LR2, LRC1, and LRC2. Problem
data also include the number of available vehicles, Euclidean distances among cus-
 2.5 Numerical Results and Discussion 47

Table 2.1 Problem sizes and CPU times for several PDPTW benchmark problem instances.

Example Nodes Vehicles Optimal CPU time Binary Continuous Constraints

value (s) variables variables

LR-101 24 8 599.1 5.83 156 112 6265

30 10 700.2 50.07 245 140 12418
36 11 833.6 12.29 312 166 19702
50 14 1138.4 527.42 1108 228 52098
LR-102 24 8 543.9 8.81 200 112 6446
30 9 673.5 24.98 295 138 11376
36 11 824.6 2476.20 415 166 20181
40 11 860.2 5821.40 563 182 25334
LR-103 24 5 432.2 120.81 228 106 4129
30 6 616.5 519.94 306 132 7602
36 7 782.8 1715.69 392 158 12658
LR-104 24 4 493.3 9.25 268 104 3411
30 5 534.4 711.98 393 130 6586
36 6 681.2a 7200a 524 156 11250
LR-201 24 2 557.1 22.83 69 100 1124
30 2 645.8 287.23 101 124 1718
36 2 760.7 715.91 136 162 2429
LC-201 24 2 257.6 0.91 36 100 1024
30 2 269.1 1.47 46 124 1552
36 2 290.0 1.77 56 162 2188
LC-102 24 3 249.8 7.83 108 102 2067
30 4 313.5 22.69 161 128 4529
36 5 370.9 121.75 282 130 8581
LR-103m 36 7 759.8 88.74 387 158 12628

a Best solution after 7200 s of CPU time.

tomer sites, and normalized vehicle speeds such that traveling times and Euclidean
distances are numerically identical.
In the PDPTW problem-generation process starting from the VRPTW Solomon
benchmark problems, a customer request is randomly assigned to either a pure
pickup (delivery) site and paired with a pure delivery (pickup) location on the same
route [16]. All problems have 100 real nodes with some dummy nodes for coupling
purpose if necessary, and a single central depot. In this way, MVPDPTW problem
instances with approximately 50 customer requests each one involving a pair of
delivery and pickup locations are generated. In order to derive problem instances
that can be solved to optimality, we take the first n/2 requests of Li and Lim’s ex-
amples to generate a problem instance with at most 25 request-pairs and 50 nodes.
They were solved by applying ILOG Studio OPL 3.5 [30] on a 2.0 GHz, 1.048 GB
RAM Pentium IV HP Workstation. The traveled distance was the selected objec-
tive function to be minimized and customer time windows were considered hard
constraints. A series of MVPDPTW problems of different types and sizes has been
solved. Numerical results for different instances of problems LR-101/102/103/104,
48 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Table 2.2 Optimal solution for the LR-101 problem instance with 50 nodes.

Vehicle Pickup/ Waiting Initial Final Service Pickup/ Vehicle Vehicle Traveled
delivery time service service duration delivery load routing distance
node time time load time

V1 P83 22.90 44.00 54.00 10.00 11 11

D37 72.00 144.00 154.00 10.00 −11 0
94.90 175.20 60.30
V2 P33 12.20 37.00 47.00 10.00 11 11
D34 67.60 127.00 137.00 10.00 −11 0
79.80 173.10 73.30
V3 P75 27.60 69.00 79.00 10.00 18 18
D22 13.90 97.00 107.00 10.00 −18 0
P55 16.80 146.00 156.00 10.00 2 2
D25 22.40 182.00 192.00 10.00 −2 0
80.70 225.50 91.20
V4 P36 0.00 41.40 51.40 10.00 5 5
P47 0.00 58.60 68.60 10.00 27 32
D19 0.00 76.70 86.70 10.00 −5 27
P08 0.60 105.00 115.00 10.00 9 36
D46 2.60 127.00 137.00 10.00 −27 9
D17 11.80 167.00 177.00 10.00 −9 0
15.00 207.40 132.40
V5 P31 32.50 50.00 60.00 10.00 27 27
P88 9.00 74.00 84.00 10.00 9 36
D07 0.00 90.30 100.30 10.00 −9 27
D10 19.60 134.00 144.00 10.00 −27 0
61.10 169.50 68.40
V6 P52 40.70 52.00 62.00 10.00 9 9
D06 23.80 99.00 109.00 10.00 −9 0
64.50 120.20 35.70
V7 P69 37.80 50.00 60.00 10.00 6 6
P76 10.00 83.00 93.00 10.00 13 19
D03 0.00 123.90 133.90 10.00 −6 13
D54 0.00 150.00 160.00 10.00 −13 0
47.80 182.80 70.80
V8 P21 44.00 62.00 72.00 10.00 11 11
D41 11.70 97.00 107.00 10.00 −11 0
P56 10.90 130.00 140.00 10.00 6 6
D04 10.80 159.00 169.00 10.00 −6 0
77.40 184.00 76.60
V9 P63 9.10 44.00 54.00 10.00 10 10
P64 4.60 73.00 83.00 10.00 9 19
D49 12.30 108.00 118.00 10.00 −10 9
D48 39.50 175.00 185.00 10.00 −9 0
65.50 212.80 107.30
 2.5 Numerical Results and Discussion 49

Table 2.2 (continued)

Vehicle Pickup/ Waiting Initial Final Service Pickup/ Vehicle Vehicle Traveled
delivery time service service duration delivery load routing distance
node time time load time

V10 P28 32.70 39.00 49.00 10.00 16 16

D12 4.80 63.00 73.00 10.00 −16 0
P40 7.90 95.00 105.00 10.00 9 9
D26 19.90 132.00 142.00 10.00 −9 0
65.30 153.20 47.90
V11 P11 33.50 67.00 77.00 10.00 12 12
P90 16.80 95.00 105.00 10.00 3 15
D20 6.90 126.00 136.00 10.00 −12 3
D32 4.20 151.00 161.00 10.00 −3 0
61.40 195.00 103.60
V12 P30 45.50 71.00 81.00 10.00 21 21
D09 1.00 97.00 107.00 10.00 −21 0
P66 14.40 135.70 145.70 10.00 25 25
D01 0.00 171.00 181.00 10.00 −25 0
60.90 196.20 95.30
V13 P14 0.00 32.00 42.00 10.00 20 20
D38 29.80 83.00 93.00 10.00 −20 0
P43 20.90 132.00 142.00 10.00 7 7
D13 3.90 169.00 179.00 10.00 −7 0
54.60 190.20 95.60
V14 P71 37.30 77.00 87.00 10.00 15 15
D50 23.70 124.00 134.00 10.00 −15 0
61.00 151.00 80.00

Total traveled distance: 1138.4.

Total waiting time: 889.9.
Total service time: 500.0.

LR-201, and LC-102/201, all involving transportation requests with a single pickup
and a single delivery site, are summarized in Table 2.1. The number of nodes and
vehicles vary with the problem instance as shown in Table 2.1. To illustrate the
application of the approach to MVPDPTW examples involving customer requests
with multiple pickup and multiple delivery points, a modified version of the LR-103
problem instance featuring 36 nodes, 2 depots, and 7 vehicles, called LR-103m, has
been tackled and the results appear on the last row of Table 2.1.
For small time window widths, the behavior of the algorithm greatly improves
due to the better performance of the exact elimination rules and, consequently, the
CPU time increases almost linearly with the problem size. This is clearly observed
at benchmark problems LC-201 and LR-101 with tight time windows for which the
solution times are much shorter than those required for solving less constrained
examples. In some cases, the CPU time is almost independent of the problem size
(see LR-101 problem instances with 24 and 36 nodes in Table 2.1). This leads to the
conclusion that the proposed formulation is capable of efficiently finding the least-
50 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Warehouse
Pick-up nodes
Delivery nodes

Fig. 2.1 A sketch of the optimal solution to the LR-101 problem instance with 50 nodes.

routing-cost tours for large MVPDPTW problems with narrow time windows. For
instance, the version of problem LR-101 with 50 nodes and 14 vehicles was solved
to optimality in just 527 s. Table 2.2 presents a detailed description of the optimal
routes/schedules found for such an instance of problem LR-101. In Table 2.2, every
node with prefix P stands for a pickup location while the prefix D indicates a de-
livery site. Figure 2.1 shows a drawing of the optimal routes for problem LR-101
(50 nodes) where hollow squares represent pickup nodes and black squares denote
delivery nodes. Petal-shaped routes are somewhat distorted by the pairing, prece-
dence, and capacity constraints. Moreover, Fig. 2.2 depicts a Gantt chart showing
the vehicles activities accomplished throughout the time horizon with light gray
boxes denoting pickup activities and dark boxes representing delivery tasks. It is
also indicated the customer ID at which the activity is performed and the amount
that is collected (a positive number) or delivered (a negative number) at each lo-
cation. The x-axis is indeed a time-axis indicating the time intervals at which the
tasks assigned to every used vehicle are performed while the resources (vehicles)
are represented in the y-axis.
The computational cost shows a larger growth with the number of nodes for
wider time windows as follows from Table 2.1. The CPU time required to solve
different instances of problems LR-102, LR-103, and LR-104 to optimality signifi-
cantly increases as the number of nodes climbs from 24 to 36 nodes, especially for
problem LR-104. For such less constrained examples, the exact elimination rules
no longer eliminate an important number of constraints and variables. In spite of
 2.5 Numerical Results and Discussion 51

Fig. 2.2 Gantt chart describing the optimal solution to LR-101 problem instance with 50 nodes.

that, instances of benchmark problems LR-102, LR-103, LR-201, LC-201, and LC-
102, involving up to 36 nodes were solved in very acceptable CPU times. Although
the algorithm may grow exponentially with the number of nodes, medium-size
problems with fairly wide and overlapping time windows can still be tackled to
optimality at a reasonable computational cost.
Since it presents a mix of narrow and wide time windows that is usual on realis-
tic problems, the benchmark problem instance LR-103 (36) looks very appropriate
to be analyzed with more detail. The optimal solution was found in a CPU time
of 1716 s. Table 2.3 describes the least-cost vehicle routes and schedules found for
such an instance of problem LR-103 with 36 nodes. In turn, a sketch of the vehi-
cle routes is shown in Fig. 2.3. Since the total traveling cost has been minimized,
similarly to the optimum of problem instance LR-101 (50), the total waiting time
(411.30) is significantly larger and even greater than the total service time (360.00).
Moreover, the total waiting time for vehicle V2 is even higher than the travel time.
This is a consequence of using a purely geographical objective function. A more ap-
propriate evaluation criterion that also prices the routing times will surely reduce
vehicle idle times, probably at the cost of somewhat increasing the traveled dis-
tance. Figure 2.4 depicts the pickup/delivery tasks carried out by any vehicle along
the time horizon.
52 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Table 2.3 Optimal solution for the LR-103 problem instance with 36 nodes.

Vehicle Pickup/ Waiting Initial Final Service Pickup/ Vehicle Vehicle Traveled
delivery time service service duration delivery load routing distance
node time time load time

V1 P30 45.50 71.00 81.00 10 21 21

P49 2.10 118.00 128.00 10 30 51
D32 16.10 173.10 183.10 10 −30 21
D01 0.00 204.00 214.00 10 −21 0
63.70 219.20 125.55
V2 P33 12.20 37.00 47.00 10 11 11
P79 41.75 94.45 104.45 10 23 34
D10 0.00 134.00 144.00 10 −11 23
D31 49.90 202.00 212.00 10 −23 0
103.85 229.50 85.50
V3 P26 19.90 31.10 41.10 10 17 17
P23 0.00 68.00 78.00 10 29 46
P55 40.55 138.25 148.25 10 2 48
D24 0.00 160.45 170.45 10 −2 46
D29 0.00 177.51 187.51 10 −29 17
D03 0.00 197.00 207.00 10 −17 0
60.45 228.45 108.90
V4 P02 0.00 18.00 28.00 10 7 7
P43 0.00 46.45 56.45 10 7 14
D15 0.00 63.70 73.70 10 −7 7
P41 11.10 97.00 107.00 10 5 12
D04 33.60 159.00 169.00 10 −7 5
D21 22.00 201.00 211.00 10 −5 0
66.70 229.00 102.30
V5 P52 17.80 29.10 39.10 10 9 9
P19 0.00 59.95 69.95 10 17 26
D11 0.00 77.00 87.00 10 −17 9
P83 30.00 145.60 155.60 10 11 20
D17 0.00 167.00 177.00 10 −9 11
D05 0.00 187.00 197.00 10 −11 0
47.80 217.60 109.80
V6 P60 0.00 18.00 28.00 10 3 3
P98 2.20 43.45 53.45 10 10 13
D14 0.00 65.15 75.15 10 −10 3
P97 0.00 90.00 100.00 10 12 15
D06 0.00 109.00 119.00 10 −3 12
D13 42.90 169.00 179.00 10 −12 0
45.10 190.10 85.00
V7 P65 1.10 51.00 61.00 10 20 20
P56 3.40 130.00 140.00 10 6 26
D25 17.60 172.00 182.00 10 −6 20
D28 1.60 213.00 223.00 10 −20 0
23.70 229.30 165.70

Total traveled distance: 782.75; total waiting time: 411.30; total service time: 360.00.
 2.5 Numerical Results and Discussion 53

Warehouse
Pick-up nodes
Delivery nodes

Fig. 2.3 A sketch of the optimal solution to the LR-103 problem instance with 36 nodes.

Fig. 2.4 Gantt chart describing the optimal solution to the LR-103 problem with 36 nodes.
54 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Table 2.4 Optimal solution for the modified LR-103m problem instance with 36 nodes.

Ware Vehicle Pickup/ Waiting Initial Final Service Pickup/ Vehicle Vehicle Traveled
house delivery time service service duration delivery load routing distance
node time time load time

P2 V1 P30 63.90 71.00 81.00 10 21 21

P49 0.00 115.90 125.90 10 30 51
D32 0.00 154.90 164.90 10 −30 21
D01 0.00 185.80 195.80 10 −21 0
63.90 212.90 109.00
P2 V2 P33 14.80 37.00 47.00 10 34 34
D79 39.30 92.00 102.00 10 −3 31
D10 0.00 131.50 141.50 10 −11 20
D31 0.00 149.60 159.60 10 −20 0
54.10 173.10 79.00
P1 V3 P26 29.90 41.10 51.10 10 48 48
D23 0.00 78.00 88.00 10 −4 44
D55 30.60 138.20 148.20 10 −5 39
D24 0.00 160.40 170.40 10 −2 37
D29 0.00 177.50 187.50 10 −20 17
D03 0.00 197.00 207.00 10 −17 0
60.50 229.40 108.90
P1 V4 P02 0.00 18.00 28.00 10 7 7
P43 0.00 46.40 56.40 10 7 14
D15 0.00 63.70 73.70 10 −7 7
P41 11.10 97.00 107.00 10 5 12
D04 23.60 149.00 159.00 10 −7 5
D21 0.00 169.00 179.00 10 −5 0
34.70 197.00 102.30
P1 V5 P52 0.00 11.30 21.30 10 9 9
P19 7.80 49.90 59.90 10 17 26
D11 0.00 67.00 77.00 10 −17 9
P83 30.00 135.60 145.60 10 11 20
D17 0.00 157.00 167.00 10 −9 11
D05 0.00 177.00 187.00 10 −11 0
37.80 207.60 109.85
P1 V6 P60 0.00 18.00 28.00 10 3 3
P98 0.00 41.20 51.20 10 10 13
D14 0.00 62.90 72.90 10 −10 3
P97 0.00 87.80 97.80 10 12 15
P06 0.00 106.80 116.80 10 −3 12
D13 35.50 159.00 169.00 10 −12 0
35.50 180.20 85.10
P1 V7 P65 1.10 51.00 61.00 10 20 20
P56 3.40 130.00 140.00 10 6 26
P25 16.60 172.00 182.00 10 6 32
D28 0.00 211.40 221.40 10 −32 0
21.10 227.70 165.60

Total traveled distance: 759.75; total waiting time: 307.60; total service time: 360.00.
 2.5 Numerical Results and Discussion 55

Warehouses
Pick-up nodes
Delivery nodes

Fig. 2.5 A sketch of the optimal solution to the modified LR-103 problem with 36 nodes.

The more general case involving several many-to-many customer requests and
multiple depots is illustrated by solving the problem instance LRC-103 with 36
nodes. The modified version includes a new warehouse with Euclidean coordinates
X = 35 and Y = 65. In addition, two one-to-many requests: {P33, D10, D31, D79}
and {P26, D03, D23, D24, D29, D55}, and a single many-to-one request {P25, P56,
P65, D28} were specified. In the new version of problem instance LRC-103m with
36 nodes, some pickup nodes turned into delivery nodes and vice versa. Moreover,
the load to be collected or delivered at some nodes were also modified to define
load-balanced requests so that vehicles depart from and arrive to their bases empty.
As shown in Table 2.4, customer requests involving multiple pickup or multiple
delivery nodes are serviced by vehicles V2, V3, and V7 in the optimal solution. In
spite of the higher complexity introduced by the new pairing/precedence condi-
tions and the presence of a new warehouse, the optimal solution to the modified
problem instance LRC-103m (36 nodes) was found in just 88.74 s. This is because
the handling of requests with multiple pickup and delivery nodes improves the
performance of the exact elimination rules and, consequently, a higher number
of binary variables can be deleted. Detailed results are included in Table 2.4, and
the sketch of the optimal vehicle routes are depicted in Fig. 2.5. The relocation of
vehicles V1 and V2 from depot P1 to P2 permits to save 23 units of traveled dis-
tance/time. Figure 2.6 depicts the pickup and delivery activities carried out by every
vehicle as a function of time.
56 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Fig. 2.6 Gantt chart describing the optimal solution to the

modified LR-103 (36) problem instance.

2.6
Conclusions

A new MILP formulation for the multiple vehicle pickup and delivery problem
with time windows (MVPDPTW) based on a generalized precedence notion has
been presented. The approach allows to separately treating depot-vehicle allocation,
vehicle-request assignment, and node sequencing decisions through different sets
of two-index binary variables. In this way, a two-index mathematical model capable
of dealing with heterogeneous vehicles, multiple depots, many-to-many requests,
and pure pickup/delivery nodes has been derived. The model utilizes a continuous
representation of the time domain, the cost domain, and the load domain. Actual
vehicle load is calculated as the difference between two positive strictly increasing
variables, the total pickup load, and the total delivery load. Different types of ob-
jective functions like vehicle fixed costs, distance-based traveling cost, time-based
traveling cost, customer inconvenience or a weighted combination of them can
be alternatively used. To tackle medium-size problems, several variable and con-
straint elimination rules disallowing noncompatible requests to stay on the same
tour have also been presented. Moreover, preordering rules to be complied for
some compatible requests if they are satisfied by the same vehicle permit to delete
further 0−1 variables and constraints. On top of that, preassignment rules aimed
at reducing the number of vehicle-request assignment variables and discard sym-
metric solutions were also developed. Results show that the proposed formulation
is capable of efficiently finding the least-cost vehicle routes for large MVPDPTW
problems with narrow time windows. Optimal solutions for such a class of bench-
 Acknowledgments 57

mark problem instances involving up to 50 nodes and 14 vehicles were found in

quite acceptable CPU times. Despite the elimination rules do not work well for less
constrained examples, problem instances with wider time windows, and up to 36
nodes were still solved at a reasonable computational cost. Finally, a more general
case study involving several many-to-many customer requests, 36 nodes, 7 vehicles,
and a pair of depots was found in just 89 s.
Areas of current research include the preprocessing of problem data to define
and incorporate mini-clusters into the problem formulation and the development
of reactive approaches aimed at improving solutions provided by incomplete for-
mulations or heuristic search.

Acknowledgments

Financial support received from FONCYT-ANPCyT under Grant PICT 11-14717,

from CONICET under Grant PIP-5729 and from Universidad Nacional del Litoral
under CAI+D 003-13 is fully appreciated.

References

1 Bodin, L., Golden, B., Assad, A., Toth, P., Vigo, D. (eds.), SIAM Mono-
Ball, M., Routing and scheduling of graphs on Discrete Mathematics and
vehicles and crews: The state of the Applications, SIAM, Philadelphia,
art. Comput. Oper. Res. 10(2) (1983), 2002, pp. 85–108.
pp. 62–212. 6 Sexton, T., Bodin, L., Optimizing
2 Savelsbergh, M. W., Sol, M., The single vehicle many-to-many opera-
general pickup and delivery problem. tions with desired delivery times: I.
Transport. Sci. 29(1) (1995), pp. 17–29. Scheduling. Transport. Sci. 19 (1985),
3 Fisher, M. L., Vehicle routing, in: Net- pp. 378–410.
work Routing. Handbooks in Operations 7 Dumas, Y., Desrosiers, J., Soumis, F.,
Research and Management Science, vol. Large scale multi-vehicle dial-a-ride
8, Ball, M., Magnanti, T. L., Monma, systems. Ecole des Hautes Commer-
C. L., Nemhauser, G. L. (eds.), North- ciales. Montreal Report G-89-30, 1989.
Holland, Amsterdam, 1995, pp. 1–33. 8 Jaw, J. J., Odoni, A., Psaraftis, H.,
4 Desrosiers, J., Dumas, Y., Solomon, Wilson, N., A heuristic algorithm
M., Soumis, F., Time constrained for the multi-vehicle advance-request
routing and scheduling, in: Network dial-a-ride problem with time win-
Routing. Handbooks in Operations Re- dows. Transport. Res. B20 (1986), pp.
search and Management Science, vol. 8, 243–257.
Ball, M., Magnanti, T. L., Monma, 9 Toth, P., Vigo, D., Heuristic algo-
C. L., Nemhauser, G. L. (eds.), North- rithms for the handicapped persons
Holland, Amsterdam, 1995, pp. 35– transportation problem. Transport. Sci.
139. 31 (1997), pp. 60–71.
5 Desaulniers, G., Desrosiers, J., Erd- 10 Thompson, P., Psaraftis, H. N.,
mann, A., Solomon, M. M., Soumis, Cyclic transfer algorithms for mul-
F., VRP with pickup and delivery, in: tivehicle routing and scheduling prob-
The Vehicle Routing Problem, vol. 9, lems. Oper. Res. 41 (1993), p. 5.
58 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

11 Van der Bruggen, L., Lenstra, J., 20 Psaraftis, H., An exact algorithm for
Schuur, P., Variable depth search the single vehicle many-to-many dial-
for the single-vehicle pickup and de- a-ride problem with time windows.
livery problem with time windows. Transport. Sci. 17 (1983), pp. 351–361.
Transport. Sci. 27 (1993), pp. 298–311. 21 Desrosiers, J., Soumis, F.,
12 Potvin, J., Rousseau, J., A parallel Desrochers, M., Sauve, M., Methods
route building algorithm for the vehi- for routing with time windows. Eur. J.
cle routing and scheduling problem Oper. Res. 23 (1986), pp. 235–245.
with time windows. Eur. J. Oper. Res. 22 Dumas, Y., Desrosiers, J., Soumis, F.,
66 (1993), pp. 331–340. Large scale multi-vehicle dial-a-ride
13 Nanry, W., Barnes, J., Solving the systems, Ecole des Hautes Commer-
pickup and delivery problem with ciales. Montreal Report G-89-30, 1989.
time windows using reactive tabu-
23 Savelsbergh, M. W., Sol, M., DRIVE:
search. Transport. Res. 34 (2000), pp.
dynamic routing of independent vehi-
107–121.
cles. Oper. Res. 46 (1998), pp. 474–490.
14 Taillard, E., A tabu search heuristic
24 Lou, Q., Dessouky, M. M., An exact
for the vehicle routing problem with
algorithm for the multiple vehicle
soft time windows. Transport. Sci. 31
pickup and delivery problem. Trans-
(1997), pp. 170–186.
port. Sci. 38 (2004), pp. 503–514.
15 Potvin, J., Robillard, C., The impact
25 Cordeau, J. F., A branch-and-cut al-
of emerging technologies on com-
gorithm for the dial-a-ride problem.
puter science and operations research,
Oper. Res. 54 (2006), pp. 573–586.
Integrating Operations Research and
Neural Networks for Vehicle Routing, 26 Cyrus, J. P., The vehicle scheduling
Kluwer Academic Publishers, Boston, problem: models, complexity and al-
MA, 1995, pp. 245–262. gorithms. PhD. Dissertation, Technical
16 Li, H., Lim, A., A meta-heuristic for University of Nova Scotia, Halifax,
the pickup and delivery problem with N.S., 1988.
time windows. J. Artif. Intel. Tools 27 Desrochers, M., Desrosiers, J.,
12(2) (2003a), pp. 173–186. Solomon, M., A new optimization
17 Li, H., Lim, A., Local search with algorithm for the vehicle routing prob-
annealing-like restarts to solve the lem with time windows. Oper. Res. 40
VRPTW. Eur. J. Opert. Res. 150(1) (1992), pp. 342–354.
(2003b), pp. 115–127. 28 Langevin, A., Desrochers, M.,
18 Tam, V., Kwan, M. C., Adapting the Desrosiers, J., Soumis, F., A two-
large neighborhood search to effec- commodity flow formulation for the
tively solve pickup and delivery prob- traveling salesman and makespan
lems with time windows. Proceedings problem with time windows. Networks
of the 16th IEEE International Confer- 23 (1993), pp. 631–640.
ence on Tools with Artificial Intelligence 29 Solomon, M., Algorithms for the ve-
(ICTAI 2004), 15–17 November, Boca hicle routing and scheduling problem
Raton, FL, USA, 2004. with time window constraints. Oper.
19 Bent, R., van Hentenryk, P., A two- Res. 32 (1987), pp. 254–265.
stage hybrid algorithm for pickup and 30 Van Hentenryck, P., ILOG OPL opti-
delivery vehicle routing problems with mization language reference manual,
time windows. Comput. Oper. Res. ILOG S.A., The MIT Press, France,
33(4) (2006), pp. 875–893. 2004.
 Appendix 59

Appendix

Nomenclature

Sets
N Nodes
I Pickup nodes
J Delivery nodes
I PP Pure pickup nodes (I PP ⊂ I )
J PD Pure delivery nodes (J PD ⊂ JI)
R Transportation requests
P Depots
V Vehicles
D Minimum-cost arcs connecting pickup and delivery nodes
Nr Nodes related to request r ∈ R
Ir Pickup nodes related to request r ∈ R
Jr Delivery nodes related to request r ∈ R

Node-related Parameters
wlir Load to be collected at pickup node i to fulfill request r
wdj r Load to be transported to delivery node j to fulfill request r
ai Earliest service time at node i
bi Latest service time at node i
dij Minimum-cost distance between locations i and j
cijv Travel cost between nodes i ∈ I and j ∈ I on vehicle v ∈ V

tijv Travel time between nodes i ∈ I and j ∈ I on vehicle v ∈ V

stiv Service time at node i

Vehicles-related Parameters
qv vth-vehicle capacity
cfv vth-vehicle fixed utilization cost
qv0 vth-vehicle initial load
tvmax Maximum allowed routing time for vehicle v

Specific Penalties
ρi Unit penalty cost on ith-node time window constraint violation
ρv Unit penalty cost on vth-vehicle routing time constraint violation

Binary Variables
Xvp Binary variable denoting the allocation of vehicle v ∈ V to depot p ∈ P .
Yrv Binary variable denoting the assignment of transportation request r ∈ R to
vehicle v ∈ V .
Sij Binary variable denoting that node i ∈ I is visited before (Sij = 1) or after
(Sij = 0) node j ∈ I whenever both nodes are serviced by the same vehicle.
60 2 Solving Multiple Vehicle Pickup and Delivery Problems in Multisite Systems

Vertexes-related Continuous Variables

Ci Accumulated vehicle routing cost at node i ∈ N
Ti Accumulated vehicle travel time after the service at node i ∈ N has been com-
pleted
ai ith-time window constraint violation due to early arrivals
bi ith-time window constraint violation due to late arrivals
Li Total cargo loaded on the assigned vehicle after completing the service at pickup
node i ∈ I
Uj Total cargo unloaded from the assigned vehicle after completing the service at
delivery node j ∈ J

Vehicles-related Continuous Variables

CVv Total vth-vehicle routing cost after returning to the assigned depot
TVv Total vth-vehicle travel time after returning to the assigned depot
Tv vth-vehicle routing time constraint violation

Other Parameters
Nd Number of depots
nv Number of vehicles
npn Number of pickup nodes
nd Number of delivery nodes
 61

3
A Real Time Approximate Dynamic Programming Approach:
A High Dimensional Supply Chain Application
Nikolaos E. Pratikakis, Matthew J. Realff, Jay H. Lee

In this chapter, we introduce the main concepts of dynamic programming (DP).

Then, we specify the computational obstacles of DP, named as the “Curse of Di-
mensionality”(COD) and focus on techniques that approximate DP. Specifically,
we delineate a Real-Time Approximate Dynamic Programming (RTADP) method-
ology for solving multi-stage decision problems under a stochastic environment.
The case study presented is a high dimensional supply chain network, which is to
be optimized via dynamic decisions. The supply chain system experiences stochas-
tic variations in demand and price of the main products, which are modeled using
Markov chains. Hence, this dynamic supply chain decision problem is formulated
as a Markov Decision Process (MDP). The main idea of the RTADP methodology is
to start with a heuristic policy derived from a Mixed Integer Linear Programming
(MILP) formulation and to gradually construct a superior quality solution by inter-
acting with the stochastic system via simulation. We will explicitly reason how one
can circumvent the COD, when using the RTADP. Specifically we will introduce
several notions including “evolving set of relevant states”, for which the value func-
tion table is built, “adaptive action set” to minimize the COD concerning the Action
space, ‘non-parametric k nearest neighbor averager’ for value function approxima-
tion, and Monte Carlo sampling to eliminate the need to accurately compute the
one step expectation operator for each action.

3.1
Introduction

Obtaining high quality solutions in supply chain case studies, where the future
uncertainty is significant is an important optimization challenge. Complexity arises
both from the details of the application itself, and the multistage nature of the
decisions, where information is received during the planning horizon and has to be
taken into account. Formulating and solving this type of multistage optimization
problem generally entails exploring a large number of scenarios or performing
multi-dimensional integrals over probability distributions, which typically increase

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
62 3 A Real Time Approximate Dynamic Programming Approach

exponentially with the number of decision stages. A quantitative discussion on the

complexity of multistage stochastic problem can be found in [1].
Historically, multistage planning and scheduling optimization problems with a
known future have been formulated via mathematical programming. Naturally,
when it comes to deterministic problems, mathematical formulations and in par-
ticular MILPs are undoubtedly the way to go. A natural temptation is for these for-
mulations to be applied to stochastic problems by replacing all random variables
by their expected value. This is called the mean value problem [2].
Such solutions are suboptimal and often infeasible. Recently, mathematical for-
mulations with uncertain future have been a subject of considerable research [3],
but they run into several obstacles:
1. Solving for an expected value by sampling the future misses
the opportunity to revise actions depending on the state.
2. Solving the full problem where actions can depend on the
state requires that the branching of the future scenarios be
taken into account; this by itself presents the following
problems:
• The number of branching points and scenarios is
exponential to the number of time periods. Therefore
even writing the problem as a mathematical program
becomes very difficult.
• One needs fairly restrictive assumptions about how
the actions and the future interact. For example, it is
very difficult to express situations in which the actions
change the nature of the underlying transitions (e.g.,
by revealing information).

Rolling horizon mathematical formulations takes care the lack of feedback, but
run into a compromise between the first two problems. Essentially, one can limit
the combinatorial explosion using the rolling horizon idea, but does not solve the
problem of having choices now, that depend on the future and are not properly
evaluated.
Our prologue, motivates to explore the usage of Dynamic Programming (DP).
Of course DP is compromised by several computational obstacles therefore we will
focus on methodologies and ideas from the evolving stream of literature named af-
ter Approximate Dynamic Programming (ADP). Briefly, the key advantage of ADP
methodologies is that they are based on a procedural representation of the problem
(essentially code that does the simulation plus a relatively small amount of book
keeping overhead). We will specifically explain the small amount of book keeping
overhead notion. ADP methodologies are very different from mathematical pro-
gramming, which runs into the fact that it must declare the alternatives explicitly
before it starts.
 3.2 Dynamic Programming 63

3.1.1
Chapter Structure

The structure of this chapter is as follows:

• First, we would like to familiarize the reader with the central
concept of dynamic programming, which is the value
function.
• Then, we will discuss the computational obstacles of DP,
that are named COD. Therefore, we will review several
state-of-the-art ADP methodologies that tend to minimize
them.
• Next, we will introduce in great detail a high-dimensional
supply chain case study. This case study will be formulated
as a MILP.
• Moreover, we will add some Markovian uncertainty to the
problem. Naturally, to address this additional degree of
complexity, the case study will be formulated as a Markov
Decision Process. Here, we will see the absolute relevance of
the MDP with respect to the MILP formulation.
• The high-dimensional case study is suitable to be addressed
by the proposed Real Time Approximate Dynamic
Programming methodology. Thus, we will delineate the
approach.
• Finally we present simulation results, useful conclusions
and potential future directions.

3.2
Dynamic Programming

We refer the reader to [4] for an excellent introduction on DP. For a very strict
mathematical description of DP we refer the reader to [5].

3.2.1
The Value Function

Very loosely speaking DP is a methodology that converges to a fixed point the value
of an unknown function for all the system states. The converged function is the
so-called optimal value function. By the term system state we mean the minimum
collection of the variables, which completely characterizes the future behavior of
the given system.
“What is so particular of this value function?” Simply, by obtaining this function
the optimal policy for deterministic or stochastic, one stage or multi stage prob-
lems is trivially defined. To retrieve this function, one usually results to numerical
methods. Nonetheless, if we add some structure to problems like linear dynamics,
64 3 A Real Time Approximate Dynamic Programming Approach

quadratic cost function and Gaussian disturbances, then there are analytical expres-
sions (algebraic Ricatti equations) that provide the optimal control (optimal policy).
The Ricatti equations work under some additional restrictions (1) that the system
is controllable, which means that the control policy affects the system states and (2)
the choice of the quadratic function must be proper, meaning that it should mini-
mize the systems unstable modes (this means that the state variables that cannot
be controlled must go to the origin). This sort of formulation is very popular to the
process control engineering society. It is trivial that by relaxing these assumptions,
these analytical expressions do not hold.
As we mentioned the main feature of DP is the “value function” denoted as J ,
which maps the system state st to its resultant expected total reward under some
policy π:

∞ 

π t
J (st ) = Eω γ r(st , π(st ), ω) , γ ∈ [0, 1), (3.1)
t=0
where r(st , π(st ), ω) represents the stage-wise or myopic reward received by exer-
cising π : st → α (α is the control) at a given state st , with the realization of the
stochastic variable ωt . The infinite horizon total reward is discounted by a factor γ .
That means that the future rewards are discounted to the present.
The usage of the discount factor guarantees the convergence of DP, under mild
assumptions. Those assumptions are: (1) Finite state S, action A and probability
space ; (2) Bounded rewards |r(st , α, ω)| < M, ∀(s, α, ω) ∈ S × A × , and
under them DP retrieves the optimal policy π = π ∗ and the optimal value function,
∗
J π (s) (Eq. (3.2)):
∗
J π (st ) = max J π (st ). (3.2)
π∈
∗
The optimal value function J π (s) satisfy the Bellman or optimality equation.
More information about the properties of this linear mapping in [4, 6].
∗ ∗
J π (st ) = max E{r(st+1 , α, ω) + γ J π (st )|St = st }. (3.3)
α∈A
Note that we loosely use the index t to imply the current state and t + 1 the succes-
sive state. By St = st , we imply that the current state is a random variable and st is
one of its realizations. The interpretation is that the optimality equation is always
conditioned on the current state realization st .
As we mentioned, the knowledge of the optimal value function ∀s ∈ S, solves
the stochastic multistage decision making problem. They are classically obtained
through value iteration (VI) or policy iteration (PI) or exact linear programming
(detail description and small numerical examples of these techniques can be re-
trieved in [4] and [6]).

3.2.2
Markov Decision Processes

Our choice of describing the future uncertainty is via Markov chains. We will re-
strict our attention to discrete Markov chains. Markov chain is a discrete stochastic
 3.2 Dynamic Programming 65

process that inherits its name from the Markov property. By the term Markov prop-
erty we mean that the conditional probability distribution of the future states of the
process, given the present state and all past states, depends only upon the present
state and not on any past states.
Before proceeding further, some specification and definitions are in order. This
chapter considers discounted infinite horizon, discrete time, Markov Decision
Processes (MDP). An MDP is formally defined by a tuple (S, A, , r), where S de-
notes the state space, A the action space,  a transition function, and r a stage-wise
reward function.
For the problems of interest, st ∈ S is typically a concatenation of deterministic
and random variables. The state is the collection of all variables which completely
characterizes any system at a given (discrete) time epoch. The latter can transi-
tion according to probabilities prescribed by a Markov chain. The transition of the
deterministic variables may also depend on the realization of their stochastic coun-
terparts. Using subscript t to denote the value of any variable at a time instance t,
and with some abuse of notation, we then write a ∈ A,  : st × α → P (st+1 |st , α),
where P represents a probabilistic mapping. Furthermore, a policy π ∈  : st → α
is a map indicating what action, α, to take for any given system state, st . Here 
represents the set of all admissible policies.
Next we will explore some computational issues concerning a classic dynamic
programming algorithm named “value iteration”.

3.2.3
From Value Iteration to Asynchronous Value Iteration to Real-Time Dynamic
Programming

The description of the classical value iteration algorithm follows:

For a given problem with a finite state space S, action space A and probability
space :
1. Initialize the value function J 0 (st ) ∀st ∈ S.
2. Iterate over all states st ∈ S for all actions α ∈ A using the
Bellman equation (Eq. (3.4)):

 
 
N 
J i+1 (st ) = max r(st+1 , α) + γ P (st+1 |st , α)J i (st+1 )|St = st ,
α∈A  
j =1

r(st+1 , α) = Eω [r̂t+1 (st+1 , αt , ωt+1 |st ))], (3.4)

where i is the iteration index; note that the entire state space
is swept through in each iteration.
3. Terminate Bellman iteration when
J i+1 (st ) − J i (st )∞ , < ε, where ε is a problem specific
constant.
66 3 A Real Time Approximate Dynamic Programming Approach

With the converged value function J ∗ , the optimal control for a given state is given
according to Eq. (3.5).
 
 N 
α ∗ = arg max r(st+1 , α) + γ P (st+1 |st , α)J i (st+1 )|St = st . (3.5)
α∈A  
j =1

The value iteration is a monotonic linear operator and guarantees the convergence
to a fixed point (optimal value function). In mathematics such an operator is called
contraction mapping. This means that irrespective of the initialization (overestima-
tion or underestimation) the value iteration will converge to the same fixed point
after a finite number of iterations. Nonetheless, the rate of convergence is affected
by the quality of the initialization.
It is a well-known fact that the value iteration algorithm is compromised by the
so-called curse of dimensionality (COD). This refers to the proportional growth in
the computational load with respect to |S| and |A|, where the |·| operator represents
set cardinality. For instance, computational load per iteration of VI scales as |S|2 |A|.
As the cardinality of the state and action space tend to grow exponentially with the
dimensions of the state and action variables, the computational requirement can
quickly become unwieldy. For almost all problems of practical interest, |S| and |A|
are too large to admit these exact DP approaches. Therefore, the use of discretiza-
tion and interpolation schemes is unavoidable, with the caveat that convergence
can no longer be guaranteed [7].
The value iteration as presented is named “synchronous”. Variants of the value
iteration are the so called asynchronous value iteration methodologies. Simply, one
uses some heuristic (prioritizing) rules to select the state that if its value function
is updated will result to a fast convergence. A type of asynchronous value iteration
scheme is the episodic learning Real Time Dynamic Programming approach as
proposed by Barto et al. [8]. The RTDP approach utilizes the Bellman equation to
pick an greedy action and the known probability distribution from the Markov
Chain to select the successive state to be updated. The proposed methodology uses
as conceptual basis the RTDP approach as presented at the seminar paper of Barto
et al. [8].

3.2.4
A Review Of Approximate Dynamic Programming Techniques

The evolving stream of the ADP literature aims at reducing the COD of DP with-
out sacrificing the solution quality in a significant manner, in order to facilitate the
usage of ADP concepts to realistic (large-scale) problems. The first and most signif-
icant source of computational burden is associated with |S|. The two other sources
of computational bottleneck are |A|, and the calculation of the expectation operator
within the maximization (or minimization) as seen in Eq. (3.4).

3.2.4.1 Minimizing the COD Concerning |S|

Most of the ADP methodologies attempt to minimize |S| by intelligently sampling
the state space and then building function approximators based on the value table
 3.2 Dynamic Programming 67

of the sampled states. An interesting research question arises: “What is the most
effective way to sample the state space S?” A naive approach would be to employ
uniformly spaced sampling of the state space. Such a scheme is not attractive be-
cause it is subject to exponential growth with respect to the state dimension. A
more efficient discretization scheme derived from the area of statistical design is
orthogonal arrays (OA) based on Latin hypercubes (LH) [9]. LH is a constrained
Monte Carlo scheme. A comprehensive tutorial of LH can be found online in [10].
This scheme has been used as a part of a stochastic DP methodology to solve a high
dimensional large water reservoir network proposed by Cerbellera et al. [11].
Yet another way to sample the state space is to simulate the system under some
a priori available policies, such as heuristics or other known suboptimal policies.
Choi et al. [12, 13] and Lee et al. [14] sampled the state space via simulations of
the system under known heuristic policies. The set of sampled states for which the
value iteration was performed was a vanishing fraction of the entire state space,
thereby removing the major source of the COD. Specifically after the sampling
Lee et al. investigated the performance of two different types of function approx-
imators: feedforward neural networks and non-parametric local approximators (k
nearest neighbor averager). Their conclusions were that the local approximators ex-
hibited more consistent and stable behavior, whereas the neural networks showed
extremely unstable behavior at times during value iteration. This finding was sup-
ported by a mathematical analysis that proved monotonic convergence of the k-
nearest neighbor averager in value iteration, as well as the empirical results from
the case studies involving a Van de Vusse Reactor and a MMA polymerization re-
actor. Note that these systems are characterized by highly non-linear dynamics,
which are common in chemical engineering applications. The approach’s major
limitation is that it does not address the COD associated with action space. Since,
it only evaluates the Bellman equation with respect to the heuristic policies, that
created the sampled state space in the first place. Hence, this approach is limited
for MDPs with relatively small action spaces.

3.2.4.2 Minimizing the COD Concerning |A|

The second source of the COD concerning A was first tackled by MacQueen [15],
where the optimal value function was bounded to eliminate some suboptimal ac-
tions. This action elimination technique has been used in VI and in PI (details in
[4, 6]). There is a recently published approach that addresses effectively this issue.
This approach is named evolutionary random policy search (ERPS) [16]. It turns out
that the structure of the Adaptive Action Set (AAS), which we analyze at a later
section is very similar to the set of actions proposed in ERPS. The main idea of
the AAS is that for each state, the value function update will be restricted to only a
small set of actions.
Their methodology works on the basis of evolutionary policy iteration. In ERPS
they prove that at each iteration the policy does not deteriorate and the learned
policy converges to the optimal policy with probability 1. There are three main
differences between the RTADP algorithm that we will propose and the ERPS:
68 3 A Real Time Approximate Dynamic Programming Approach

1. The proposed approach constructs a value table with an

increasing number of entries starting from an empty value
table, while the ERPS needs to initialize a fixed number of
states;
2. In the RTADP the exploration rate can be tuned via
initialization of the value function, whilst in the ERPS the
exploration rate is fixed from the beginning;
3. The RTADP is based on asynchronous value iteration, while
ERPS on the policy iteration.

3.2.4.3 Minimizing the COD Concerning the Expectation Operator

The classical way of circumventing this computational obstacle is to use Monte
Carlo sampling, while evaluating each decision. We will implement this idea in
our approach later; therefore it will be extensively discussed.
A different very attractive perspective that addresses the same issue is to inter-
change the expectation with the maximization operator at the traditional Bellman
equation by posting these equations around a post state variable. Post state variable
is the state of the system after a decision. A detailed description about this notion
can be retrieved in [17].

3.2.4.4 Value Function Approximators

As an endnote for this paragraph, we must admit that from a theoretical point
of view, accurate approximation of the value function in high-dimensional appli-
cations is an impossible task [18]. Approximating the value function surface with
simplistic basis functions, e.g. piecewise linear functions, is definitely a rough com-
promise, and is justified only if one knows a-priori that the optimal value function
structure.1) Even if one uses non parametric methodologies e.g. k-NN is subject to
the fact that all sample points are close to an edge of the sample.
To be more specific consider 500 data points uniformly distributed in a 10-
dimensional ball centered at the origin. The mean distance to the nearest point
is ≈0.52; this means the data are closer to the boundary than to the point, hence
one must extrapolate from a neighboring sample and not interpolate! That exam-
ple shows that our logic collapses, when dealing with high dimensional spaces.
Extensive discussion of local methods in high dimensions can be retrieved in [18].
The issue of structure of the optimal value function is critical. In fact to derive
such a structure may be the only way to devise fast algorithms for large MDPs. To
speculate and use complex approximation schemes for high-dimensional spaces
that update the coefficients of localized basis function, like multiple adaptive re-
gression splines [11], is currently the best bet to achieve a good online performance.
Nonetheless, this sort of an approximation cannot produce any mathematical guar-
antees for the convergence of the value function surface.

1) In Partially Observable Markov Decision Processes (POMDP) one knows that the optimal value
function is piecewise linear and convex.
 3.3 A High-Dimensional Supply Chain Case Study 69
Import Facilities

BY PRODUCTS

C3
Crude
Oil Top Customers in need of main products

C4

Light R
LN B C5 Unit Type Connected wih
Naphta Reoil
Product Tanks

Pyrolysis
C9 Reformer (R) C3, C4, C5, C9,
Oil Pyoil
Extraction (B) Mix-xyl, Bz, Raf
C10

C9 Tatoray (TT) C10, Bz

Raf Parmex (P) p-xyl
TT

MAIN PRODUCTS Xylene o-xyl , C9

Toluene Tol Fractorization
BENZENE (XF)
Product tanks connected with units
Iso_mix TOLUENE
P

P-xyl
P - XYLENE
Mix_xy
TT
Iso_mix
Mix-Xyl O - XYLENE

Mix-xyl XF
B
Main MIX - XYLENE
Products
Mix-xyl I
I

Fig. 3.1 Flow Diagram of a simplified BTX Supply Chain.

3.3
A High-Dimensional Supply Chain Case Study

In this section, we formulate a supply chain case study as a MILP. This and the next
section serve to establish a firm connection between mathematical programming
and MDPs. Note that the main goal of this chapter is to help the reader under-
stand, how one can circumvent the COD associated with DP. This will be shown at
Section 5, where we present a general ADP methodology.
Having said that, the supply chain case study is inspired from the work of Kuo
and Chang [19]. We address a simplified version of the supply chain of the light aro-
matics (benzene (Bz), toluene (Tol), mixed xylenes (Mx), P-xylene (Px) and O-xylene
(Ox)). Thus, the supply chain is named BTX and is represented at the schematic at
Fig. 3.1. The reader is referred to their work for a more complex paradigm of the
BTX supply chain.

3.3.1
Introduction

The main units of this particular BTX supply chain instance are: (1) one topping
(Top) unit, (2) one reformer (R) unit, (3) one BTX (B) or the so-called extraction unit,
(4) one tatory (TT) unit, (5) one xylene fractorization (XF) unit, (6) one iso xylene (I)
unit, (7) one parex (P) unit, (8) one iso xylene (I) unit, (9) one terminal (T) unit, and
70 3 A Real Time Approximate Dynamic Programming Approach

(10) five customers (C) for the main products. The structure of the supply chain also
involves a facility (Im) that imports raw materials to specific supply chain units.
For visual purposes the connections between the units and their output tanks are
omitted.

3.3.2
Mathematical Modelling of the Supply Chain

The mathematic description accommodates: (a) definition of the sets, (b) material
flow equations, (c) hard constraints, (d) objective function.

3.3.3
Sets

• U = {Top, R, B, TT, XF, I, P , T , C} is the set that contains

all the supply chain units. An element of this set is denoted
as u.
• Fu = {FTop , FR , FB , FT T , FXF , FI , FT , FC }. Each element of
Fu is a set that represents the feed tanks of a specific unit u.
An element of such a set is denoted as s. We will define
explicitly s for each unit at the next subsection.
• Pu = {PTop , PR , PB , PT T , PXF , PI }. Each element of Pu is a
set that represents the products of a specific u. An element of
such a set is denoted p. We will define explicitly p for each
unit at the next subsection.
• Ku,s is the set that defines the operational tasks that unit u
can perform given a specific feed s.
• tp is the set that defines the planning horizon.

3.3.4
Material Balances at Each Tank

Each unit u ∈ U has a specific number of tanks that act as its input and a specific
number of output tanks. The output tanks are as many as the products of a partic-
ular unit. By introducing control volumes at an input and output tanks of each unit
we derive from first principles the following equations:
Control Volume at an input tank of a specific unit u

Yu,s (t + 1) = Yu,s (t) + yi,u,s (t) − xu,s (t),
i
u ∈ U, s ∈ Fu , t ∈ tp. (3.6)
• The amount of feed stocked at the feed tank s at the next
period t + 1 is denoted as Yu,s (t + 1).
• The total amount of feeds added from the supply chain units
i at the feed tank s at period t is denoted as i yi,u,s (t).
(i: This index can take any element in the set U .)
 3.3 A High-Dimensional Supply Chain Case Study 71

yi ,u ,s(t) yu,i,p(t)
xu,s(t) Pru,p (t)
i
Yu,s(t) u YU,On (t)
i

Fig. 3.2 Control volumes on the input and output tanks of each unit.

• The amount of feed that feeds u from the feed tank s at

period t is denoted as xu,s (t).

If there is no connection yi,u,s (t) from a unit i to the tank s, then yi,u,s (t) = 0.
Control Volume at an output tank of a specific unit u

Yu,p (t + 1) = Yu,p (t) + Pru,p (t) − yu,i,p (t),
i
u ∈ U, p ∈ Pu , t ∈ tp. (3.7)
• The amount of product p stocked at its output tank at the
next period t + 1 is denoted as Yu,p (t + 1).
• The amount of product p produced at its output tank at
period t is denoted as Pru,p (t).
• The amount of product p forwarded to the ith unit from its
output tank at period t is denoted as yu,i,p (t).

If there is no connection yu,i,p (t) to a unit i from the tank p, then yu,j,p (t) = 0.
The active connections regarding each unit are represented at Table 3.1. We in-
tentionally use a loose repetitive notation concerning each units feed and output
tanks. In that table the only elements missing for the sake of space are the byprod-
uct tanks, namely C3 , C4 , C5 , C9 , C10 , Raf of the terminal unit.

3.3.4.1 Constraints on yT ,C,(P m) (t)

For the MILP, the market demands for the final and the byproducts of the supply
chain are deterministic and are denoted as Dc,Pm .
The actual amounts of the main products shipped to the customers is allowed
to deviate from the requested demand values Dc,Pm . The amounts of backlog and
surplus are represented by the corresponding variables LC,P m (t) and EC,P m (t) into
the following material balance equation:
Dc,p (t) − Lc,p (t) + Ec,p (t) = yT ,c,p (t), c ∈ C, p ∈ Pm , t ∈ tp. (3.8)
The above equation makes sense only if the following inequalities are present. The
following equations describe the surplus and backlog:
 72

Table 3.1 Active connections concerning the feed and product tanks of each unit in the BTX supply chain.

Unit s i : Feed p i : Product s i : Active Connection p i : Active Connection

Top s1 = {CO} – yi,T op,s (t) = 0∀i ∈ U \{I m} –

R s1 = {LN } p1 = {C3 , C4 , C5 } yi,R,s (t) = 0∀i ∈ U \{I m, T op} yR,i,p1 (t) = 0∀i ∈ U \{T }
– p2 = {C9 } – yR,i,p2 (t) = 0∀i ∈ U \{T , T T }
– p3 = {Re.Oil} – yR,i,p2 (t) = 0∀i ∈ U \{B}
B s1 = {Re.Oil} p1 = {Bz} yi,B,s1 (t) = 0∀i ∈ U \{R} yB,i,p1 (t) = 0∀i ∈ U \{T }
s2 = {P y.Oil} p2 = {T ol} yi,B,s2 (t) = 0∀i ∈ U \{I m} yB,i,p2 (t) = 0∀i ∈ U \{T T }
– p3 = {Mx} – yB,i,p3 (t) = 0∀i ∈ U \{T , XF }
– p4 = {Raf } – yB,i,p4 (t) = 0∀i ∈ U \{T }
TT s1 = {T ol} p1 = {Bz} yi,T T ,s1 (t) = 0∀i ∈ U \{B, I m} yT T ,i,p1 (t) = 0∀i ∈ U \{T }
s2 = {C9 } p2 = {Mx} yi,T T ,s2 (t) = 0∀i ∈ U \{R} yT T ,i,p2 (t) = 0∀i ∈ U \{XF }
– p3 = {C10 } – yT T ,i,p3 (t) = 0∀i ∈ U \{T }
XF s1 = {Mx} p1 = {P x} yi,XF,s1 (t) = 0 yXF,i,p1 (t) = 0∀i ∈ U \{P }
3 A Real Time Approximate Dynamic Programming Approach

– p2 = {C9 , Ox} ∀i ∈ U \{B, I m, I, T T } yXF,i,p2 (t) = 0∀i ∈ U \{T }

– p3 = {I x} – yXF,i,p4 (t) = 0∀i ∈ U \{P }
P s1 = {I x, P x} p1 = {P x} yi,XF,s1 (t) = 0 ∀i ∈ U \{XF } yXF,i,p1 (t) = 0∀i ∈ U \{T }
– p2 = {I so} – yT T ,i,p2 (t) = 0∀i ∈ U \{I }
I s1 = {I so} p1 = {I x} yi,I,s1 (t) = 0 ∀i ∈ U \{P } yI,i,p1 (t) = 0 ∀i ∈ U \{XF }
T s1 = {Bz} – yi,T ,s1 (t) = 0∀i ∈ U \{B, T T , I m} yT ,C,s1
s2 = {T ol} – yi,T ,s2 (t) = 0∀i ∈ U \{B, I m} yT ,C,s2
s3 = {Ox} – yi,T ,s3 (t) = 0∀i ∈ U \{XF, I m} yT ,C,s3
s4 = {P x} – yi,T ,s4 (t) = 0∀i ∈ U \{I m, P } yT ,C,s4
s5 = {Mx} – yi,T ,s5 (t) = 0∀i ∈ U \{I m, T T } yT ,C,s5
 3.3 A High-Dimensional Supply Chain Case Study 73

U
0
Lc,p (t)
NBc,p (t)(Dc,p (t) − Dc,p (t)),
c ∈ C, p ∈ Pm , t ∈ tp, NBc,p (t) ∈ {0, 1}, (3.9)

L
0
Ec,p (t)
NEc,p (t)(Dc,p (t) − Dc,p (t)),
c ∈ C, p ∈ Pm , t ∈ tp, NEc,p (t) ∈ {0, 1}, (3.10)
NEc,p (t) + NBc,p (t)
1.
L (t) and D U (t) are design parameters denoted as the upper and lower bounds
Dc,p c,p
of the amount of main products delivered to customer c in period t.
It will be obvious at the next section that the MILP formulation serves to our
proposed approach as a policy maker. We will demonstrate how our approach will
gradually build on the such policy makers performance.

3.3.5
Reaction and Separation Processes – The Determination of Pru,p (t)

The formulation presented here follows the paper of Kuo and Chang [19]. This
paragraph determines the amount of product p produced at a unit u as defined in
Eq. (3.7). To achieve that we discriminate the units of the supply chain with respect
to the nature of their main operation. They can perform either (a) reaction or (b)
separation processes.

3.3.5.1 Reaction Processes

The set Ua = {Top, TT, I, R} defines all the reaction processing units within the
system. It should be noted that the reaction yields (YDu,s,k,p ) of every product of
each unit is assumed to be dependent upon the feedstock compositions and opera-
tion modes k. They are regarded as design parameters. The material balance of the
reaction processes can be written as:
 
Pru,p (t) = xu,s (t) Binu,s,k (t)YDu,s,k,p , u ∈ Ua, p ∈ Pu , (3.11)
s∈Fs k∈Ku,s

where Fs is the set of all allowable feeds of unit u ∈ Ua . The Ku,s set represents the
operational modes k of unit u for processing feedstock s. Bin are binary variables.
These variables enforce that each feed s will be processed by a certain operational
mode. YDu,s,k,p is a system design parameter describing the reaction yield of prod-
uct p, given the input tank Fs with operational mode k. The yield parameters are
invariant.
Notice, that this equation is not linear, but it can be formulated as such by intro-
ducing a new decision variable. In this case the decision variables are Bxu,s,k (t) =
xu,s (t)Binu,s,k (t). Then Eq. (3.11) can be written in linear form.
In [19] we retrieve the following information. The operability of the supply chain
is ensured, if at least one the reforming and one isomar units are in operation at
each time period. Nonetheless, at this version of this example we will ensure with
logical constraints that every unit must be in operation at each time period and
74 3 A Real Time Approximate Dynamic Programming Approach

only one operation mode can be adopted for the feed streams in each unit. The
logical constraints used in this formulation are similar with the Kuo and Chang
formulation and therefore we will not go into further details.

3.3.5.2 Separation Processes

The set of Ub = {B, XF, P } defines all the performing separation processes units
within the system.
The linear equation that quantifies Pru,p (t) is

Pru,p (t) = xu,s (t)RFu,s,p (t), u ∈ Ub, p ∈ PUb . (3.12)
s∈Fs

The RFu,s,p (t) variables are design parameters for each of these units. Notice that
these units carry a particular operational task.

3.3.5.3 Hard Constraints

The hard constraints associated with the MILP formulation are with respect to (a)
tank dimensions, (b) input flow constraints of the units, (c) transportation con-
straints, and (d) constraints on imported materials.
Tank dimensions: The Yu,mL , Y U are system parameters, that represent low and
u,m
upper bounds on the tank dimensions.
L U
Yu,m
Yu,m (t)
Yu,m , u ∈ U, m ∈ (Fu ∪ Pu ). (3.13)

Input flow constraints: The FBL U

u , FBu are system parameters that represent low
and upper bounds on the input flows of the separation units:

FBLu
xu,s (t)
FBU
u, u ∈ Ub, s ∈ Fu . (3.14)
s

The hard input flow constraint for the reaction processes are:
 
FBLu
Bxu,s,k (t)
FBUu, u ∈ Ua, s ∈ Fs . (3.15)
s∈Fs k∈Ku,s

Transportation constraints: The process materials are transferred from a unit to

another. The corresponding transportation capacities are in practice, the following
inequalities are included in the model:
L U
yu,i,p
yu,i,p (t)
yu,i,p , u, i ∈ U, p ∈ Pu . (3.16)
L , yU
yu,i,p u,i,p are design parameters.

Constraints on the imported materials: The model also includes constraints on

the quantity of the imported materials. Those are expressed via the following in-
equalities:

yILm,u,s
yI m,u,s (t)
yIUm,u,s , ∀u ∈ U, s ∈ Fu . (3.17)

yILm,u,s , yIUm,u,s are system parameters.

 3.3 A High-Dimensional Supply Chain Case Study 75

3.3.6
Decision Variables

The decision variables for this problem are as follows:

• yi,u,s (t), i, u ∈ U, s ∈ Fu
• yu,i,p (t), i, u ∈ U, p ∈ Pu
• xu,s (t), u ∈ U, s ∈ Fu

3.3.7
Objective Function

The MILP takes into consideration the dynamics of the system under a specific
scenario and maximizes the reward over a specified horizon h.

h
max r(t) , (3.18)
t=1
r(t) is the net profit represented as:
r(t) = p(t) − c(t), t ∈ tp. (3.19)

3.3.7.1 Net Profit p(t)

This variable represents the net income for this system at each time period. It
represents the satisfaction of the demand of the main and byproducts products of
the supply chain. The revenue produced at t is

p(t) = yT ,c,p (t)SPT ,p (t), ∀p ∈ (Pm ∪ Pb), c ∈ C, (3.20)
c p

SPT ,p (t) is the selling price of each product at time t.

3.3.7.2 Net Cost c(t)

We define two sources of cost factors for this system. These are: (a) the total in-
ventory cost (Cs(t)), (b) the total backlog cost (Cb(t)). These are described by the
following equations (Eqs. (3.21, 3.22)):
• Storage costs
 
Cs(t) = Yu,s (t)Cu,s + Yu,p (t)Cu,p ,
u s u p
u ∈ U, s ∈ Fu , p ∈ Pu , (3.21)
Cu,s (Cu,p ) is the storage cost associated with a particular
feed (product) s(p) tank at unit u.
• Backlog cost

Cb(t) = (Dc,p (t) − yT ,c,p (t))+ Cbmp ,
c p
c ∈ U, c ∈ C, p ∈ Pm , (3.22)
Cbmp is the invariant backlog cost for a specific p.
76 3 A Real Time Approximate Dynamic Programming Approach

3.4
Formulating the Problem as an MDP

The formulation of this problem as a formal MDP includes the following ele-
ments: state variables, decision variables, transition function, and objective func-
tion. These elements can be identified from the mathematical programming for-
mulation.

3.4.1
State Variables

Given a mathematical programming formulation, the state variables are those in-
dexed by time, t, and which have equality constraints indicating how the variable
at t + 1 should be calculated from the value at t. These are supplemented by the
variables that evolve randomly according to a Markov process (namely SPT ,p and
Dc,p , c ∈ C, p ∈ Pm ). For the deterministic problem these random variables are
predetermined constants at each period of the time horizon.
For the BTX supply chain instance modeled as an MDP, the state variable is the
high-dimensional vector defined below:
 
Yu,s (t)
 
 Yu,p (t) 
 
SPT ,p (t) =Selling price of the main products to the customers
st =   (3.23)
 at time period t. 
 
 Dc,p (t) = Demand of the main products of the customers 
at time period t.

3.4.2
Decision Variables

The decision variables of the MDP coincide with the decision variables of the MILP
formulation as described in Section 3.3.6.

3.4.3
Transition Function

The stochastic parameters are SPT ,p and Dc,p . These random variables are jointly
modeled with a first-order Markov model. This probabilistic model represents the
probability distribution that governs the transitions among given discrete sets of
values.
The chosen model proposes four joint states for the main products except
the mixed xylene. Each state is represented jointly by a particular demand and
price realization SPT ,p and Dc,p . For instance, the probability to transition from
SPT ,Bz = 85 and Dc,p = 100 to SPT ,Bz = 85 and Dc,p = 85,000 is 0.8.
Notice from Fig. 3.5 that the demand states of the main products transition from
very low values to pretty high values. These demand outliers will make the MILP
 3.4 Formulating the Problem as an MDP 77

with deterministic future not to be optimal with respect to a big number of scenar-
ios. Since the sample path MILP actions are optimal only for a particular scenario.
Such a fact is demonstrated via our numerical experiments in Section 3.6.
In the following section we will provide a brief discussion of how one can use
identification techniques and estimate its elements.
The key element in modeling the systems state transition is to take into account
the physical constraints as described in Eq. (3.6) and Eq. (3.7) among the interde-
pendent units and also the probabilistic state transitions of the random variables.

3.4.4
Objective Function

The objective in this problem is to maximize the discounted expected profit over an
infinite horizon. This will be accomplished, when we find a stationary2) decision
function π(st ) such that each state is mapped to the best possible action. The total
return of a policy π starting from an initial state s0 is

∞

F0π = γ t r(st , π(st ))|s0 ), (3.24)
t=0

r(st , at ) is defined at Section 3.3.7. Trivially, the optimal policy π ∗ as discussed

before is the one instructed by the optimal value function:
 
π ∗ = arg max F0π . (3.25)
π∈

3.4.5
Concluding Remarks

In conclusion, one can observe a very natural correspondence between mathemati-

cal programming and markov decision process formulations. When addressing an
MDP the state variables are a concatenation of deterministic and random variables
and the state transitions are defined over a probabilistic space. Moreover, the ob-
jective is to maximize the reward over an infinite horizon, while in mathematical
programming the horizon length is predefined. The decision variables of a given
problem are the same in both formulations. Next, we analyze the proposed real
time approximate dynamic programming algorithm.

2) There is ambiguity regarding the nature of the resulting policy (stationary or not stationary). Since,
if one uses ADP methods that utilize only a vanishing fraction of the entire state space the resulting
policy may not be stationary [20].
78 3 A Real Time Approximate Dynamic Programming Approach

3.5
A Real-Time Approximate Dynamic Programming Algorithm

Real-Time Dynamic Programming (RTDP) is an asynchronous dynamic program-

ming approach based on -greedy simulation. RTDP operates over finite duration
trials, meaning that each trial starts from a starting state within the value table and
accumulates valuable learning along its trajectories. After a finite number of trials
the visited states that are involved in trial trajectories become saturated and belong
to a closed set of states [21, 12], because the actions taken from states are heuristi-
cally limited. We refer the reader to [8] for the definition of the simulated relevant
state space denoted as Ssim : This set contains all the states that belong to the trajec-
tory of the optimal policy. There is no algorithm that can exactly identify this set of
states for general problems, except full dynamic programming. Therefore the main
function of our framework is to approximate Ssim . In this section, we describe our
variant, RTADP, of the RTDP approach. The original version RTDP was developed
in Barto et al. [8].
For an idea of the computational savings accrue by RTDP, we refer the read-
ers to [8]. The numerical illustration of the RTDP approach involves a stochastic
shortest path application, also known as race track problems, with 9115 discrete
states. Results indicate that the value functions converge and instructs the optimal
policy, when only 2% of the total states gets updated more than 100 times, 20%
gets updated more than 10 times, and 3% of the spate space is not updated at all.
This inspires further development of a computationally tractable ADP architecture
based on the fact that not the entire state space needs to be updated, in order to
construct a high quality policy.
The key issue in RTDP is that comprehensive exploration of the full state and
action spaces must be avoided in order to avoid unmanageable computation. How-
ever, enough exploration must be attempted in order to find good policies to exploit,
and to enable reasonable approximations of the performance of actions in parts of
the state space that might not have been encountered in the initial phases of the
algorithm. To orient the reader, a high level description of the RTADP approach is
given. First, an initially empty value table is populated with entries using a -greedy
episodic RTDP. Next, we have to define a mechanism to tune the exploration vs ex-
ploitation tradeoff. The goal of exploration is to retrieve more valuable entries for
the evolving value table (Ssim ). Generally, this Ssim improves on the previous at-
tained performance, since better actions and states will be added to the table. In
exploitation the values of the actions and states are refined to generate better esti-
mates. In [21], this tradeoff was controlled by choosing a particular value function
initialization scheme. A different way to control this tradeoff is to use a type of a
simulated annealing schedule and bias selection at the beginning of the simulation
towards random rather than greedy actions. Last, each state’s value function update
requires the evaluation of a subset of the action space termed the adaptive action
set. During the evaluation we suggest the use of a non parametric approximator
 3.5 A Real-Time Approximate Dynamic Programming Algorithm 79

based on the one from [22]. This approximator has some guarantees that the value
function approximations at least will not diverge.
Thus, the RTADP is a forward dynamic programming approach; it combines ele-
ments of the classical RTDP approach and several novel notions including; “evolv-
ing set of relevant states”, for which the value function table is built, an “adaptive
action set” to minimize the COD concerning the Action space, a “non-parametric
k nearest neighbor (k-NN) averager” for value function approximation, and Monte
Carlo sampling to eliminate the exact computation of the one step expectation op-
erator when evaluating each action.

3.5.1
The Greedy RTADP Algorithm

The procedure below samples the state space using a greedy policy and constructs a
value table denoted as Ssim starting from an empty one by gradually adding entries,
as states are encountered in the simulation. The following steps are involved in
each iteration of the algorithm and a schematic representation appears in Fig. 3.3.

While |Ssim | < , where is defines the maximum number of states we allow the
constructed value table to have
For iterations i = 1, 2, . . . , h, where h is the horizon length of each episode.
Step 1 Start from a random state st ∈ Ssim .
Step 2 Construct set of actions (denoted by Asub ) for st . Asub ⊂ A, where A is the
set of all possible controls that the decision maker can exercise at any time
instance for a given state.
Step 3 Update the J π (st ) after evaluating every control in Asub . The decision-
maker follows a policy that is greedy with respect to the most recent estimate
of the value table (Eq. (3.5)). This evaluation (Eq. (3.4) requires knowledge of,

Adaptive Action Set ( )

sub

2
Candidate optimal action for Initial state
Possible successive next state
3
1
si Evaluate sub using 4
*
Bellman equation

Monte Carlo
7 Sampling
5
st st 1

st 1

6
Sim Sim st 1 or Sim Sim

Fig. 3.3 Schematic representation of sequential calls on RTADP algorithm.

80 3 A Real Time Approximate Dynamic Programming Approach

if not an estimation of, Jiπ (st+1 ) for all possible successor states st+1 from
st . Further details concerning this are outlined in the following subsection.
Step 4 Pick the greedy control and sample state st+1 is sampled according to the
probability distribution p(st+1 |st , α ∗ (st )) as defined from the Markov model
of the random variables.
Step 5 Check a parameterized distance (Section 3.5.3) to identify k neighboring
registered entries in the simulated value table Ssim . If (k < ) → Ssim =
Ssim ∪ st+1 else→ Ssim = Ssim EndIf
Step 6 Set t = t + 1 and we go back to Step 1.
End For
End While

If the algorithm circulates over a small cyclic graph of states, the algorithm
is restarted from a random state si ∈ Ssim . Empirically, one terminates with
π (s ) − J π (s ) < ε ,
(Ji+1 ∀si ∈ Ssim ⊆ S) like in VI, where ε is a tolerance
i i i ∞
parameter. The user can apply this termination criterion, only if the state space is
saturated and the number of entries does not grow.
To implement this approach, the user will need to choose several parameters,
like , , or the distance metric. The parameters are tailored by trial and error
for each application. The earlier statement that ADP usually consists of simulation
plus a relatively small amount of book keeping overhead can now be clarified. This
small amount of book keeping is essentially the value table and their associated
value function estimates.
In the next subsections we describe with detail the steps of the RTADP approach.

3.5.2
Key Elements of Asub

Recall that the purpose of constructing Asub is to significantly reduce the “curse of
dimensionality” (COD) arising from having to evaluate the entire action space A.
Simulation results [21] indicate that after a number of iterations the quality of the
approximation does not deteriorate and converge to good approximations of the
optimal value functions.
The key elements of the AAS are:
1. Mathematical programming actions. If we can describe the
problem of interest with a deterministic mathematical
model (e.g., MILP), then it is highly advisable to try actions
resulting from deterministic math program formulations.
This typically constitutes a suboptimal policy.
2. Best known actions. This action is a product of the a-priori
learning with respect to the value functions of all states in
the “evolving” value table. If the state to be updated is a state
never visited before, then its best known action is empty. If
the state has been revisited, a best known action should have
 3.5 A Real-Time Approximate Dynamic Programming Algorithm 81

been stored with respect to the prior estimate of the value

function.
3. Random actions. Random controls ensure that we effectively
explore the entire action space and exclude the possibility of
not visiting any portion of the state space. We generated the
random controls by random perturbations to the heuristic
and mathematical programming actions for state si .
4. Other candidate optimal actions, as a part of the Asub , are
the best known actions of the k − NN of state si or actions
coming from genetic algorithms.

3.5.3
On Calculating Jtπ (st )

Exact DP techniques (e.g.,VI, PI) can be applied only to MDPs with finite S, A
and require an initial estimation of the value function ∀s ∈ S. The optimal policy
created by these techniques is not sensitive to any initial estimation scheme. Since
the proposed approach does not consider the entire state space in order to address
large MDP problems, it foregoes the formal convergence of the value functions.
This section provides rules of thumb for structuring such an estimation scheme,
so as to achieve empirical convergence and maximize the performance.
In Eq. (3.4) the calculation of Jtπ (st ) involves the knowledge of the value func-
tion of all possible successive states Jtπ (st+1 ) for each action in Asub . During this
calculation we will encounter one of these three possible scenarios. A schematic
representation of the following scenarios appears in Fig. 3.4.
• Scenario 1. All st+1 s have values registered in the value table.
We use these values to calculate Jtπ (st ).
• Scenario 2. Some of the st+1 s are not found in the value table.
In this case we first need to find the set of states within δ
distance of st+1 (denoted by Nδ (st+1 )). Here we use the a
Euclidean distance metric d, with a user defined design
parameter δ.

  
def
F ind Nδ (st+1 ) = s ∈ Ssim : (s − st+1 )T W (s − st+1 ) < δ ,

(3.26)

where W is a feature weighting diagonal matrix. It will be

apparent in Section 3.6.1 that before applying the previous
equation we use a filter with two characteristics to reduce the
search space of the nearest neighbors. If |N δ (st+1 )|  k,
then we can approximate the value function of st+1 from the
k nearest states, by utilizing a local k-nearest neighbor
82 3 A Real Time Approximate Dynamic Programming Approach

approximator. The mathematical expression for

approximating the value function for each st+1 is as follows:

1 
Jtπ (st+1 ) = Jtπ (x), (3.27)
k
x∈Nk (st+1 )

where Nk (st+1 ) denotes the set of states representing the k

nearest neighbors. The value k = 4 was used for the
numerical results in [22] where the authors proved the
numerical stability and convergence of the approximated VI
using the k − NN as approximation scheme.
In the case that sj has k < k = 4 neighbor states
within Ssim , we utilize Eq. (3.27) with the k number states to
approximate Jiπ (sj ).
• Scenario 3. The case where |Nδ (st+1 )| = 0. Apparently
Eq. (3.27) cannot be used, and therefore we suggest an initial
estimation scheme of underestimating the value functions
with respect to the optimal one. Nonetheless, in this
application the Jtπ (st+1 ) of the st+1 s that belong to Scenario
3 are initialized with 0. The reader is prompted to [21] for an
extensive study on the exploration versus exploitation due to
the initialization with respect to the Scenario 3. Taking this
further, the decision maker may introduce a priori
knowledge, similar to [12, 13].

In Eq. (3.4) the calculation of Jtπ (st ) involves the knowledge of the value func-
tion of all possible successive states Jtπ (st+1 ) for each action in Asub . During this
calculation we will encounter one of these three possible scenarios. A schematic
representation of the following scenarios appears in Fig. 3.4.

3.6
Simulation Results

Simulations generate a significant volume of data and for the purposes of this
chapter we will focus on the relative performance of three possible algorithms
rather than the specific action and state trajectories that are generated. Section 3.6.1
presents, (a) the simulation procedure, (b) how Asub is numerically constructed
and, (c) the parameters that characterize the creation of the neighborhood for st . In
Section 3.6.2 we evaluate the performance of RTADP relative to an “ideal” solution
derived from an MILP formulation, and a heuristic solution based on a sample
path MILP. An “ideal” solution is one produced by the deterministic MILP when
the random variable values are known beforehand. A sample path MILP is one
where the random variables are sampled at specific values, but then the resulting
actions are evaluated against a broader sample of scenarios.
 3.6 Simulation Results 83

st 1
st 1
st 1

s st
st 1
1

Scenario 1
Scenario 2
Scenario 3

Sim

Fig. 3.4 Schematic representation of a state s going under a

decision to st+1 . During this transition we will encounter one
of the 3 scenarios corresponding to a legal state transition
inside the state space. S is the entire state space. Ssim is the
already sampled state space from the -greedy simulation.

3.6.1
Simulation Procedure

The value of was set to 10,000 states and each simulation trial was executed
with a horizon length of h = 13 time periods. We chose this number, because the
performance of this approach will be compared to other methodologies over a three
time period horizon. The periods beyond the first three influence the decisions,
which is the reason for using a longer horizon for each episode. A discount factor
(γ = 0.9) was used, which makes the impact of any states beyond 10 future time
periods negligible.
The discrete event simulation of the supply chain system starts given a value
table with one entry. At each trial one can choose as starting state any state within
the table. For convenience, each trial starts with the same st .
To construct the Asub at each loop we include: 10 “legal” random actions, 10
sample path MILP actions, a best known action, when it is available and the k best
stored actions from the k nearest neighbors of st (k = 4). The term legal implies
that one must be almost certain that the given action will not violate any hard
constraints on operation, in this case the tank constraints. Specifically, each action
leads to 512 successive realizations,3) each of which is accompanied by a certain

3) The successive states are 512, since the random variables are 5. The first four have four states,
while the fifth enumerates only two states.
84 3 A Real Time Approximate Dynamic Programming Approach

probability. If the cumulative probability of illegal states due to this action is more
than 10% then the action is considered to be illegal and it is not included to the
adaptive action set. Given this one can make sure that the RTADP will not violate
frequently those tank constraints.
Next, we describe the parameters δ and W that define the neighborhood of st .
The Euclidean measure δ demarcates the “neighborhood” of a particular st ; the
thus defined neighboring states necessarily have existent entries in the evolving
value function table.
We choose δ in a way that neighboring states lie solely in one dimension. To be
more specific a neighboring state sn , with respect to st , belongs to a set of states Sn ,
that has the following characteristics.
Characteristic 1: States such that all tank levels are the same bar one particular
tank level Y , and sn is within 100 gallons of the corresponding tank level of st .
Characteristic 2: The same realization of the random variables.
The value table can be filtered to find states that satisfy the above two characteris-
tics, e.g., search the table for existing states sn within Ysn = Yst ±100, or, given these
characteristics one can quantify W and δ. Those entries are then used in Eq. (3.26)
to infer the value function of st .

3.6.2
Performance Comparison

The parameters used to generated the following results are identical with [19]. Due
to space constraints we refer the reader to retrieve those in their paper in Tables 9 to
12. The states of the random parameters follow the probability transition matrices
in Fig. 3.5.
A way to measure the performance of the proposed architecture is to let it pro-
vide the controls for the system for a large number of time periods and find an
average performance over a smaller number of periods repeatedly sampled from
this longer trace. In this example we could not run the MILP for a very large num-
ber of time periods, since it will become infeasible as the random variables diverge
from the values assumed to calculate the actions of the MILP. In addition it is un-
likely that the MILP actions will be retained for long periods if there is significant
divergence between the assumed parameter values and their actual realizations, a
rolling horizon approach is more reasonable. Hence, we run the MILP for three
time periods and consider 500 out of the 5122 total possible scenarios.
The methods were tested over the same 500 different scenarios and Table 3.2
summarizes the results. The full information MILP solution represents an ideal-
istic policy tailored for that scenario. The comparison between the MILP versus
RTADP shows the performance gap of 13.1% in favor of the MILP per three
time periods. This performance gap represents both the lack of optimality of the
RTADP and the value of perfect information on what will happen in the next three
time periods. At this time we cannot separate these two effects, but this provides
an upper bound on the quality of the solution that can be obtained through the
RTADP approach.
 3.6 Simulation Results 85

Dc, Bz 1.8 55 10 35 10 3 SPT, Bz 85 85 91 91 10

Dc,Tol 1.6 30 20 40 10 3 SPT,Tol 65 65 70 70 10
Dc, Px 8 40 50 25 10 3 SPT, Px 68 68 72 72 10
Dc,Ox 25 45 5 25 10 3 SPT, Bz 65 65 680 68 10
Dc, Mx 1 1 10 3 SPT, Bz 540 570 10

0 .1 0 .8 0.08 0.02 0 .1 0 .8 0.08 0.02

0.06 0 .1 0.8 0.04 0.06 0 .1 0.8 0.04
PDBZ SPBZ PDTol SPTol
0.02 0.08 0.1 0 .8 0.03 0.07 0.1 0 .8
0.8 0.04 0.06 0 .1 0.8 0.04 0.06 0 .1

0.1 0.8 0.08 0.02 0.1 0.8 0.08 0.02

0.06 0.1 0.8 0.04 0.06 0.1 0.8 0.04
PDPx SPPx PDOx SPOx
0.02 0.08 0.1 0.8 0.02 0.08 0.1 0.8
0.8 0.04 0.06 0.1 0.8 0.04 0.06 0.1

0 . 2 0 .8
PDMx SPMx
0 .9 0 .1

Fig. 3.5 Probability transition matrices and state variables concerning the uncertain variables.

We can develop some intuition about the importance of the multistage stage
uncertainty by comparing the RTADP solution with that generated by an MILP so-
lution that chooses actions based on one particular realization of the future. The
sample path MILP is optimal only with respect to one particular scenario. Its av-
erage performance is derived by applying the same action trace to many different
scenarios. Nonetheless, on average its action trace is feasible only for only 20% of
the sampled 500 scenarios. The infeasibility lies at the tank of the main products.
The performance calculated for the sample path MILP at Table 3.2 is for the entire
scenario population. The infeasibility penalty is explicitly calculated by the excess
storage cost at the main product tanks. On average these solutions perform 27%
per 3 time periods worse that the MILP approach and 16% worse than the RTADP
approach given that we start from the same state.
We tried to implement an MILP approach using expected values of the random
variables, but its performance cannot be judged appropriately since by applying its
solution, on average only 2 out of 500 scenarios do not violate the tank constraints
for the main products at the terminal unit.
The performance of the RTADP is dependent on the number of sampled states.
As the number of sampled states increase, we gain a finer resolution of the value
function surface. A nice feature of RTADP is that by appropriate creation of the
adaptive action set we can guarantee feasibility.In this instance we sampled 1137
states since the rate of increasing the cardinality of Ssim was significantly reduced.
Also, we acknowledge that the number of random scenarios is significantly greater
than this.
86 3 A Real Time Approximate Dynamic Programming Approach
Table 3.2 Evaluating the average performance per three time
periods on 500 scenarios of (a) MILP with full information,
(b) RTADP – 1137 sampled states, (c) sample path MILP.

Comparison Average Performance per 3 Time Periods ×108 Performance Gap

MILP full information 1.91 ± 0.13 –

RTADP-1137 1.66 ± 0.17 13.1%
MILP sample path 1.39 ± 0.33 27.2 %

3.7
Conclusions

When one wants to address multistage stochastic problems, the main tools are DP
and stochastic mathematical programming. In this chapter, we delineated the com-
putational obstacles of dynamic programming and developed an approximate dy-
namic programming approach for a BTX supply chain problem. We did not address
the stochastic programming alternative, or approximations that can be developed
based on it [23, 24].
The real-time approximate dynamic programming approach as presented here is
attractive, because one can superimpose it on a range of decision-making method-
ologies from complex and computationally demanding deterministic mathemati-
cal programming or simple heuristics or even exploratory randomized actions. It is
not clear how much effort to devote to developing good initial action and state tra-
jectories versus allowing random exploration to find them. The proposed iterative
methodology promises to choose the greedy control at each time period and to not
discard valuable information about how that choice plays out in the future. More-
over, because this methodology is iterative and is based on asynchronous value
iteration it can improve online performance as the number of iterations increase.
The on-going research involves to couple this RTADP approach with local regres-
sion methods to minimize the COD concerning the k-nearest-neighbor local value
function approximation scheme. Also, one could use mathematical programming
to bound the value function, when we evolve the value table and take advantage
existing in the literature ways that those bounds guide the exploration vs the ex-
ploitation in RTDP [25]. Finally, modifying this approach to solve problems involv-
ing multiple agents is another topic of interest, since these problems inherit the
challenge of a non-stationary environment.
 References 87

References

1 Shapiro, A., Nemirovski, A., On salesman problem. Comput. Chem.

complexity of multistage stochastic Eng. 28(8) (2004), pp. 1285–1296.
programs. http://www.optimization- 13 Choi, J., Lee, J. H., Realff, M. J., An
online.org. algorithmic framework for improv-
2 Birge, J. R., Louveaux, F., Intro- ing heuristic solutions: Part II. A new
duction to Stochastic Programming, version of the stochastic traveling
Springer, Berlin, 1997. salesman problem. Comput. Chem.
3 http://stoprog.org/index.html/ spin- Eng. 28(8) (2004), pp. 1297–1307.
troduction.html 14 Lee, J. M., Kaisare, N. S., Lee, J. H.,
4 Putterman, M. L., Markov Decision Choice of approximator and design of
Processes: Discrete Stochastic Dynamic penalty function for an approximate
Programming, Wiley, New York, 1994. dynamic programming based con-
5 Stokey, N. L., Lucas, R. E., Prescott, trol approach. J. Process Control 16(2)
E. C., Recursive Methods in Economics (2006), pp. 135–156.
Dynamics, Harvard University Press, 15 MacQueen, J., A modified dynamic
Cambridge, MA, 1989. programming method for Markov-
6 Bertsekas, D. P., Dynamic Pro- ian decision problems. J. Math. Anal.
gramming and Optimal Control, vol. Appl. 14 (1966), pp. 38–43.
2, Athena Scientific, Nashua, New 16 Ramezani, M. C., Hu, V., Fu, J., Mar-
Hampshire, 1995. cus, S. I., An evolutionary random
7 Thrun, S., Schwartz, A., Issues in policy search algorithm for solving
using function approximation for re- Markov decision processes. University
inforcement learning. Proceedings of of Maryland, ISR Technical Report.
the Fourth Connectionist Models Sum- 17 Powell, W. B., Approximate
mer School (Hillsdale, NJ), Lawrence Dynamic Programming for
Erlbaum, Mahwah, NJ, 1993. Operations Research: Solving
8 Barto, A., Bradtke, S., Singh, S., the curses of dimensionality.
Learning to act using real-time dy- www.castlelab.princeton.edu. (Email
namic programming. Artif. Intell. 72 [email protected] for a link to a
(1995), pp. 81–138. review copy), 2007.
9 Cervellera, C., Chen, V. C. P., Wen, 18 Trevor, H., Tibshirami, R., Fried-
A., Optimization of a large-scale wa- man, J., The Elements of Statistical
ter reservoir network by stochastic Learning, Springer, Berlin, 2001.
dynamic programming with efficient 19 Kuo, T. H., Chang, C. T., Optimal
state space discretization. Eur. J. Oper. planning strategy for the supply
Res. 171(3) (2006), pp. 1139–1151. chains of light aromatic compounds
10 http://www.cs.sandia.gov/DAKOTA/ in refineries. Accepted at Comput.
papers/LhsManual2004. pdf) Chem. Eng.
11 Tsai, J. C. C., Chen, V. C. P., Beck, 20 McManan, B. H., Likhachev, M.,
M. B., Chen, J., Stochastic dynamic Gordon, G. J., Bounded Real-Time
programming formulation for a Dynamic Programming: RTDP with
wastewater treatment decision- monotone upper bounds and perfor-
making framework. Ann. Oper. Res., mance guarantees. Appearing in Pro-
Special Issue on Applied Optimization ceedings of the 22 International Confer-
Under Uncertainty 13, (2004). ence on Machine Learning (Bonn, Ger-
12 Choi, J., Realff, M. J., Lee, J. H., An many), 2005.
algorithmic framework for improving 21 Pratikakis, N. E., Realff, M. J., Lee,
heuristic solutions: Part I. A deter- J. H., Strategic capacity decisions in
ministic discount coupon traveling
88 3 A Real Time Approximate Dynamic Programming Approach

manufacturing using stochastic dy- Chem. Res. (2004), pp. 3695–3713.

namic programming. Under Review http://egon.cheme.cmu.edu/Papers/
at Naval Res. Log. BalasubMultistage.pdf
22 Lee, J. M., Lee, J. H., Approximate 24 Balasubramanian, J., Grossmann,
dynamic programming strategies and I. E., Scheduling optimization under
their applicability for process control: uncertainty – an alternative approach.
A Review and future directions. Int. Computers & Chemical Engineering
J. Control Auto. Syst. 2(3) (2004), pp. 27(4) (2003), pp. 469–490.
263–278.
25 Smith, T., Simmons, R., Focused
23 Balasubramanian, J., Grossmann, real-time dynamic programming
I. E., Approximation to multistage for MDPs: Squeezing more out of a
stochastic optimization in multi- heuristic. National Conference on Artifi-
period batch plant scheduling un- cial Intelligence, 2006.
der demand uncertainty. Ind. Eng.
 89

4
Robust Supply-Chain Operations through Rescheduling
Arief Adhitya, Rajagopalan Srinivasan, Iftekhar A. Karimi

4.1
Introduction

Today’s supply chains (SCs) operate in a complex environment. Globalization has

expanded the coverage of SCs with suppliers and customers distributed around
the globe, numerous modes of logistics are needed, and outsourcing introduces
more players and links into SCs. Customers are increasingly demanding and com-
petition is rife. Amidst this complexity, disruptions arising from acts of nature or
human-caused incidents occur frequently. In recent years, we have witnessed hur-
ricanes, earthquakes, floods, SARS, avian flu, labor strikes, terrorist attacks, wars,
transport accidents, etc. In today’s highly connected world, these events adversely
impact SC operations, directly and indirectly, and pose numerous challenges.
Some recent real-world examples bring forth the multifaceted nature of these
challenges. In 2000, a fire, which lasted only 10 min in a Philips semiconductor
plant in Albuquerque, New Mexico, tested the SC resilience of two of the world’s
largest mobile communications companies. Nokia and Ericsson depended on this
plant to produce radio frequency chips, a key component in their mobile telephone
handsets. The fire ruined eight trays of silicon wafers, which were enough to pro-
duce chips for thousands of handsets. But more importantly, it damaged the plant’s
equipment and that it would take many weeks for the plant to resume normal pro-
duction. The varying effect of this incident on Nokia and Ericsson is instructive.
Nokia was quick to detect the problem in the flow of chips even before being in-
formed by Philips of the incident and took immediate measures to secure alternate
sources of radio frequency chips. In contrast, Ericsson was slow to react even after
Philips notified them. Due to the delay, they were not successful in their efforts
to find alternate supplies. As a net result, due to good supply chain visibility and
timely response, Nokia was able to tide through the supply disruption and meet
production targets while Ericsson reported a major loss of US$1.68B that year and
lost significant market share, mostly to Nokia [1].
Another similar dichotomy was visible in the automobile industry as it struggled
to recover from the consequences of the September 11, 2001 terrorist attacks. The

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
90 4 Robust Supply-Chain Operations through Rescheduling

immediate and sharp increase in security at US borders with Canada and Mex-
ico led to long delays in supply flows, which affected all major automobile man-
ufacturers. Ford and Daimler–Chrysler, like many other automobile manufactur-
ers relied heavily on just-in-time inventory systems and frequent replenishment of
components and sub-assemblies to ensure steady but efficient production. Due to
the restricted traffic, shipments of engines and other parts to US plants stalled at
the borders. Daimler–Chrysler reacted promptly to the disruption by arranging for
parts that were normally air-shipped to be transported by an expedited truck ser-
vice. Ford, on the other hand, did not make major course corrections to manage
the disruption. As a result, in contrast to Daimler–Chrysler, Ford produced 13%
fewer vehicles than planned in the fourth quarter of 2001 [1].
The chemical industry has not been insulated from disruptions. More recently,
in August 2005, Chevron Oronite, one of the world’s largest suppliers of lube ad-
ditives, issued a global force majeure notice and stopped product shipments. Sev-
eral supply chain factors contributed to this: rising global demand and tightened
availability of raw materials, a flash fire in their Singapore plant, and 6-week shut-
down of their plant in Belle Chasse, Louisiana due to Hurricane Katrina. The con-
sequence of this confluence of “rare-events” extended beyond Chevron Oronite,
other lube additive suppliers had to implement product rationing and the industry
was tipped into a crisis.
It is evident from the above examples that numerous eventualities can lead to
supply chain disruptions. Disruption can be defined as any event or situation that
causes deviation from normal or planned operations of entities and flows. Exam-
ples include operational difficulties, emergency shutdown, natural disasters, terror-
ist incidents, industrial actions (such as strikes or protests), accidents (in-plant or
during transportation), or uncertainties in supply and demand. Disruptions could
bring about adverse effects, such as blockage of material, information, and finance
flows, loss of production, off-spec products, loss of efficiency, under- or over-supply,
etc. All of these directly or indirectly translate into operational problems, loss of
business opportunities, and financial losses. There is therefore a need for decision
support systems to aid in disruption management and enable robust SC opera-
tions.
A supply chain is essentially a network of entities exchanging material, informa-
tion, and finance flows. One important part of supply chain management (SCM)
is scheduling activities of entities and exchange of flows among them. A deviation
between plan and actual realization occurs during disruptions. Depending on the
magnitude of the deviation, the necessary response would have to differ. This mo-
tivates rescheduling, the focus of this chapter. The petroleum refinery SC is used
to identify the salient characteristics of the problem and review solution methods.
The crude oil supply chain is described in Section 4.2 and the rescheduling prob-
lem stated in Section 4.3. Existing literature on rescheduling is also reviewed there.
In Section 4.4, we review two rescheduling methods and illustrate them through
case studies in Section 4.5.
 4.2 Refinery Supply Chain 91

4.2
Refinery Supply Chain

Petroleum refining can be considered as the mother of the chemical industry. A

petroleum refinery SC as illustrated in Fig. 4.1 begins from oil reservoirs. Refining
is a complex process involving a series of operations to transform crude oil into
valuable products, such as gasoline, heating oil, jet fuel, lubricants, and intermedi-
ates for further processing to produce plastics, paints, fertilizers, detergents, etc. A
large number of chemical products thus originate from crude oil. At the center of
the petroleum SC lies the oil refining operations. Crude is transported to refineries
around the world mostly by large ships called Very Large Crude Carriers (VLCCs) or
pipelines. Transportation times are relatively long despite extensive networks and
carefully planned schedules; typically it takes 4–6 weeks for a VLCC carrying crude
oil from the Middle East to reach refineries in Asia. The refining industry is highly
capital-intensive, and mostly high-volume and low-margin, where profitability de-
pends on efficient operations and the best mixture of products. SCM plays a key
role in achieving this. SCM in a refinery involves activities, such as crude procure-
ment, oil trading, logistics management, production planning, scheduling, etc.
Due to the complexity of refinery operations, the literature on refinery scheduling
has addressed this in the form of three smaller sub-problems:
1) crude oil operations from unloading up to charging into a
crude distillation unit (CDU),

Electricity

Jet Fuel

Refining

Oil & Gas Transportation

Production Petrol & Diesel

Storage & Transportation

Petrochemicals

Fig. 4.1 The petroleum refinery supply chain.

92 4 Robust Supply-Chain Operations through Rescheduling

2) blending of intermediate products from the CDU into

finished products, and
3) lifting or delivery of the finished products.

Kelly and Mann [2] report that crude oil costs account for 80% of refinery turnover
and efficient scheduling could make a difference of millions of dollars every year.
Consequently, most of the literature in refinery scheduling is on the crude oil oper-
ations sub-problem. An optimal crude operation schedule is predetermined based
on crude arrival data, production targets, and operational constraints. The typical
scheduling horizon in a refinery is 1–2 weeks.
The crude oil scheduling problem is typically stated as a mixed integer nonlin-
ear programming (MINLP) problem. At the present time, no algorithms can find
the global optimal solution of an MINLP problem in a tractable run-time. Most
research [3–6] approximates the nonlinearity to simplify the problem to a mixed
integer linear programming (MILP) problem. However, it still takes a significant
amount of time to solve the MILP problem and generate an optimal schedule for a
real-world refinery. Disruptions add a further challenge to this.
It is not uncommon for unexpected events to occur and disrupt the schedule. On
the contrary, the long crude supply chain is highly susceptible to disruptions, and
4–5 occurrences every month is common [7]. These lead to crude stock-out and
failure to meet customer demands. Other disruptions arise due to the delivered
crude oil being out-of-spec, unexpected changes in product yields, unavailable or
constrained plant units, demand fluctuations, etc. In many cases, disruptions cause
the current schedule to become infeasible; for example, a ship arrival delay could
lead to an out-of-crude situation. Efficient rescheduling is therefore necessary to
ensure robust SC operations.
Total rescheduling, i.e. running the whole scheduling optimization again every
time a disruption occurs is not practical. The time required is a critical issue since
rescheduling has to be performed in the actual course of the operation execution
and any delay in responding to the disruptions could have significant financial im-
pact. Further, a small change in data could lead to large changes from the original
schedule since the MILP solver in general would pursue a very different path. Such
large changes are undesirable during operation. Rescheduling should take into ac-
count the initial schedule to minimize such changes.

4.3
Rescheduling Problem Statement

Rescheduling is an integral part of disruption management to achieve robust SC

operations. The rescheduling work presented here is based on a discrete-time for-
mulation of crude oil scheduling [4] in the refinery SC, which considered the un-
loading of crude oil from crude carrier vessels up to the charging of crude oil to
crude distillation units (CDUs). Figure 4.2 shows the configuration of a typical
marine-access refinery, which includes crude offloading facilities, such as a Single
 4.3 Rescheduling Problem Statement 93

Multi-parcel VLCCs
C
VLCC-2 VLCC-1 T-101
SBM
D
U
J-1 T-102 1
V-1

T-103
C
J-2
Single-parcel vessels

V-2
D
U
J-3 T-104 2
V-3

T-105 C
J-4
V-4
D
J jetties
U
T-106
3
V vessels I tanks U CDUs
Fig. 4.2 Schematic of refinery configuration in a typical marine-access refinery.

Buoy Mooring (SBM) station and/or one or more jetties, storage facilities, such
as storage tanks and/or charging tanks, and processing facilities, such as CDUs.
The SC operation involves unloading crudes into multiple storage tanks from the
ships/tankers arriving at various times and feeding the CDUs from these tanks
at various rates. Crudes arrive in either large multiparcel tankers or small single-
parcel vessels. A VLCC has multiple compartments to carry several large parcels of
different crudes. However, due to its huge size, a VLCC must dock offshore at an
SBM, which connects to the crude tanks in the refinery via one SBM pipeline. How-
ever, from time to time, a refinery may also receive small parcels of single crudes
via small ships that dock at an onshore jetty. A refinery may have such multiple jet-
ties. When there are multiple jetties, multiple ships can dock at the same time and
transfer crude parcels simultaneously. The operating rules and the assumptions
regarding the refinery operations are summarized in Table 4.1.
We assume that the refinery operations have been a priori scheduled but have
been disrupted. The proposed method is effect-based and not root-cause-based,
as ultimately any source of disruption will impact the refinery by disrupting the
crude unloading operations, the CDU charging operations, or both. We consider
five types of refinery SC disruptions: (1) delays in crude arrival, (2) unavailability of
offloading facilities (SBM or jetty), (3) unavailability of storage tanks, (4) unavail-
ability of CDUs, and (5) changes in demand. These disruptions are either detected
94 4 Robust Supply-Chain Operations through Rescheduling
Table 4.1 Operating rules and assumptions regarding refinery operations.

Operating rules
1 A tank receiving crude from another tank, a ship, or a tanker cannot feed a CDU at the same
time.
2 Each tank needs some time (8 h) for brine settling and removal after receiving crude.
3 Multiple tanks can feed a single CDU. Most refiners allow at most two tanks to feed a CDU,
as the operating complexity increases and controllability becomes a problem for more than
two tanks.
4 A tank may feed multiple CDUs. Again, a tank normally does not feed more than two CDUs.

Assumptions made regarding refinery operations

1 Only one VLCC can unload at any moment. This is reasonable, as there is only one SBM.
2 The sequence in which a VLCC unloads its parcels is known a priori.
3 A parcel can unload to only one storage tank at any moment.
4 The SBM pipeline holds only one type of crude at any time and crude flow is plug flow. This
is valid, as parcel volumes in a VLCC are much larger than the SBM pipeline holdup.
5 Crude mixing is perfect in each tank and time to changeover tanks between processing units
is negligible.
6 For simplicity, only one key component decides the quality of a crude feed to CDU.

by an automated disruption management system [8–10] or by operations person-

nel. Other types of disruptions can be handled by observing their effects on the
operations.
The rescheduling problem is then stated as follows. Given:
1) Refinery specifications: configuration details, information
about modes of crude segregation in storage and processing,
quantity limits (flow rates from the SBM station and jetties to
tanks and from tanks to CDUs, CDU processing rates,
holdup in the SBM pipeline, storage tank capacities), quality
limits (key component concentration limits during storage
and processing).
2) Production demands during the scheduling horizon.
3) Initial refinery state (initial crude type in the SBM pipeline,
initial inventory levels, and initial volume fractions of crudes
in each tank).
4) Original operation schedule.
5) Disruption specifications (detection time, disrupted object,
disruption duration, etc.).
6) Economic data, such as sea waiting (demurrage) cost, safety
stock penalty, and crude changeover cost.

Within an acceptable solution time, generate one or more new schedules that
are:
1) feasible after accommodating the disruption, and
 4.3 Rescheduling Problem Statement 95

2) efficient, i.e., having profit close to the initial schedule

(within 5–10% is reasonable).

In the crude oil scheduling formulation considered here, profit is calculated as

the sum of margins (netbacks) of crude throughput minus the operating costs re-
lated to changeover, demurrage or sea-waiting cost, and safety stock penalty.

4.3.1
Literature Review

Although the literature on refinery rescheduling is very sparse, batch-plant (flow

shop and job shop) rescheduling has drawn much attention in the literature [11].
In general, the shop-scheduling problem involves assigning jobs, consisting of a
number of operations, for processing in a set of machines in order to produce cer-
tain products, and meet orders with different priorities under some technological
precedence (sequence) and resource constraints. Disruptions to shop operations
could come from internal or external sources. Examples of internal disruptions
are machine breakdown, process time variation, and manpower unavailability. Ex-
ternal disruptions include unavailability of raw material, arrival of urgent jobs, and
cancellation of orders. In the literature, several different approaches have been pro-
posed to address the shop-rescheduling problem.
Since the dynamic shop-scheduling problem is nondeterministic polynomial
time (NP)-complete, heuristic methods have been proposed to solve this problem.
Kunnathur et al. [12] develop a simulation-based expert system, which incorporates
a rescheduling heuristic to reschedule operations when there is any variation from
the expected value of flowtime. Akturk and Gorgulu [13] present a rescheduling
strategy that reschedules part of the initial schedule in the case of machine break-
down. The affected operations rescheduling (AOR) heuristic of Abumaizar and
Svestka [14] propagate disruptions and reschedule operations affected by the dis-
ruptions. Building on this approach, Subramaniam and Raheja [15] present a more
robust heuristic in modified AOR (mAOR). Henseler [16] proposes a reschedul-
ing algorithm guided by rules and heuristics, which efficiently repairs broken con-
straints by iteratively revising the schedule until there are no more violated con-
straints.
Exact MILP-based methods have also been proposed for schedule revision in
case of disruptions. Mendez and Cerdá [17] develop a MILP-reactive scheduling al-
gorithm to revise the short-term schedule of resource-constrained multistage batch
facilities in response to unexpected disruptions. The size of the problem formula-
tion remains reasonable because a large part of the scheduling decisions are un-
changed and rescheduling actions are applied gradually. Roslöf et al. [18] use an
MILP-based algorithm to efficiently improve an existing feasible, but nonoptimal,
production schedule or to reschedule jobs in the case of changed operational para-
meters. Vin and Ierapetritou [19] formulate a MILP model for reactive scheduling
in multiproduct batch plants, where a two-stage solution procedure is used. The
96 4 Robust Supply-Chain Operations through Rescheduling

first stage is deterministic and the second stage incorporates all different reschedul-
ing alternatives based on the previous stage.
On the other hand, Rangsaritratsamee et al. [20] propose a scheduling methodol-
ogy based on a stochastic approach. Their methodology simultaneously considers
both efficiency and stability measures of the schedule. Scheduling is performed
at fixed rescheduling points using a multiobjective genetic algorithm. Belz and
Mertens [21] couple expert systems and simulation in a prototype decision support
system for short-term rescheduling in case of production disturbances. Li et al. [22]
integrate simulation, artificial neural network, expert knowledge, and dispatching
rules in their production rescheduling expert simulation system. Their system has
been implemented in a Chinese manufacturing firm.
Other shop rescheduling approaches, which have been proposed include case-
based reasoning [23, 24], constraint-based [25, 26], fuzzy logic [27, 28], and neural
network [29–31]. The review of these methods, their advantages and disadvantages
is summarized in Subramaniam and Raheja [15]. In addition, the reader is referred
to Aytug et al. [32] and Herroelen and Leus [33] for a more detailed review of the
vast literature on production scheduling under uncertainties, robust scheduling,
and reactive scheduling procedures.
However, these shop-rescheduling methods cannot be directly applied to con-
tinuous systems like a refinery due to inherent differences. A batch process runs
per order, whereas a continuous process runs continuously. The decisions to be
made in batch scheduling involve allocation of resources, whereas in the contin-
uous case additionally throughputs have to be specified. The typical objectives in
batch scheduling are minimum makespan or minimum due date penalties as op-
posed to maximum profit or minimum cost in continuous processes. As a result,
refinery rescheduling requires a different method from that of batch plants case.
Two methods for crude oil rescheduling are discussed in the next section.

4.4
Rescheduling Methodologies

Filling the gap in refinery rescheduling amidst the plentitude of work in the batch
counterpart, we have proposed two crude oil rescheduling methods for robust SC
operations. The first is heuristic-based [34], while the second is model-based [35].
The motivation, framework, procedures, and features of both methods are dis-
cussed below.

4.4.1
Heuristic-Based Rescheduling

As previously discussed, total rescheduling by solving the complete MINLP (or

approximate MILP) would take a significantly long computational time; a lux-
ury which the refinery can hardly afford during actual operation. In addition, to-
tal rescheduling without considering the original schedule would result in a new
 4.4 Rescheduling Methodologies 97

Original Data

Plant Hardware
Model
• Operations (Inventories,
Original Schedule throughput, transfer rates, etc.)

Schedule Heuristic
Evaluator Rescheduler

Disruptions
/ Changes New Schedule(s)

Fig. 4.3 Heuristic-based rescheduling framework.

schedule most likely to be very different from the original. These two limitations
are the key motivation for the heuristic-based method.
The framework of the heuristic-based rescheduling method is schematically
shown in Fig. 4.3. Inputs to the rescheduler are the original data – consisting of
the original schedule and the initial refinery state – and the disruption details. The
original schedule contains the crude (ship) arrival data, the production targets, and
the schedule of the two kinds of operation – parcel unloading from ship to tank
and crude charging from tank to CDU. The plant hardware model represents the
evolving refinery state (tank inventory, CDU throughput, unit availabilities) over
the scheduling horizon. The disruption details include the disruption source, mag-
nitude, duration, and detection time. The heuristic rescheduler first determines the
effect of the disruption on the original schedule and detects any infeasibility. New
schedules are then generated based on heuristics that take into account operational
and capacity constraints from the plant hardware model. The schedule evaluator cal-
culates the profit of each new schedule generated by the rescheduler.
The key concept in the heuristic-based rescheduling method is operation block.
In general, the crude oil schedule in the refinery specifies the rates and timings
of two kinds of operations: parcel unloading and CDU charging. The schedule im-
plicitly establishes a configuration – parcel-to-tank connections and tank-to-CDU
connections – in the refinery at every period over the scheduling horizon. Based
on this, a schedule can be decomposed into blocks of operation. One or more op-
erations spanning one or more periods is considered to be a block if it involves no
intervening change in configuration.
98 4 Robust Supply-Chain Operations through Rescheduling

The heuristic-based rescheduling method is based on the blocks in the original

schedule. A disruption may disrupt one or more operation block(s). A block is mod-
ified only when it is affected by a disruption. For example, a ship delay may disrupt
a parcel unloading block. Disrupted blocks are removed from the schedule to a tem-
porary space called the D-space. Rescheduling involves inserting variations of these
blocks to form a new schedule so as to correct the violations arising from the dis-
ruptions. Heuristics are used to identify disrupted blocks and to generate variations
that overcome the effects of the disruption. Since there are two kinds of operations
(parcel unloading and CDU charging) and both can be removed or rescheduled,
four heuristics are needed: (1) removing disrupted CDU charging operations, (2)
removing disrupted parcel unloading operations, (3) rescheduling disrupted CDU
charging operations, and (4) rescheduling disrupted parcel unloading operations.
As an example, Fig. 4.4 shows the heuristic for rescheduling disrupted CDU
charging operations. The main idea in rescheduling a CDU charging operation is
to (1) rectify any possible minimum throughput violation caused by the removal
of the originally scheduled operation, and (2) maximize profit. This is performed
by identifying all eligible tanks that can charge the CDU and calculating the charg-
ing volume based on the aforementioned criteria. Overall, there are two distinct
procedures involved in rescheduling a CDU charging operation: reconfiguring,
which involves changes in configuration, and updating volume, which involves
only changes to transfer rates. Reconfiguring could result in new CDU charging
block(s) whereas updating volume would not. The heuristic for reconfiguring CDU
charging operations may result in one or more proposed schedules as correspond-
ing to a different configuration. If no configuration changes are necessary, only
the charging rates can be varied following the heuristic for updating volumes of
CDU charging blocks. The reader is referred to Adhitya et al. [34] for details of the
heuristics.
The heuristics are designed such that the blocks in the original scheduled are pre-
served to the extent possible, i.e., configuration changes are minimized, while gen-
erating new schedules that are feasible in the face of disruption. Since the method
is rule-based, it requires much less time to generate new (near optimal) schedules
compared to total MILP rescheduling. Furthermore, a number of feasible sched-
ules can be identified rather than just one.
However, the heuristic-based method is not without limitations. Heuristics by
nature lack generality and only work for the types of disruptions considered. It
could take a considerable effort to extend or modify them. They may not be able to
accommodate unforeseen disruption types. Furthermore, in some cases, heuristics
may fail to identify feasible solutions if they lie outside the heuristics’ search space.
The model-based rescheduling method seeks to overcome the above shortcomings.

4.4.2
Model-Based Rescheduling

The model-based rescheduling approach covers all types of disruptions (general)

and considers the whole search space (complete) to identify the solution. The
 4.4 Rescheduling Methodologies 99

ASSIGNMENT
CC CDU charging operation in the D-space to be rescheduled
U The CDU CC is scheduled for
[t1, t2] The periods CC is scheduled for
[Qmin, Qmax] Throughput limits of U
V Total transfer volume of CC
tdet Disruption detection time
thor End of scheduling horizon
Q(u, t) Total charging volume to CDU u at the period [t-1, t]
V1 Charging volume required to rectify min. throughput violation
V2 Remaining volume after rectifying min. throughput violation

Check for minimum throughput violation

Is Q(U, t) Q min? t, t [ t1 , t 2 ]

No Yes

More charging volume to U is V can be rescheduled to

needed in the period [t1, t2] any period in [tdet, thor]

Calculate V 1
V1 (t 2 t1 )(Qmin Q(U , t 2 )) Update volume of CDU charging blocks
following Heuristic 3b for (V, tdet, thor)
Note: Q(U, t) for each period in [t1, t2] are constant
by definition of operation.

Reconfigure for CC following Heuristic 3a Result

Schedule
for (V1, t1, t2)

There may be more than one

proposed schedule from Heuristic 3a
Schedule

Calculate V 2
V2 V V1

For each No more schedule left

proposed End of heuristic
schedule

Update volume of CDU charging blocks Result

Schedule
following Heuristic 3b for (V2, tdet, thor)

Next schedule

Fig. 4.4 Flowchart of the heuristic for rescheduling disrupted CDU charging operations.
100 4 Robust Supply-Chain Operations through Rescheduling

Refinery Module
Configuration Library
CO-Graph

Original Schedule

Scope-Graph
Disruptions

Graph Propagation of Graph Search for Schedule

Disruption Consequences Rectification Options Evaluator

New Schedule

Fig. 4.5 Model-based rescheduling framework.

model-based rescheduling framework is shown in Fig. 4.5. It uses two graphs to

represent the operations: (1) the composite-operations graph (CO-graph), which rep-
resents all possible material-flows among SC entities (Fig. 4.6) and (2) the scheduled-
operations graph (Scope-graph), which represents scheduled operations (Fig. 4.7).
The latter is a subgraph of the former. For refinery crude oil operations, the mate-
rial is crude oil and the entities are parcels, offloading facilities (SBM and jetties),
tanks, and CDUs. Both graphs consist of nodes, which represent variables, and
arcs, which represent the cause-and-effect relationships between them.
There are two types of node: (1) state node, which represents the state of an entity,
in this case the volume of crude remaining in a parcel, the volume of crude in a
tank, and the volume of crude throughput into the CDU and (2) flow node, which
represents material flow, in this case the flows from crude vessels (parcels) to tanks
via the offloading facilities and from tanks to CDUs. There are two types of flow
node: (2a) transfer node, which represents the physical transfer of material from one
entity to another and (2b) total-flow node, which represents the cumulative transfer
into (out of) an entity from (to) all connected entities.
There are three types of arcs: (1) undirected arc, which represents the equal-
ity relationship between a transfer node and a total-flow node. (2) Directed arc,
which represents how a total-flow node impacts a state node. The impact can be
positive/additive (as is the case between total-flow-in and a state node) or nega-
tive/subtractive (as between total-flow-out and a state node). The former is shown
as a solid directed arc and the latter a broken directed arc. (3) Constraint arc, which
imposes a mass balance or operational constraint between any two nodes.
 Parcels module Tank-farm module CDUs module

Parcel-1 module 3
1 P1-jetty-T2

P1-jetty-T3 T2 total flow-in T2 total flow-out

P1 volume P1 total flow-out 1 T2-U1
P1-jetty-T4
4 T2 volume
3 CDU-1 module
P1-jetty-T5
T3-U1
T3 total flow-in 1 T3 total flow-out
1 4
P2-jetty-T1 T3 volume U1 total flow-in U1 fulfillment
Tank-4 module
P2-jetty-T6 T4-U1
P2 volume P2 total flow-out 4 1
P2-jetty-T7
T4 volume
P2-jetty-T8 T4 total flow-in T4 total flow-out T5-U1
3

Fig. 4.6 CO-graph for the case study.

1 P3-jetty-T2 4 1

P3-jetty-T3 T5 volume T2-U2

P3 volume P3 total flow-out
P3-jetty-T4 T5 total flow-in T5 total flow-out
3
P3-jetty-T5 T3-U2
4 1
U2 total flow-in U2 fulfillment
1 P4-jetty-T1 T1 volume
T4-U2
3
P4-jetty-T6 T1 total flow-in T1 total flow-out
P4 volume P4 total flow-out
P4-jetty-T7
4 1 T5-U2
P4-jetty-T8
T6 volume
3
1 P5-jetty-T2 T6 total flow-in T6 total flow-out

P5-jetty-T3 T1-U3
P5 volume P5 total flow-out 4 1
P5-jetty-T4
T7 volume
P5-jetty-T5 T6-U3
3
T7 total flow-in T7 total flow-out
U3 total flow-in U3 fulfillment
1 P6-jetty-T1
4 1 T7-U3
P6-jetty-T6
P6 volume P6 total flow-out T8 volume
P6-jetty-T7 3
T8 total flow-in T8 total flow-out T8-U3
4.4 Rescheduling Methodologies

P6-jetty-T8
101

Legend Flow node Positive effect Constraint arc

State node Inverse effect Equality relationship

 Parcels module Tank-farm module CDUs module

Parcel-1 module 3
1 P1-jetty-T2

P1-jetty-T3 T2 total flow-in T2 total flow-out

102

P1 volume P1 total flow-out 1 T2-U1

P1-jetty-T4
4 T2 volume
3 CDU-1 module
P1-jetty-T5
T3-U1
T3 total flow-in 1 T3 total flow-out
1 4
P2-jetty-T1 T3 volume U1 total flow-in U1 fulfillment
Tank-4 module
P2-jetty-T6 T4-U1
P2 volume P2 total flow-out 4 1
P2-jetty-T7
T4 volume
P2-jetty-T8 T4 total flow-in T4 total flow-out T5-U1
3

1 P3-jetty-T2 4 1

P3-jetty-T3 T5 volume T2-U2

P3 volume P3 total flow-out
P3-jetty-T4 T5 total flow-in T5 total flow-out
3
P3-jetty-T5 T3-U2
4 1
U2 total flow-in U2 fulfillment
1 P4-jetty-T1 T1 volume
T4-U2
3
P4-jetty-T6 T1 total flow-in T1 total flow-out
P4 volume P4 total flow-out
P4-jetty-T7
4 1 T5-U2
P4-jetty-T8
T6 volume
4 Robust Supply-Chain Operations through Rescheduling

3
1 P5-jetty-T2 T6 total flow-in T6 total flow-out

P5-jetty-T3 T1-U3
P5 volume P5 total flow-out 4 1
P5-jetty-T4

Fig. 4.7 Scope-graph for the original schedule in the case study.
T7 volume
P5-jetty-T5 T6-U3
3
T7 total flow-in T7 total flow-out
U3 total flow-in U3 fulfillment
1 P6-jetty-T1
4 1 T7-U3
P6-jetty-T6
P6 volume P6 total flow-out T8 volume
P6-jetty-T7 3
T8 total flow-in T8 total flow-out T8-U3
P6-jetty-T8

LegendL Flow node Positive effect Constraint arc

State node Inverse effect Equality relationship

the tank module consists of a volume node, a total-flow-in node, and a total-flow-
the parcel module consists of a parcel volume node and a parcel total-flow-out node;
a state node, a total-flow-in node and/or a total-flow-out node. As shown in Fig. 4.6,
grouped into modules based on the SC entities they belong to. A module consists of
To recognize the modularity of the SC entities, the nodes in the CO-graph are
 4.4 Rescheduling Methodologies 103

Table 4.2 Value of a typical tank-volume node in CO-graph.

Time / Period
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
c1 0 0 0 0
Crude Volume

c2 0 0 0 0
(kbbl)

c3 0 0 0 0
c4 100 89 77 66 54 43 37 31 29 27 25 23 21 19 16 14
c5 150 133 116 99 81 64 56 47 44 41 38 34 31 28 25 21
c6 100 89 77 66 54 43 37 31 29 27 25 23 21 19 16 14
350 310 270 230 190 150 130 110 103 95 88 80 73 65 58 50
Max limit
(kbbl) 400 400 400 400 400 400 400 400 400 400 400 400 400 400 400 400

Min limit
(kbbl) 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

out node; the CDU module consists of a fulfillment node and a total-flow-in node.
Modules of the same type are grouped together to derive collections – a parcels
module, a tank-farm module, and a CDUs module, which are the superset of all parcel
modules, all tank modules, and all CDU modules, respectively. Hence, the CO-
graph can be viewed as three superset modules in series interfaced by transfer
nodes. The CO-graph can be constructed for any SC by using standard modules
from the module library and connecting them according to the SC configuration
and practices (Fig. 4.5).
Each node in the CO-graph has an associated two-dimensional matrix, with time
on the x-axis and crude type on the y-axis (Table 4.2). Discrete-time representation
is used; values of variables are observed at fixed time points or periods within the
scheduling horizon. For state variable nodes, the time index is time point, whereas
for transfer nodes, the time index is period. Period 1 spans Time 0 to Time 1. The
fact that multiple crudes (e.g., Brent, Dubai, Kuwait, etc.) make up the material
flows and states of each entity is accounted for in the y-axis. For ease of analy-
sis, related parameters from the refinery configuration, such as total volume and
maximum/minimum allowable limits can also be included in the y-axis. Hence,
matrix calculations for each node in the CO-graph can be performed based on the
relationship specified by the arcs.
The original schedule dictates the transfer flows from parcels to tanks and from
tanks to CDUs, i.e., the transfer nodes in the CO-graph. The schedule can then be
simulated by propagation in the CO-graph starting from the transfer nodes to all
other nodes. Simulation is executed starting from the upstream module moving to
the downstream ones for all periods in the scheduling horizon. Once the simula-
tion is completed, the values of all the nodes in the CO-graph have been updated
according to the original schedule. This forms a subset of the CO-graph called the
Scope-graph. Only scheduled material flows are represented in the Scope-graph
(Fig. 4.7). As shown in Fig. 4.5, the Scope-graph is derived from the CO-graph fol-
lowing the original schedule. The Scope-graph has all the information needed to
calculate the profit of the schedule: parcel unloading data (from parcel total-flow-
out nodes) for demurrage calculation, tank inventories (from tank volume nodes)
104 4 Robust Supply-Chain Operations through Rescheduling

for safety stock penalty calculation, CDU charging schedule (from tank-to-CDU
transfer nodes) for CDU changeover cost calculation, and CDU throughput (from
CDU total-flow-in nodes) for netbacks calculation.
The disruptions are then imposed on the Scope-graph and their effects de-
termined by graph propagation. The result could be violations of transfer flow
limits, CDU-flow-in limits, incomplete unloading of parcel, and under- or over-
production. The disrupted schedule becomes the base-case for rescheduling. A
greedy best-first graph search strategy [36] is used to generate new schedules, which
rectify the violations caused by the disruptions. The schedule evaluator calculates
the utility values of the schedule, such as profit, number of changes from the origi-
nal schedule, and other user-specified utility functions. The number of rescheduled
operations or changes from the original schedule is defined as the number of op-
eration blocks that were in the initial schedule but not in the new schedule, that is,
changes in configuration (parcel to tank and tank to CDU connections). The same
configuration with different start time or duration is also counted as a rescheduled
operation, while an identical configuration with a difference in transfer rates is not.
Rectification of violations to generate new feasible schedules is performed
through rectifications-graph. It charts the path from the base-case schedule to new
schedules through modification of variables to rectify violations. The rectifications-
graph consists of scheds, which is a set of the complete values of all the decision
variables in the schedule, i.e., parcel-to-tank transfers and tank-to-CDU transfers.
Each sched also has a set of descriptions, including utility values (e.g., profit, num-
ber of changes, etc.), number of violations, depth-level in the rectifications-graph,
and a pointer to the parent sched. Children sched are derived from a parent sched
by changing the decision variables (i.e., the transfer nodes) related to an operation
to rectify a particular violation. The amount of modification is linear to the violated
quantity because all the nodes in the model are linearly related. Discretization of
the continuous decision variables allows the enumeration of different values of the
variables and makes the search domain finite. The objective is to find scheds with
zero violation, in other words feasible schedules.
A parent sched will have multiple children scheds since there are many possible
modification options of the transfer variables to rectify for a particular violation.
For example, in Fig. 4.7, any one or a combination of T2-U1, T3-U1, T4-U1, and
T5-U1 transfer nodes can be modified to rectify for CDU 1 under-production viola-
tion. The scheds are sorted to determine which sched to be expanded first to rectify
for the next violation. Sorting is based on the utility function specified by the user,
which could be profit, number of changes from the original schedule, number of
violations, depth-level, or combinations of these. The rectifications-graph is contin-
ually expanded until a termination criterion has been met. A variety of termination
criteria in addition to zero violation can be used, such as prespecified search time,
profit decrease below a certain limit, number of feasible schedules identified, etc.
Since the rectifications-graph provides all possible rectification options and any
possible schedule can be generated by manipulation of the transfer nodes, the
model-based rescheduling method is general and complete. It also provides the
flexibility of employing various search strategies, including the heuristic-based
 4.5 Case Study 105

method in Section 4.4.1. The user can define his preferred utility function and
termination criteria.

4.5
Case Study

In this section, a case study involving two different disruptions – a parcel disruption
and a tank disruption – is described. This case study considers a refinery with the
configuration shown in Fig. 4.2. It has 8 tanks, 3 CDUs, and is initially scheduled
for 15 periods. Crude arrives in six single-parcel vessels. The parcels are unloaded
over the horizon through the jetty. Total production demand for the CDUs over
the horizon is 375, 400, and 400 kbbl, respectively. Table 4.3 shows the complete
data used in this case study and Table 4.4 shows the initial schedule. In Table 4.4,
a parcel unloading operation is depicted by the name of the parcel source and a
positive volume as it increases the crude volume in a tank, whereas a CDU charging
operation is depicted by the name of the destination CDU and a negative amount
as it decreases the crude volume in a tank. For example, Tank 3 is charging 10 kbbl
to CDU 1 at Periods 1–15 and Tank 2 is receiving 100 kbbl from Parcel 1 at Period
3. For ease of reference, parcel unloading operations and CDU charging operations
to different CDUs are shaded differently. The CO-graph and Scope-graph for this
case study are shown in Figs. 4.6 and 4.7, respectively. Rescheduling results from
the heuristic-based method, model-based method, and total MILP rescheduling [4]
are reported and compared for two examples.

Example 1. The refinery is informed at Time 1 that because of adverse weather at

sea, the ship carrying Parcel 1 is delayed by 24 h (three periods) and will arrive only
at Time 5. Parcel 1 is initially scheduled to be unloaded to Tank 2. However, at the
new arrival time, Tank 2 is scheduled to charge CDU 2. If Parcel 1 is not unloaded,
there will be a crude stock-out in Tank 2. Hence, this crude arrival disruption re-
quires rescheduling of operations.

Rescheduling was done using three different methods: heuristic-based, model-

based, and total MILP rescheduling following Reddy et al. [4]. The heuristic method
identifies one schedule, which is shown in Table 4.5. In the model-based method,
the utility function used is a combination of number of changes and profit, with
the former taking the higher priority. The discretization used is 100 kbbl for parcel
unloading operations and 5 kbbl for CDU charging operations. The termination
criterion used is zero violation and profit decrease of at most 20 from the original
schedule. The first schedule that satisfied this criterion is shown in Table 4.6. Total
MILP rescheduling with zero gap fails to give any schedule after 48 h. Hence, 1%
gap is used and the resulting schedule is shown in Table 4.7.

Example 2. Consider the case where heavy overnight rainfall causes water entrain-
ment in Tank 6, detected at Time 9. As a result, Tank 6 needs to undergo a 16-h
Table 4.3 Problem data for the case study.
106

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6

2 250 100 1 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1

4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 6 -35 3 -35 3 -35 3 -35 3 -40 3 -40 3

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

4 Robust Supply-Chain Operations through Rescheduling

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 285 285 285 285 285 285
2 250 250 250 350 350 325 300 275 250 225 200 175 150 125 100 75
3 300 290 280 270 260 250 240 230 220 210 200 190 180 170 160 150
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 355 320 285 250 210 170
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 70 70 70 70 70 70

CDU Throughput
CDU Demand Min Max Total
/period /period Throughput
1 375 20 50 375
2 400 20 50 400
3 400 20 50 400
Table 4.4 Initial schedule for the case study (profit = 1817).

Flow Rate Limits (kbbl / period) (Max Vol) Initial Inventory Initial Crude Composition (kbbl)
Tank
Min Max Capacity (kbbl) (kbbl) C1 C2 C3 C4 C5 C6
Parcel-Tank 10 250 T1 400 250 100 100 50 0 0 0
Tank-CDU 0 50 T2 400 250 0 0 0 50 100 100
CDU throughput 20 50 T3 400 300 0 0 0 100 100 100
T4 400 350 0 0 0 100 150 100
T5 400 250 0 0 0 100 75 75
Demurrage Cost (k$/period) 15 T6 400 100 25 25 50 0 0 0
Changeover Loss (k$/instance) 5 T7 400 100 50 25 25 0 0 0
Safety Stock Penalty ($/bbl/period) 0.2 T8 400 250 75 75 100 0 0 0

Crude Key Component Margin ($/bbl) Parcel Arrival Time Crude Type, Amount (kbbl)
C1 0.005 1.5 P1 2 C6, 100
C2 0.008 1.75 P2 3 C1, 100
C3 0.004 1.85 P3 3 C4, 90
C4 0.015 1.25 P4 4 C2, 125
C5 0.01 1.45 P5 4 C5, 125
C6 0.02 1.65 P6 6 C3, 100
CDU Key Comp. Limits Demand (kbbl)
CDU1 0.0125-0.0185 375
(Min-Max)
CDU2 0.0125-0.0175 400
(Min-Max)
CDU3 0.0040-0.0070 400
(Min-Max)
4.5 Case Study
107
Table 4.5 Proposed schedule for Example 1 by heuristic-based method (profit = 1807).
108

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6

2 250 100 1 -26.9 2 -26.9 2 -26.9 2 -26.9 2 -26.9 2 -26.9 2 -26.9 2 -26.9 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1
-20 2 -20 2 -20 2
4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 6 -35 3 -35 3 -35 3 -35 3 -40 3 -40 3

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 Robust Supply-Chain Operations through Rescheduling

Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 285 285 285 285 285 285
2 250 250 250 250 250 250 350 350 323.125 296.25 269.375 242.5 215.625 188.75 161.875 135
3 300 290 280 270 260 230 200 170 160 150 140 130 120 110 100 90
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 355 320 285 250 210 170
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 70 70 70 70 70 70

CDU Throughput
CDU Demand Min Max Total
/period /period Throughput
1 375 20 50 375
2 400 20 50 400
3 400 20 50 400
Table 4.6 Proposed schedule for Example 1 by model-based method (Profit = 1804).

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6

2 250 -25 2 100 1 -25 2 -25 2 -25 2 -25 2 -25 2 -20 2 -30 2 -35 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1
-20 2 -20 2
4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 6 -35 3 -35 3 -35 3 -35 3 -40 3 -40 3

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 285 285 285 285 285 285
2 250 250 250 250 250 225 325 325 300 275 250 225 200 180 150 115
3 300 290 280 270 260 250 220 190 180 170 160 150 140 130 120 110
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 355 320 285 250 210 170
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 70 70 70 70 70 70

CDU Throughput
CDU Demand Min Max Total
4.5 Case Study

/period /period Throughput

1 375 20 50 375
109

2 400 20 50 400
3 400 20 50 400
Table 4.7 Proposed schedule for Example 1 by total rescheduling (Profit = 1815).
110

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250

2 250 90 3 -20 2 -20 2 -45 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1

4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -20 2 -20 2 -20 2 -20 2 125 5 100 1 -50 2 -50 2 -35 2 -20 2 -20 2 -20 2 -20 2

6 100 125 4 100 6 -34.3 3 -34.3 3 -34.3 3 -34.3 3 -34.3 3 -34.3 3 -34.3 3

7 100 100 2

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 Robust Supply-Chain Operations through Rescheduling

Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 250 250 250 250 250 250 250 250
2 250 250 250 250 340 340 320 300 255 255 255 255 255 255 255 255
3 300 290 280 270 260 250 240 230 220 210 200 190 180 170 160 150
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 210 190 170 150 275 375 375 325 275 240 220 200 180 160
6 100 100 100 100 100 100 225 325 325 290.714 256.428 222.142 187.856 153.57 119.284 85
7 100 100 100 100 200 200 200 200 200 200 200 200 200 200 200 200
8 250 230 210 190 170 150 130 110 90 90 90 90 90 90 90 90

CDU Throughput
CDU Demand Min Max Total
/period /period Throughput
1 375 20 50 375
2 400 20 50 400
3 400 20 50 400
 4.5 Case Study 111

dewatering at Periods 11 and 12. If no action is taken, CDU 3 will be down dur-
ing those periods, as Tank 6 cannot operate while being dewatered. Consequently,
70 kbbl of CDU 3 throughput would be lost and shutdown/startup costs incurred.
Therefore, rescheduling is critical to avoid unnecessary losses.

The heuristic method proposes one schedule in this tank disruption example,
shown in Table 4.8. In the model-based method, the graph search explores modifi-
cation only of nodes that are directly related to the violations. The utility function
used is profit. The termination criterion is complete exploration of modification of
nodes directly related to the violations. The method identifies 1983 feasible sched-
ules; the one with the highest profit is shown in Table 4.9. Total MILP rescheduling
with zero gap gives the schedule in Table 4.10.
The number of feasible schedules identified, profit decrease, number of resched-
uled operations, and computational time of the three methods for both exam-
ples are compared in Table 4.11. In terms of number of proposed schedules, total
rescheduling would only give one schedule. Heuristic-based method can propose
multiple schedules, but in these two examples only one schedule is identified for
each because the others are infeasible. The model-based method gives one sched-
ule in Example 1 because of the termination criterion used, stopping after the first
feasible schedule with a profit decrease of at most 20 is identified. In Example 2,
multiple feasible schedules are identified due to the search strategy used.
As expected, total rescheduling gives the highest profit for both examples. The
other two methods are reasonably good as the differences are within 1% of the total
profit. The heuristic-based method gives higher profit than the model-based one as
the latter only modifies nodes that are directly related to nodes with violations.
Unlike the former, it pays no consideration to minimizing changeover. Hence, a
different search strategy with such consideration would result in a better profit.
Total rescheduling has the highest number of rescheduled operations for Exam-
ple 1 because the heuristic-based method by design attempts to minimize number
of changes and the model-based method accounts for number of changes in its
utility function. In Example 2, the disruption is detected at Time 9, which means
that the search space becomes significantly smaller than that of the full horizon. As
a result, total rescheduling has only one change from the original schedule, same
as the heuristic-based and model-based methods.
Heuristic-based method is the most efficient of the three in terms of computa-
tional time, taking a mere 1 s in both the examples. It should be noted that total
rescheduling with zero gap fails to identify a feasible schedule after 48 h in Ex-
ample 1. Due to the much smaller search space in Example 2, total rescheduling
with zero gap proposes a schedule in just 10 s. The computation time taken by
the model-based method is large but can be improved by proper adjustment of the
search strategy employed.
As a summary, a comparison of the solution approach, advantages, disadvan-
tages, and suitable situations for the three methods is given in Table 4.12. In sit-
uations where time is scarce, the heuristic-based method would serve well. The
model-based method is called upon when the heuristic-based method fails. Total
Table 4.8 Proposed schedule for Example 2 by heuristic-based method (Profit = 1812).

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
112

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6 -23.3 3 -23.3 3 -23.3 3

2 250 100 1 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1

4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 50 3 -50 3 -50 3

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 Robust Supply-Chain Operations through Rescheduling

Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 261.667 238.334 215 215 215 215
2 250 250 250 350 350 325 300 275 250 225 200 175 150 125 100 75
3 300 290 280 270 260 250 240 230 220 210 200 190 180 170 160 150
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 390 390 390 340 290 240
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 70 70 70 70 70 70

CDU Throughput
CDU Demand Min Max Total
/period /period Throughput
1 375 20 50 375
2 400 20 50 400
3 400 20 50 400
Table 4.9 Proposed schedule for Example 2 by model-based method (Profit = 1808).

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6 -10 3 -10 3

2 250 100 1 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1

4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 6 -35 3 -45 3 -50 3 -50 3

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -10 3 -10 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 285 275 265 265 265 265
2 250 250 250 350 350 325 300 275 250 225 200 175 150 125 100 75
3 300 290 280 270 260 250 240 230 220 210 200 190 180 170 160 150
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 355 355 355 310 260 210
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 70 60 50 50 50 70

CDU Throughput
CDU Demand Min Max Total
4.5 Case Study

/period /period Throughput

1 375 20 50 375
113

2 400 20 50 400
3 400 20 50 400
Table 4.10 Proposed schedule for Example 2 by total rescheduling (Profit = 1816).
114

Parcel Unloading (positive volume) and CDU Charging (negative volume) (u: CDU index, p: parcel index)
Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Vol Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p Vol u/p
1 250 35 6 -40 3 -40 3 -40 3 -40 3 -40 3

2 250 100 1 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2 -25 2

3 300 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1 -10 1

4 350 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1 -15 1

5 250 -20 2 -35 2 -35 2 -35 2 90 3

125 5
6 100 100 2 125 4 65 6

7 100

8 250 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3 -20 3

Tank Volume (min = 50, max = 400)

Tank 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 Robust Supply-Chain Operations through Rescheduling

Vol Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume Volume
1 250 250 250 250 250 250 250 250 285 285 285 245 205 165 125 85
2 250 250 250 350 350 325 300 275 250 225 200 175 150 125 100 75
3 300 290 280 270 260 250 240 230 220 210 200 190 180 170 160 150
4 350 335 320 305 290 275 260 245 230 215 200 185 170 155 140 125
5 250 230 195 160 125 340 340 340 340 340 340 340 340 340 340 340
6 100 100 100 100 200 200 325 325 390 390 390 390 390 390 390 390
7 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
8 250 230 210 190 170 150 130 110 90 70 50 50 50 50 50 50

CDU Throughput
CDU Demand Min Max Total
/period /period Throughput
1 375 20 50 375
2 400 20 50 400
3 400 20 50 400
Table 4.11 Comparison of the three rescheduling methods in Examples 1 and 2.

Method Example 1 Example 2

No. of
Profit No. of CPU No. of
Profit No. of CPU

sch. (s) resch. op. timea sch. resch. op. time
proposed (s) proposed (s)

Heuristic-based 1 −10 2 ∼1 1 −5 1 ∼1

Model-based (first feasible) 1 −22 2 35 1 −16 1 60

Model-based (best) 1 −13 2 ∼3,600 1983 −9 1 ∼21,600

Total reschedulingb (0% gap) Fails to find a solution in 48 h 1 −1 1 10

Total rescheduling (1% gap) 1 −2 9 354 Not applicable

a On Intel Pentium 2.80 GHz, 1 GB RAM running Windows

XP.
b CPLEX 10.0.1 solver within GAMS.
4.5 Case Study
115
 116

Table 4.12 Summary of the three rescheduling methods.

Rescheduling Approach Advantages Disadvantages Suitable Situations

method

Heuristic-based Block + Speed − Sub-optimal result + Common disruption

preservation + Multiple proposed schedules − Lack of generality + Answer required immediately
heuristic + Few changes from original schedule − Lack of completeness
+ Extensible

Model-based Graph + Generality − Effort required in defining + Heuristic fails

search + Completeness search strategy
4 Robust Supply-Chain Operations through Rescheduling

+ Flexibility in defining search strategy − Computational performance + Adequate time available

+ Multiple proposed schedules depend on search strategy

Total rescheduling MILP solver + Optimal (approximated) − Effort required in formulation + Search space relatively small
+ Completeness − Long computational time + Adequate time available
+ Minimum effort required, same − At most one solution
method as scheduling − Many changes from original
schedule
 4.6 Discussions 117

rescheduling is suitable in situations where the search space is small and time is
available.

4.6
Discussions

This chapter has highlighted the challenges presented by uncertainties and the
vulnerability of today’s SCs. The fact that today’s SCs are increasingly vulnerable to
disruptions is becoming widely accepted and companies are beginning to acknowl-
edge that a trade-off exists between efficiencies and risk. Outsourcing practices are
cost-efficient but they increase the complexity of the SC by adding the number
of players. Complex and lengthy SCs lack visibility and are more susceptible to
disruptions. Single sourcing reduces raw material costs but production continuity
becomes heavily, if not entirely, dependent on that sole supplier. Just-in-time and
lean practices reduce inventory costs but also contingency stocks, which could help
tide through periods of SC glitches. Specialized production sites limit continuity
of the whole SC in the event of disruption to any one site. Strategies that focus
solely on improving efficiency will inevitably increase risk if the supply chain be-
comes fragile as a consequence. A holistic evaluation of the trade-off is therefore
necessary. Disruption management would enable robust SC operations.
One important component of disruption management is rescheduling, as dis-
ruptions may render the current operational schedule infeasible. Robust schedul-
ing is an alternative approach to this problem. However, it improves the extent of
deviations that can be tolerated, at the cost of quality of solution (profit) for the
nominal case. The predictive robust scheduling approach differs from the reactive
rescheduling approach in the way it deals with the solution flexibility versus quality
tradeoff. The former values flexibility over solution quality, while the latter make
amends for uncertainty as necessary as and when it is actually realized. There-
fore, for large deviations that usually result in infeasibility of the original schedule,
rescheduling would be attractive.
Using the refinery crude oil operations as a motivator, we have reviewed three
rescheduling approaches: heuristic-based, model-based, and total rescheduling.
Features, advantages, disadvantages of the three methods and how they comple-
ment one another are described. From the crude oil operations example, it is clear
that four aspects of robust SC operations are important: Flexibility: This could be in
the form of wider operating range of a CDU, which allows more room for through-
put adjustment in response to a disruption. Flexibility could also come from con-
nectivity. When more tanks are connected to a CDU, the refinery has the flexibility
of switching to another tank in case a particular tank is unavailable. It could also
come in the form of raw-material sourcing and product delivery contracts (an issue
that has not been discussed here). Resilience: Rescheduling improves the resilience
of the SC because CDU productions can be maintained in the event of a disrup-
tion. This minimizes the propagation of the disruption to downstream units and
118 4 Robust Supply-Chain Operations through Rescheduling

enables resumption of normal operation. Agility: Robust SC operations require that

a disruption be detected early and rescheduling done quickly to minimize the im-
pact of the disruption. The fast heuristic-based rescheduling method is suitable for
this purpose. Agility also means prompt implementation of the changes brought
by the new schedule. Redundancy: Safety stocks protect against crude stock-out in
case of a crude arrival delay. Backup pumps prevent operational disruptions in case
of a pump failure. Redundancy involves cost but provides protection against uncer-
tainties. The cost-benefit trade-off of redundancy needs to be considered for robust
SC operations.

References

1 Pickett, C. B., Strategies for maxi- 8 Mishra, M., Srinivasan, R., Karimi,
mizing supply chain resilience: learn- I. A., Managing disruptions in refin-
ing from the past to prepare for the ery supply chain using an agent-based
future. M. Eng (Logistics) Thesis. decision support system. Presented in
Massachusetts Institute of Technol- the AIChE Annual Meeting, San Fran-
ogy, 2003. cisco, November 16–21, 2003.
2 Kelly, J. D., Mann, J. L., Crude-oil 9 Mishra, M., Srinivasan, R., Karimi,
blend scheduling optimization: an I. A., A model-based framework for
application with multi-million dol- detecting, diagnosing, and rectifying
lar benefits – Part 1. Hydrocarbon supply chain disruptions. Presented
Process. 82 (2003), pp. 47–54. in the AIChE Annual Meeting, Austin,
3 Reddy, C. P., Karimi, I. A., Srini- TX, November 7–12, 2004.
vasan, R., A new continuous-time 10 Bansal, M., Adhitya, A., Srinivasan,
formulation for scheduling crude oil R., Karimi, I. A., An online decision
operations. Chem. Eng. Sci. 59 (2004a), support framework for managing
pp. 325–1341. abnormal supply chain events. In:
4 Reddy, C. P., Karimi, I. A., Srini- Computer-Aided Chemical Engineer-
vasan, R., Novel solution approach ing, vol. 20, Puigjaner, L., Espuna,
for optimizing crude oil operations. A. (eds.), 2005, pp. 985–990. Elsevier
AIChE J. 50(6) (2004), pp. 1177–1197. Science.

5 Jia, Z., Ierapetritou, M., Efficient 11 Yu, G., Xi, Q.: Disruption Manage-
short-term scheduling of refinery op- ment: Framework, Models and Appli-
erations based on a continuous time cations, World Scientific Publishing,
formulation. Comput. Chem. Eng. 28 Singapore, 2004.
(2004), pp. 1001–1019. 12 Kunnathur, A. S., Sundararagha-
6 Lee, H., Pinto, J. M., Grossmann, van, P. S., Sampath, S., Dynamic
I. E., Park, S., Mixed-integer linear rescheduling using a simulation-
programming model for refinery based expert system. J. Manuf. Tech-
short-term scheduling of crude oil nol. Manage. 15 (2004), pp. 199–212.
unloading with inventory manage- 13 Akturk, M. S., Gorgulu, E., Theory
ment. Ind. Eng. Chem. Res. 35 (1996), and methodology – match-up schedul-
pp. 1630–1641. ing under a machine breakdown. Eur.
7 Singapore Refining Company. Per- J. Op. Res. 112 (1999), pp. 81–97.
sonal communication, 2003. 14 Abumaizar, R. J., Svestka, J. A.,
Rescheduling job shops under ran-
 References 119

dom disruptions. Int. J. Prod. Res. 35 26 Spargg, J. E., Fozzard, G., Tyler,
(1997), pp. 2065–2082. D. J., Constraint based reactive
15 Subramaniam, V., Raheja, A. S., rescheduling in a stochastic environ-
mAOR: a heuristic-based reactive ment. Proceedings, Recent Advances
repair mechanism for job shop sched- in AI Planning. 4th European Con-
ules. Int. J. Adv. Manuf. Technol. 22 ference on Planning, ECP, 1997, pp.
(2003), pp. 669–680. 403–413. Toulouse, France.
16 Henseler, H., Reaktion: a system for 27 Dorn, J., Kerr, R., Thalhammer,
event independent reactive schedul- G., Reactive scheduling in a fuzzy-
ing. In: Artificial Intelligence in Reactive temporal framework. Knowledge
Scheduling, Chapman and Hall, UK, based reactive scheduling. IFIP Trans
1994, pp. 19–31. B (Appl. Technol.) B-15 (1994), pp. 39–
54.
17 Mendez, C. A., Cerdá, J., An MILP
framework for batch reactive schedul- 28 Schmidt, G., How to apply fuzzy
ing with limited discrete resources. logic to reactive production schedul-
Comput. Chem. Eng. 28 (2004), pp. ing. Knowledge based reactive
1059–1068. scheduling. IFIP Trans B (Appl. Tech-
18 Roslöf, J., Harjunkoski, I., nol.) B-15 (1994), pp. 57–66.
Björkqvist, J., Karlsson, S., West- 29 Garner, B. J., Ridley, G. J., Applica-
erlund, T., An MILP-based reorder- tion of neural network process models
ing algorithm for complex industrial in reactive scheduling. Knowledge
scheduling and rescheduling. Comput. based reactive scheduling. IFIP Trans
Chem. Eng. 25 (2001), pp. 821–828. B (Appl. Technol.) B-15 (1994), pp. 19–
19 Vin, J. P., Ierapetritou, M. G., A new 28.
approach for efficient rescheduling of 30 Rovithakis, G. A., Poerrakis, S. E.,
multiproduct batch plants. Ind. Eng. Christodoulou, M. A., Application of
Chem. Res. 39 (2000), pp. 4228–4238. a neural network scheduler on a real
20 Rangsaritratsamee, R., Ferrell, manufacturing system. IEEE Con. Sy.
Jr., W. G., Kurz, M. B., Dynamic 9 (2001), pp. 261–270.
rescheduling that simultaneously con- 31 Qi, J. G., Burns, G. R., Harrison,
siders efficiency and stability. Comput. D. K., The application of the parallel
Ind. Eng. 46 (2004), pp. 1–15. multi population genetic algorithms
21 Belz, R., Mertens, P., Combining to dynamic job shop scheduling. Int.
knowledge-based systems and simula- J. Adv. Manuf. Technol. 16 (2000), pp.
tion to solve rescheduling problems. 609–615.
Decision Sup. Sys. 17 (1996), pp. 141– 32 Aytug, H., Lawley, M. A., McKay,
157. K., Mohan, S., Uzsoy, R., Executing
22 Li, H., Li, Z., Li, L. X., Hu, B., A pro- production schedules in the face of
duction rescheduling expert simu- uncertainties: a review and some fu-
lation system. Eur. J. Op. Res. 124 ture directions. Eur. J. Op. Res. 161(1)
(2000), pp. 283–293. (2005), pp. 86–110.
23 Dorn, J., Case-based reactive schedul- 33 Herroelen, W., Leus, R., Robust and
ing. In: Artificial Intelligence in Reactive reactive project scheduling: a review
Scheduling, Chapman and Hall, UK, and classification of procedures. Int. J.
1994, pp. 32–50. Prod. Res. 42(8) (2004), pp. 1599–1620.
24 Szelke, E., Markus, G., A learning 34 Adhitya, A., Srinivasan, R., Karimi,
reactive scheduler using CBR/L. Com- I. A., Heuristic rescheduling of crude
put. Ind. 33 (1997), pp. 31–46. oil operations to manage abnormal
25 Miyashita, K., Case based knowledge supply chain events. AIChE J. 52(2)
acquisition for schedule optimization. (2007), pp. 397–422.
Artif. Intell. Eng. 9 (1995), pp. 277–287.
120 4 Robust Supply-Chain Operations through Rescheduling

35 Adhitya, A., Srinivasan, R., Karimi, 36 Russell, S. J., Norvig, P., Artifi-
I. A., A model-based rescheduling cial Intelligence: A Modern Approach,
framework for managing abnormal Prentice-Hall, Upper Saddle River,
supply chain events. Comput. Chem. 2003.
Eng. 31(5/6) (2007), pp. 496–518.
 121

5
Supply Chain Tactical Optimization in the Fruit Industry
Guillermo L. Masini, Aníbal M. Blanco, Noemí Petracci, J. Alberto Bandoni

The pip fruit industry is a major economic activity in many countries. The sup-
ply chain of this industry is a complex system, which involves interactions among
many productive, processing, and storage nodes. Fresh fruit harvested in farms is
processed in packaging plants to produce packed fruit and in juice and cider plants
to produce concentrated juice and cider. Storage facilities within the system permit
to keep the quality of the excess of fruit until it can be processed. Large volumes
of fruit have to be transported between the different nodes of the chain during the
whole year. Opposite to many supply chain which are “demand-driven” systems,
the fruit industry supply chains is typically a “production-driven system” since the
flows of goods are “pushed” by fresh fruit availability rather than “pulled” by client
order placement.
The pip fruit business can be thought as having a “negotiation instance” and
an “operation instance.” In the “negotiation instance,” before the beginning of
the business cycle, the company managers pre-establish delivery commitments of
products (packed fruit, concentrated juice, and cider) with different clients based
on estimations of future fruit availability from farms. In the “operation instance”
of the system, the pre-established commitments are sought to be satisfied as close
as possible with the actual fruit income profile. Due to the large amount of in-
volved decisions on what, how much, and when to purchase, store, and allocate
fruit, planning models constitute a valuable tool in the decision-making process.
In this chapter, a detailed and complete tactical planning model to aid in the
negotiation instance of a typical large company that operates several nodes of the
fruit industry supply chain is presented. The proposed linear programming model
considers various interactions of the real network and the typical operative prac-
tices of the business. The model is devised to estimate the “production profiles”
of different final products of the system (packed fruit, juice, and cider) as well as
the “resources profiles” (fresh fruit, storage capacity, and transportation logistics)
required to feasibly operate in order to maximize the total net profit of the company.
Results are presented for a typical large company operating in the supply chain
of the pip fruit industry of Argentina. A very realistic scenario for a negotiation
instance of the company is developed and discussed. A complete set of data is

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
122 5 Supply Chain Tactical Optimization in the Fruit Industry

provided to reproduce the entire case study formulated as a linear programming

model with about 34,000 equations and 143,000 variables. Optimum profiles for
the maximum provision of different products at a predetermined delivery schedule,
along with an estimation of raw material and other resources requirements in order
to reach such a level of production, are determined by the model. These results
constitute valuable information for company managers to negotiate contracts with
clients and suppliers in the coming business cycle.

5.1
Introduction

The pip fruit industry is one of the major agricultural economic activities in several
countries like Germany in Europe, New Zealand in Oceania, South Africa in Africa,
and Chile and Argentina in South America. This industry essentially involves the
growing and processing of apples and pears for commercialization as fresh packed
fruit (the major component of the business), concentrated juice, and bottled cider.
A typical pip Fruit Industry Supply Chain (FISC) essentially comprehends farms
where apples and pears are grown, Packaging Plants (PPs) where fresh fruit is
washed, classified, packed, and cold stored, and processing plants like Concen-
trated Juice Plants (CJPs) and Cider Plants (CPs). There also exist Cold Storage Fa-
cilities (CSFs) to provide product quality maintenance, Fruit Reception Sites (FRSs)
to store raw fruit for juice, and cider production and Milling Plants (MPs) to pre-
process fruit for juice production.
In this chapter, we address the Argentinean pip fruit industry, which is a growing
regional economic activity in this country. The core of this industry is geographi-
cally located in the so-called “Alto Valle de Rio Negro y Neuquen,” a narrow and
very extended valley along two rivers and across two states, placed southwest of the
country. The valley has a cultivated area of 53,320 ha (about two-third of apples and
one-third of pears). The reported production by the Fruit Provincial State Depart-
ment for the season 2005/2006 was 1,568,550 tonnes of apples and 946,643 tonnes
of pears, which constitute the 87% of the total national production. The activity
employs about 51,000 people permanently and more than 136,000 in an indirect
way.
There exist three important markets for a typical large company operating in this
region (local, regional, and overseas), which demands different types of packed
fresh fruit, concentrated juice, and cider. For the Argentinean pip FISC, the lo-
cal market is conformed essentially by some important fruit centers in Argentina,
which distribute the fruit to the rest of the country. The regional market is con-
formed by fruit centers in some important cities of Brazil. The overseas market
comprehends several European countries and the United States of America.
The above raw description of the pip FISC in Argentina can be extended with
nuances to other fruit-type industries (other deciduous varieties, citrus, etc.) and
countries. Therefore, the rigorous model presented in this work can be useful as a
 5.1 Introduction 123

base for the development of others defined by different types of fruits and locations
across the world.
Usually there are only a few large companies that operate along the whole FISC.
A typical such company owns several farms, PPs, CJPs, CPs, MPs, FRSs, and CSFs.
The rest of the industry is conformed by independent farmers, and several inde-
pendent PPs, CJPs, CPs, MPs, and CSFs. Typically, and mainly during harvest peri-
ods, there is a large degree of interconnection among the nodes of the system, and
therefore, many logistic possibilities regarding “when,” “where,” and “how much”
to purchase and allocate of the different goods naturally arise, leading to a very
complex supply chain optimization problem.
Classic good chains are “pull systems” [1] driven by the orders that the customers
place to retailers, which propagates backward through distribution centers, ware-
houses, and processing plants to finally reach raw material suppliers. In such sys-
tems, product availability (adequate inventory) is the key factor to achieve good
customer satisfaction. Unlike classic chains, FISCs are “push systems,” driven by
the fresh fruit produced in farms each season, which determines the spectrum of
final products that will be available for commercialization.
Before the beginning of each harvest season, many business and operational
decisions have to be made by company managers in order to establish final product
delivery contracts with local, regional, and overseas clients. Additionally, to achieve
feasible operation, fruit purchase contracts with third party fruit suppliers and cold
storage renting to third party, CSFs have to be decided, depending on own fruit
availability and processing and storage capacities. We refer to this decision-making
process as the negotiation instance of the business. Such planning activity is based
on estimations of raw material availability (fruit quantity and quality) and forecasts
of costs and prices, and it is naturally driven by a profit maximization objective.
Once the season starts, the system has to be operated within the actual busi-
ness scenario (actual fruit availability, current prices, offer/demand tradeoffs, etc.)
and unexpected episodes have to be considered (poor harvests, out of operation
of plants, on-spot fruit purchase opportunities, etc.). This operation instance of the
business has a target fulfillment objective in order to satisfy as close as possible the
pre-established product delivery commitments.
The complex decisions involved in each business instance make the FISC a chal-
lenging application to computer-aided decision tools; in particular to mathematical
programming modeling. Due to the increasingly competitive worldwide scenario,
managers are compelled to improve the efficiency of the system to keep on busi-
ness. The implementation of mathematical planning tools represents a step in such
direction.
In this chapter, a mathematical programming model for the negotiation instance
of the FISC business is presented. Such a model is considered to be a valuable tool
for a large fruit company that operates several nodes of the FISC, at the time that
contractual obligations with product demanders and third party raw material and
services suppliers must be decided. The proposed model is therefore “tactical” in
nature and relies in estimations of future fruit quality and availability.
124 5 Supply Chain Tactical Optimization in the Fruit Industry

It should be noticed that although the “mass balance model” for the operation in-
stance of the FISC is essentially the same as that for the negotiation instance, the im-
plementation mode would be radically different. At the operation instance, current
data (quality and quantity of the actual incoming fruit, disrupting episodes, etc.)
determines the evolution of the system. Therefore, a “rolling horizon” or “model
predictive control” strategy should be considered. The detailed description of such
a model is beyond the scope of the present article and will motivate future contri-
butions.
For the purposes of the present chapter, it is considered that the company pos-
seses a limited knowledge of third party operation policies and resources, repre-
sented by bounds on the maximum availabilities in certain periods of the season.
In this sense, the supply chain is considered as an individual link of own resources
rather than an integrated network of the whole productive, processing, and storing
capacities of the region. However, competitive coordinated supply chains, which
operate large amounts of goods, are increasingly required to capture international
markets. The proposed model can be easily extended to include third party re-
sources (farms, PPs, CJPs, CPs, MPs, CSFs, and CPs) in order to model such a
coordinated network.
This chapter is structured as follows. In Section 5.2 an overview on planning
contributions in the food industry with particular emphasis in the fruit industry is
provided. Then, in Section 5.3, a typical FISC is described in some detail in order to
motivate the planning model that is presented in Section 5.4. In Section 5.5 results
for a typical company are presented and discussed in some extent. A Conclusions
and Future Work section closes the article.

5.2
Literature Review

Probably due to the lower profit levels regarding other activities (oil, chemical,
petrochemical, etc.) the application of planning models in the fruit industry, and in
general in the food industry, is relatively recent and scarce. However, the need to
become increasingly competitive also reached the food sector and some contribu-
tions on the application of operational research tools to improve the management
of such systems appeared in the last years.
In 2005, the Journal of Food Engineering devoted an entire issue to “Operational
research and food logistics” (Issue 70), where applications of many classic opera-
tions research problems (cutting stock, vehicle routing, planning, scheduling, etc.)
to different food industries where presented. In particular, several contributions
addressed the supply chain management issues of several food systems.
Regarding the fruit industry, decision-making computer-aided tools, in particu-
lar mathematical programming, have been applied to several particular instances
of the FISC. Kearney [2] proposed an intertemporal linear programming model
to explore optimal variety mixes in apple orchards. Since there exist a significant
 5.2 Literature Review 125

range of market prices regarding variety in the pip fruit business, the selection of
an adequate variety mix is an important strategic decision.
In Vitoriano et al. [3], models to schedule farm activities related to the produc-
tion of different crops were proposed. Essentially the models provide the optimal
assignment of resources (staff, machinery, etc.) to tasks (pruning, fertilizing, har-
vesting, etc.) within the seasonal time schedule.
An economic–biophysical model of apple orchards to determine the rate of thin-
ning in a 15-year period in order to maximize the net present value of the system
is described in Hester and Cacho [4].
In Broekmeullen [5] a tactical decision model for Distribution Centers of perish-
ables (vegetables and fruit) is described. The aim of such a model is to provide the
optimal assignment of material within the various storage zones of the facility in
order to minimize the keep quality loss.
A detailed tactical model of a packaging plant in the pip fruit industry was pre-
sented in Blanco et al. [6]. In that contribution a yearly multiperiod planning model
was applied to predict the optimal material flows and inventories within a typical
facility given maximum processing and storage capacities.
According to the authors’ knowledge, only two contributions on FISC operations
have been presented so far: Masini et al. [7] and Ortmann [8].
In Masini et al. [7] a preliminary planning model for a typical large company
operating in the Argentinean FISC was described. The proposed model considered
a 1-year time horizon divided in 12 monthly periods. Given fruit production and
processing and storage capacities in each period, the model provides the optimal
flows and allocation of goods within the available facilities in order maximize total
profit while satisfying a certain demand profile. Although such model captured the
major aspects of the FISC operations, it did not consider many possible scenarios,
for example, the fact that the company might sell part of its production of fresh
or waste fruit to third party manufacturers instead of processing it in their own
plants, or assign cold fruit to juice production if it is convenient due to market
reasons. Besides, daily information on fruit production amount and quality was
lumped into only 12 periods, disregarding meaningful fluctuations associated to
the rather short harvesting slots per variety. Finally, it did not discriminate between
negotiation and operation instances, which better represents the real scenarios of
the business.
In Ortmann [8], the South African fresh fruit export supply chain was described
in detail and models to optimize the fruit business with emphasis in foreign mar-
kets were proposed. The models, developed on a graph theoretic description of
the system, essentially seek to maximize the fruit flow through the existing ex-
port infrastructure and to minimize the transportation costs from storage to ports.
The study was aimed to identify bottlenecks in the whole export supply chain in-
frastructure rather than to provide a management tool for the system operation. In
this sense, the analysis did not consider, for example, local markets, derivative in-
dustries, such as juice production or peripheral costs, such as fruit production and
packaging. The proposed models belong to the multicommodity problems family
and result in single-period linear programs.
126 5 Supply Chain Tactical Optimization in the Fruit Industry

In this contribution, a mathematical programming model for the negotiation in-

stance of the FISC business is presented. It is a detailed and realistic general model
that establishes for a given company the maximum potential annual profiles for
product delivery to the different markets while simultaneously determines profiles
for fruit purchasing and cold storage capacities, all these in an economical optimal
fashion.

5.3
Fruit Industry Supply Chain Description

In this section, the nodes of the FISC system are described and a global picture of
the industry business is provided.

5.3.1
Farms

In farms, fruit of different varieties are grown and harvested according to partic-
ular harvest periods. For mature orchards, the production of each own farm can
be described by daily production average and standard deviation values per vari-
ety, obtained from historical production records. A production cost per variety can
be assigned to each farm either. Such a cost comprehends fruit growing and la-
bor expenses. From the company’s point of view, the production of its farms can
be processed in its own PPs, CJPs, CPs, MPs, sold to third party customers or
cold stored for later processing or selling. Third Party Farms (TPF) supply a large
amount of the fruit processed by the company.

5.3.2
Packaging Plants

The PPs are the core of the FISC and its detailed operations management has been
described elsewhere [6]. In a PP, fresh fruit is received, washed, classified, packed,
and cold-stored or directly delivered to clients.
Regarding classification, nontradable fruit (damaged or too imperfect) is sepa-
rated as waste to be sent to juice and cider production in own or third party plants.
The remaining fruit is further classified by quality (according to the degree of im-
perfection) and gauge (depending on the size or weight of the fruit). The classified
fruit is finally packed before storage or delivery. Packaging options comprehends
several types of crates, boxes, plates, wrapping paper, etc. A given company typi-
cally produces a large amount of different products depending on variety, quality,
size, and packaging type. For the sake of simplicity, such a complex profile will be
lumped into a few categories per variety, one for each market (client). Each product
has a corresponding selling price.
Each PP possesses a maximum processing capacity depending essentially on the
number of classification/packaging lines and has an operating cost corresponding
 5.3 Fruit Industry Supply Chain Description 127

to the number of 8 h labor shifts per day (a minimum of 1 and a maximum of 3)

and maintenance expenses.

5.3.3
Cold Storage

Fresh and packed fruit, as well as concentrated juice, have to be cold stored since
they are perishable goods. Cold storage is performed in CSFs. Packed fruit directly
delivered to clients from the PPs processing lines requires refrigerated transporta-
tion.
Each PP possesses a certain cold storage capacity. There also exist own nonin-
tegrated (outside the packaging plant) and third party cold storage facilities within
the system where fruit can be allocated if PPs’ cold storage capacity is exceeded.
The company can either hire third party cold storage capacity if its own is exceeded
or rent some of its own cold storage capacity if available.
Third party CSFs are described by a hiring cost and own CSFs by a capacity
bound and an operating cost related to the energy required to maintain the fruit at
an adequate temperature.

5.3.4
Fruit Reception Sites

FRSs are open storage sites where large amounts of waste fruit are received from
farms and distributed to MPs, CJPs, and CPs. FRSs have maximum storage capac-
ities and specific operating costs.

5.3.5
Milling Plants

MPs produce milled fruit to be used in CJPs for juice production. The fruit received
at MPs is milled and the produced fruit pulp is delivered to CJPs. Each MP is
described by a maximum processing capacity and an operating cost.

5.3.6
Concentrated Juice Plants

In these facilities, juice of the different types of fruit is produced. Just one type of
concentrated juice is elaborated per type of fruit. Therefore, all apple varieties are
assigned to apple juice production and all pear varieties to pear juice production.
After reception of fruit from own facilities and third party suppliers, batches of
fruit are processed to produce the juice, which is stored in own and third party Cold
Juice Storage Facilities (CJSFs).
Each CJP is described by a fruit-to-juice conversion factor, a maximum process-
ing capacity, and an operating cost corresponding to labor and maintenance ex-
penses.
128 5 Supply Chain Tactical Optimization in the Fruit Industry

5.3.7
Cider Plants

CPs produces cider from apples. Each CP is described by a fruit-to-cider conver-

sion factor, a maximum processing capacity and an operating cost corresponding
to labor and maintenance expenses.

5.3.8
Clients

For any FISC, there exist three major markets that can be classified as local, re-
gional, and overseas. Most of the products of the chain (packed fruit, concentrated
juice, and cider) are preallocated before the start of the season. The different mar-
kets have particular delivery schedules. Packed fruit and concentrated juice are de-
livered by trucks to the local and regional markets and by ship to the overseas
market.

5.3.9
Third Party Suppliers and Customers

In order to consider “on spot” purchase opportunities of available goods within the
system (raw material as well as final products), Third Party Suppliers (TPSs) and
Third Party Customers (TPCs) are included in the chain. The role of such actors is
described later in this section.

5.3.10
Transportation

There exists significant transportation of goods among different nodes of the chain
performed by trucks. Transportation costs depend on the distances between the
nodes and the type of transportation (refrigerated or not refrigerated).

5.3.11
Global FISC Operations

Typical FISC operation is as follows. Fresh fruit harvested in own farms is preclas-
sified in the farms as waste and pretradable fruit. Pretradable fruit is directly sent
to own PPs or temporarily allocated at CSFs for later processing at PPs.
An additional amount of waste fruit is produced in PPs. Waste fruit from own
farms and PPs is used for concentrated juice and cider production. Waste fruit
may be directly sent to MPs, CJPs, and CPs or temporarily stored in FRSs for later
distribution to MPs, CJPs, and CPs.
Pretradable fruit for processing at PPs as well as waste fruit for juice and cider
production can also be purchased from TPFs if required, to maximize PPs, CJPs,
and CPs production levels. Usually, the whole fruit production of several TPFs is
 5.3 Fruit Industry Supply Chain Description 129

prepurchased in order to ensure an adequate supply of raw material to the sys-

tem during the whole year. There also exist along this period a certain “on spot”
availability of raw fruit within the system. Such fruit is generally provided by TPSs,
typically independent farmers or fruit traders who purchase harvested fruit and
keep it cold stored for later selling.
In a common market situation, tradable fruit is processed and sold as packed
fruit and only the waste fruit is used for juice and cider production. However, it
may happen that, for example, concentrated juice reaches a relative high selling
price, turning convenient to elaborate as much juice as possible. Tradable fruit,
and even packed fruit, can be devoted to juice production in such cases.
There also exist TPSs of final products: packed fruit, concentrated juice, and
cider. Although the aim of the company is to produce such goods, it could be neces-
sary to purchase some amount of third party products, for example, to attend some
important delivery commitment.
Regarding product selling, a large amount of the estimated production of packed
fruit, concentrated juice, and cider is allocated to different markets before the be-
ginning of the harvest season. Any excess of production can be sold to TPCs, typ-
ically fruit, juice, and cider retailers. Excess of pretradable fruit produced in own
farms or stored at CSFs as well as waste fruit from own farms and PPs can also be
sold to TPCs, usually independent PPs, CJPs, MPs, and CPs.

5.3.12
Tactical Model Scope

As previously stated, the tactical developed model is intended to aid in the negoti-
ation instance of the business, when contracts with clients and services providers
have to be established.
The model requires estimations of fresh fruit availability from own farms in each
period of the business cycle, also contemplating the possibility that the necessary
amount of fruit can be purchased from TPSs.
The resulting solution, which responds to a profit maximization objective, should
essentially provide, for the installed storage and processing capacity of the com-
pany, a temporal estimation of:
• how much product of the different types (packed fruit, juice,
and cider) will be available each period of the delivery
schedule;
• how much fruit (fresh and waste) should be purchased from
third party suppliers;
• how much cold storage capacity should be required from
third party CSFs;
• how much volume of goods should be transported within the
different nodes of the system.

With these estimations, the company managers can settle two major types of
arrangements:
130 5 Supply Chain Tactical Optimization in the Fruit Industry

1. Final product delivery commitments with the different

packed fruit, juice, and cider clients. The objective is to
allocate as much of the estimated amount of product as
possible in each period of the delivery schedule in order to
maximize the total profit of the business.
2. Third party raw material purchase (fresh and waste fruit) and
services rental (cold storage and transportation) in order to
ensure the feasible operation of the system.

In a first stage, it is considered that the whole final production is allocated as

soon as possible in different markets according to the particular schedules. In other
words, the model is not constrained downstream with particular demands of final
products. The resulting solution is the “production profile” which constitutes the
“business target” of the company.
With that information on maximum potential production of final products, man-
agers can negotiate final selling contracts with the clients establishing definitive
delivery volumes (selling profile) and prices. Hopefully the “selling profile” should
look similar to the optimal “production profile.”
In a second stage, the model should be run again constrained by the actual “sell-
ing profile.” As a result, updated profiles of raw material, cold storage, and trans-
portation requirements will be obtained, providing the required information to es-
tablish contracts with third party suppliers.
In the following section, a mathematical programming model that captures the
above-described FISC business is presented.

5.4
Mathematical Programming Model

In this section, a general model for a large company of a typical FISC is proposed.
Figure 5.1 shows a detailed flow diagram of the supply chain system, where the
names of nodes and streams are indicated. Observe that for clarity sake, several
nodes in Fig. 5.1 have been repeated to avoid excessive line crossing. In particular,
there are two icons that represent PPs and two icons to represent FRSs, each being
source of different streams.
Following the detailed resulting linear programming model is presented.

5.4.1
Fruit Production in Farms

The main input data to the model is the amount of raw material from own sources,
Pf,v,t . Such data is described by average (AP) and standard deviation (SD) values,
obtained form historical records. In order to generate a deterministic value for this
 5.4 Mathematical Programming Model 131

X1
TPC PP Packaging plants
C
X11
X10 X34
TPS OF OF CJP Concentrated juice plants
TPS
TPRS OF
TPF OF
OF
TPPS C
X29
PP MP Mill plants
X4 X3 X2 PFS
X5 X6 X8 X28
TPC CP Cider plants
X7 X9 X30
X23
PP PFS
X12 C Customers
X31
X15
TPS
TPS OF OF TPS OF OF TPS
TPRS OF OF TPS
TPRS OF OF TPS TPC
TPF OF TPF OF TPMS TPC
TPC
X55 X59
X20 X19 X18 X39 X39 FRS CJP X58 X63
NPFS X38 X37 X36 X35 X45 CJS
X22 X24 X40 X41 X42 X56 X61
X21
NPFS TPS OF OF FRS X57 CJS
TPS
TPRS OF OF
OF X43 X53 CJP
TPF
X25
TPC
X50 X50 X49 X48 X47 X46 MP TPS
TPS
X26 X60 TPJS X62 C
X52
X27 TPC
X51 X54 X64
TPS OF
OF OF X32 X13
TPS TPF OF
OF TPC X16 MP
TPRS C

FRS X69 X69 X68 X66 X65 X73

CP C
X67
X41 X44 X70 X71 X72
NPFS Not processed fruit storage OF
OF Own farms
X17 OF
FRS X14 X33 CP
PFS X74 C
X28 PFS Processed fruit storage
OF
OF Third party farms
TPS
TPS TPF
PP TPCS CJS Concentrated juice storage
X6 PFS
TPS Third party supplier
FRS Fruit reception sites
X7
PP TPC Third party consumers

Fig. 5.1 FISC flow diagram.

quantity, a Monte Carlo simulation (MCS) is performed to estimate the production

per variety (Eq. (5.1)) during the corresponding harvest period (HPv )
Pf,v,t = MCS{APf,v , SDPf,v }, ∀f, ∀v, t ∈ HPv . (5.1)

5.4.2
Allocation of Fresh Fruit from Own Farms

Fresh fruit harvested in own farms (Pf,v,t ) can be sold to TPCs (X1) or allocated in
PPs (X2), CSFs (X18), FRSs (X35), MPs (X46), and CPs (X65, only apples). Waste
fruit produced at farms is directly transported to CJPs (X36), MPs (X47), and CPs
(X66). Equation (5.2) represents this balance.




Pf,v,t = X1f,v,t + X2f,p,v,t + X18f,s,v,t + X35f,r,v,t
p s r




+ X36f,r,v,t + X46f,m,v,t + X47f,m,v,t
r m m



+ X65f,cp,v  ,t + X66f,cp,v  ,t ,
cp cp
∀f, ∀v, t ∈ HPv , v  ∈ VWw1 . (5.2)
Set v  ∈ VWw1 comprises only apple varieties. Equation (5.3) calculates the waste
production at farms estimated as a fraction of the whole production and Eq. (5.4)
define this fraction based on average and standard deviation values per variety
(ywff,v ):
132 5 Supply Chain Tactical Optimization in the Fruit Industry



ywff,v,t Pf,v,t = X36f,r,v,t + X47f,m,v,t
r m


+ X66f,cp,v  ,t , ∀f, ∀v, t ∈ HPv , v  ∈ VWw1 . (5.3)
cp

ywff,v,t = MCS{Aywff,v , SDywff,v }, ∀f, ∀v. (5.4)

5.4.3
Fruit Reception at PPs

Fresh fruit entering PPs (X5), Eq. (5.5), has four different sources: own farms
(X2), third party farms (X3), cold nonprocessed fruit from CSFs (X23), and TPRSs
(X4). PPs also possess a maximum and minimum processing capacity as stated in
Eq. (5.6).



X5p,v,t = X2f,p,v,t  + X3p,v,t  + X4p,v,t + X23s,p,v,t , (5.5)
f s

MinProcCapPPp
X5p,v,t
MaxProcCapPPp ,

∀p, ∀v, t  BHPv , t  ∈ HPv . (5.6)
Each fruit variety possess a particular harvest starting day (BHPv ) which deter-
mines the time period in which that variety is available within the system.

5.4.4
Fruit Processing at PPs

From the point of view of the PPs, the whole fruit income can be described by four
parameters per variety (each represented by an average and a standard deviation). A
fraction of nontradable fruit (waste), Eq. (5.7), is devoted for juice and cider produc-
tion, Eq. (5.8). The flows of packed fruit (X6), Eq. (5.11), are calculated as a fraction
per variety for each client, Eqs. (5.9) and (5.10).
ywv,t = MCS{Aywv , SDywv }, ∀v, ∀t. (5.7)


(ywv,t X5p,v,t ) = X7p,w,t , ∀p, ∀w, t  BHPv . (5.8)
v∈w
ycv,c,t = MCS{Aycv,c , SDycv,c }, (5.9)


ycv,c,t = 1, ∀v, ∀c. (5.10)
c
ycv,c,t (1 − ywv,t )X5p,v,t = X6p,v,c,t , ∀p, ∀v, ∀c, t  BHPv . (5.11)

5.4.5
Waste Fruit from PPs

Waste fruit generated at PPs (X7) is sent to FRSs (X15), MPs (X16), and CPs (X17):




X7p,w,t = X15p,r,w,t + X16p,m,w,t + X17p,cp,w1,t ,
r m cp
∀p, ∀w, t  BHPv . (5.12)
 5.4 Mathematical Programming Model 133

5.4.6
Packed Fruit from PPs and TPSs

Packed fruit from PPs (X6) can be delivered to clients (X10), allocated in CSFs (X9)
or sold to TPCs (X11). If convenient, it can also be assigned to juice (X12 and X13)
and cider (X14) production:



X6p,v,c,t = X9p,s,v,c,t + X10p,v,c,t + X11p,v,c,t + X12p,r,v,c,t
s r



+ X13p,m,v,c,t + X14p,cp,v  ,c,t ,
m cp
∀p, ∀v, ∀c, t  BHPv , v  ∈ VWw1 . (5.13)

5.4.7
Fruit Balance at CSFs

In order to keep track of the fruit inventories, a mass balance is performed for
Packed Fruit (PFS) and Nonprocessed Fruit (NPFS) in the different CSFs. It is as-
sumed that each CSF can accommodate nonprocessed (Eq. (5.14)) as well as packed
(Eq. (5.15)) fruit until its maximum storage capacity:




TSs,t = NPFSs,v,t + PFSs,v,c,t , ∀s, t  BHPv . (5.14)
v v c
TSs,t
MaxStoreCapCSCs , ∀s, t  BHPv . (5.15)

Nonprocessed fresh fruit entering NPFS (X21), has three different sources,
Eq. (5.16): own farms (X18), third party farms (X19), and TPSs (X20).


X21s,v,t = X18f,s,v,t  + X19s,v,t  + X20s,v,t ,
f
∀s, ∀v, t  BHPv , t  ∈ HPv . (5.16)

Nonprocessed fruit leaves NPFS (X22), Eq. (5.17), to feed PPs (X23), FRSs (X24),
TPCs (X25), MPs (X26), and CPs (X27):




X22s,v,t = X23s,p,v,t + X24s,r,v,t + X25s,v,t + X26s,m,v,t
p r m


+ X27s,cp,v  ,t , ∀s, ∀v, t  BHPv , v  ∈ VWw1 . (5.17)
cp

Therefore, the nonprocessed fruit inventory is defined by Eq. (5.18):

NPFSs,v,t = NPFSs,v,t−1 + X21s,v,t − X22s,v,t , ∀s, ∀v, t  BHPv . (5.18)

Packed fruit that enters PFS has two different sources, Eq. (5.19): PPs (X9) and
TPSs (X8). There also exist a maximum constraint in the amount of fruit that can
enter the CSFs each period (MaxProcCapCSCs ), Eq. (5.19):



X9p,s,v,c,t + X8s,v,c,t
MaxProcCapCSCs ,
p,v,c v,c
∀s, t  BHPv . (5.19)
134 5 Supply Chain Tactical Optimization in the Fruit Industry

Packed fruit (X28), leaves PFS, to satisfy clients (X29) and, if convenient, is also
assigned to juice and cider production: FRSs (X31), MPs (X32), CPs (X33), and
TPCs (X30) Eq. (5.20):



X28s,v,c,t = X29s,v,c,t + X31s,r,v,c,t + X32s,m,v,c,t
r m


+ X33s,cp,v  ,c,t + X30s,v,c,t ,
cp
∀s, ∀v, ∀c, t  BHPv , v  ∈ VWw1 . (5.20)
Equation (5.21) defines the inventory balance of packed fruit:


PFSs,v,c,t = PFSs,v,c,t−1 + X9p,s,v,c,t + X8s,v,c,t − X28s,v,c,t ,
p
∀s, ∀v, ∀c, t  BHPv . (5.21)

5.4.8
Fruit Balance at FRSs

Waste fruit from own and third party farms (X36 and X38), own PPs (X15), and
TPSs (X39’ ), is usually sent to FRSs (X40), Eq. (5.22), as raw material for juice and
cider plants. Eventually, tradable fruit can also be used for this purpose if turns
economically convenient: from NPFS (X24), own and third party farms (X35 and
X37), PPs (X12), PFSs (X31), and TPSs (X39). In the model, we consider that fruit
at the FRSs, is classified only as apples (sub index w2) and pears (sub index w1).
There also exist constraints for the maximum amount of fruit that can enter the
FRSs in each period (MaxProcCapFRSr ), Eq. (5.23):




X40r,w,t  = X24s,r,v,t + X35f,r,v,t  + X37r,v,t 
v∈VWw s f





+ X12p,r,v,c,t + X31s,r,v,c,t + X39r,v,t
p c s c



+ X36f,r,v,t  + X38r,w,t  + X39r,w,t  + X15p,r,w,t  ,
f p
∀r, ∀w, t  BHPv , t  ∈ HPv , t   BHPw ,
t  ∈ HPv∈VWw . (5.22)


X40r,w,t
MaxProcCapFRSr , ∀r, t  BHPw . (5.23)
w
BHPw is the starting period for the harvest of apples and pears (any variety). Fruit
leaving FRSs (X41), Eq. (5.41) is allocated in CJPs (X42), MPs (X43), CPs (X44), or
eventually sold to TPCs (X45). Only apples (w2) are devoted to cider production.




X41r,w,t = X42r,j,w,t + X43r,m,w,t + X44r,cp,w=w1,t
j m cp
+ X45r,w,t , ∀r, ∀w, t  BHPw . (5.24)
Each FRS possesses a maximum storage capacity (MaxStoreCapFRSr ), Eqs. (5.25)
and (5.26)
FRSr,w,t = FRSr,w,t−1 + X40r,w,t − X41r,w,t , ∀r, ∀w, t  BHPw . (5.25)


FRSr,w,t
MaxStoreCapFRSr , ∀r, t  BHPw . (5.26)
w
 5.4 Mathematical Programming Model 135

5.4.9
Fruit Balance at MPs

Fruit received at MPs (X51), Eq. (5.27), is preferably waste fruit from own and third
party farms (X47 and X49) and from PPs (X16) and TPSs (X50). Eventually, tradable
fruit can be milled for juice production if convenient: from NPFSs (X26), own and
third party farms (X46 and X48), PPs (X13), PFSs (X32), and TPSs (X50). Each MP
possesses a maximum processing capacity (MaxProcCapMPm ), Eq. (5.28):




X51m,w,t  = X26s,m,v,t + X46f,m,v,t  + X48m,v,t 
v∈w s f





+ X13p,m,v,c,t + X32s,m,v,c,t
p c s c



+ X47f,m,v,t  + X50m,v,t + X49m,w,t  + X16p,m,w,t 
f p


+ X43r,m,w,t  + X50m,w,t  , ∀m, ∀w, t  BHPv ,
r
t  ∈ HPv , t   BHPw , t  ∈ HPv∈VWw . (5.27)


X51m,w,t
MaxProcCapMPm , ∀m, t  BHPw . (5.28)
w
Milled fruit from MPs (X52), Eq. (5.29), can be used for juice production at CJPs
(X53) or sold to TPCs (X54):


X52m,w,t = X53m,j,w,t + X54m,w,t , ∀m, ∀w, t  BHPv∈w . (5.29)
j

Each MP possesses a maximum storage capacity, Eqs. (5.30) and (5.31):

MPm,w,t = MPm,w,t−1 + X51m,w,t − X52m,w,t ,

∀m, ∀w, t  BHPv∈w . (5.30)


MPm,w,t
MaxStoreCapMPm , ∀m, t  BHPv∈w . (5.31)
w

5.4.10
Raw Material Reception at CJPs

Raw material for juice production has three different sources: fruit from FRSs
(X42) and TPSs (X55) and milled fruit from MPs (X53). CJPs have maximum ca-
pacity bounds, Eq. (5.32):




X42r,j,w,t + X53m,j,w,t + X55j,w,t
MaxProcCapCJPj ,
w r m
∀j, t  BHPv∈w . (5.32)

5.4.11
Fruit Processing at CJPs

Each CJP can be described by a conversion factor (ycjw, j , in gallons/kg) of raw fruit
or milled fruit into juice per juice variety:
136 5 Supply Chain Tactical Optimization in the Fruit Industry



X56j,w,t = ycjw,j X42r,j,w,t + X53m,j,w,t + X55j,w,t ,
r m
∀j, ∀w, t  BHPv∈w . (5.33)
The produced juice (X56) can be delivered directly to clients (X60), cold-stored
(X58) or sold to TPCs (X59):



X56j,w,t = X58j,sj,w,t + X60j,w,cj,t + X59j,w,t ,
sj cj
∀j, ∀w, t  BHPv∈w . (5.34)

5.4.12
Juice Storage at CJSs

Cold-stored juice has two sources: juice produced in CJPs (X58) and juice pur-
chased from TPSs (X57). Juice leaves CJSs (X61), Eq. (5.35), to attend client com-
mitments (X62), and eventual TPCs (X63). CJSs possess a maximum storage ca-
pacity, Eqs. (5.36) and (5.37).


X61sj,w,t = X62sj,w,cj,t + X63sj,w,t , ∀sj, ∀w, t  BHPv∈w . (5.35)
cj


CJSsj,w,t = CJSsj,w,t−1 + X58j,sj,w,t + X57sj,w,t − X61sj,w,t ,
j
∀sj, ∀w, t  BHPv∈w . (5.36)


CJSsj,w,t
MaxStoreCapCJSsj ,
w
∀sj, t  BHPv∈w . (5.37)

5.4.13
Raw Material Reception at CPs

Fruit devoted to cider production (X70), Eq. (5.38), is preferably waste fruit from
own and third party farms (X65 and X67), PPs (X17), FRSs (X44), and TPSs (X69).
Eventually, tradable fruit can be used for cider production if convenient: from NPFS
(X27), own and third party farms (X66 and X68), PP (X14), PFS (X33), and TPSs
(X69). CPs have a maximum capacity bound (Eq. (5.39)):





X70cp,t  = X27s,cp,v,t + X14p,cp,v,c,t
v∈VWw1 s p c





+ X33s,cp,v,c,t + X65f,cp,v,t  + X66f,cp,v,t 
s c f f


+ X67cp,v,t  + X69cp,v,t + X17p,cp,w1,t 
p


+ X44r,cp,w1,t  + X68cp,w1,t  + X69cp,w1,t  ,
r
∀cp, t  BHPv , t  ∈ HPv , t   BHPw1 , t  ∈ HPv∈VWw1 .
(5.38)
X70cp,t
MaxProcCapCPcp . (5.39)
 5.4 Mathematical Programming Model 137

5.4.14
Fruit Processing at CPs

Each CP can be described by a conversion factor of apples into cider (ycpcp in l/kg).
In each CP, cider is stored (CScp ) until delivery Eq. (5.40). Each cider storage facility
has a maximum storage capacity (MaxStoreCapCScp ), Eq. (5.41):
CScp,t = CScp,t−1 + ycpcp X70cp,t − X71cp,t , ∀cp, t  BHPw1 . (5.40)
CScp,t
MaxStoreCapCScp , ∀cp, t  BHPw1 . (5.41)
Final product, produced at the CPs (X71) and purchased from TPSs (X74), is
directly sent to clients (X72) and TPCs (X73), Eq. (5.42):


X71cp,t + X74cp,t = X72cp,cc,t + X73cp,t , ∀cp, t  BHPw1 . (5.42)
cc

5.4.15
Packed Fruit Delivery to Clients

Packed fruit to clients (X34), Eq. (5.43), comes directly from PPs (X10) and
PFS (X29):



X34v,c,t = X10p,v,c,t + X29s,v,c,t , ∀v, ∀c, t  BHPv . (5.43)
p s

5.4.16
Concentrated Juice Delivery to Clients

Concentrated juice to clients (X64), Eq. (5.44), can be delivered directly from CJPs
(X60) and from CJSs (X62):



X64w,cj,t = X60j,w,cj,t + X62sj,w,cj,t , ∀w, ∀cj, t  BHPw .(5.44)
j sj

5.4.17
Third Party Material Availability

It is considered that there exists a maximum availability of goods from third

party sources. Equations (5.45), (5.46), (5.47), and (5.48) correspond to fresh fruit,
Eq. (5.49) to packed fruit, Eq. (5.50) to milled fruit, Eq. (5.51) to concentrated juice,
and Eq. (5.52) to cider.
Raw fruit from TPFs:





X3p,v,t + X19s,v,t + X37r,v,t + X48m,v,t
p s r m


+ X67cp,v  ,t
MaxFFTPFv,t ,
cp
∀v, t ∈ HPv , v  ∈ VWw1 . (5.45)




X38r,w,t + X49m,w,t + X68cp,w1,t
MaxWFTPFw,t ,
r m cp
∀w, t ∈ HPv∈w . (5.46)
138 5 Supply Chain Tactical Optimization in the Fruit Industry

Raw fruit from TPSs:

X4p,v,t + X20s,v,t + X39r,v,t + X50m,v,t
p s r m


+ X69cp,v  ,t
MaxFFTPRSv , ∀v, t  BHPv , v  ∈ VWw1 . (5.47)
cp

X39r,w,t + X50m,w,t + X69cp,w1,t
MaxWFTPRSw ,
r m cp
∀w, t  BHPw . (5.48)
Packed fruit from TPSs:



X8s,v,c,t
MaxPFTPPSv , ∀v, t  BHPv . (5.49)
s c

Milled fruit from TPSs:

X55j,w,t
MaxMFTPMSw , ∀w, t  BHPw . (5.50)
j

Concentrated juice from TPSs:

X57sj,w,t
MaxCJTPJSw , ∀w, t  BHPw . (5.51)
sj

Cider from TPSs:

X74cp,t
MaxCTPCS, t  BHPw1 . (5.52)
cp

5.4.18
Product Delivery Constraints

Products to particular markets can only leave the system according to given
schedules, therefore, delivery variables are set to zero, for other than certain pre-
established time periods. See Eqs. (5.53), (5.54), and (5.55):
X34v,c,t = 0, ∀v, ∀c, t ∈
/ DSc . (5.53)
X64w,cj,t = 0, ∀w, ∀cj, t ∈
/ DScj . (5.54)
X72cp,cc,t = 0, ∀cp, ∀cc, t ∈
/ DScc . (5.55)

5.4.19
Raw Material Costs

This item comprehends production costs at own farms and purchase costs of fresh
and waste fruit, each with a particular specific cost, Eqs. ?eq558??eq559??eq560??eq561??eq562?(5.56)–
(5.62). Farms production:

OFFFProdCt = ffprcf,v Pf,v,t . (5.56)
f v

TPFs fresh fruit purchase cost:

 5.4 Mathematical Programming Model 139





TPFFFPurchCt = tpfffpucv X3p,v,t +
X19s,v,t + X37r,v,t
v p s r



+ X48m,v,t + X67cp,v∈VWw1,t . (5.57)
m cp

TPSs fresh fruit purchase cost:




TPRSFFPurchCt = tprsffpucv X4p,v,t + X20s,v,t
v p s




+ X39r,v,t X50m,v,t + X69cp,v∈w1,t .
r m cp

(5.58)

TPFs waste fruit purchase costs:




TPFWFPurchCt = tpfwfpucw X38r,w,t + X49m,w,t
w r m



+ X68cp,v∈w1,t . (5.59)
v∈w cp

TPRSs waste fruit purchase costs:




TPRSWFPurchCt = tprswfpucw X39r,w,t + X50m,w,t
w r m


+ X69cp,w1,t . (5.60)
cp

TPRSs milled fruit purchase costs:

TPRSMFPurchCt = tprsmfpucw X55j,w,t . (5.61)
w j

Total raw material costs:

TRMCt = OFFFProdCt + TPFFFPurchCt + TPRSFFPurchCt

+ TPFWFPurchCt + TPRSWFPurchCt
+ TPRSMFPurchCt . (5.62)

5.4.20
Final Products Purchase Cost

Equations (5.63)–(5.66) define the single terms and total fruit purchase costs.
Packed fruit purchase cost:

PFPurchCt = pfpucv,c X8s,v,c,t . (5.63)
v c s

Concentrated juice purchase cost:

CJPurchCt = cjpucw X57sj,w,t . (5.64)
w sj
140 5 Supply Chain Tactical Optimization in the Fruit Industry

Cider purchase cost:

CPurchCt = tpcscpuc X74cp,t . (5.65)
cp

Total final products purchase costs:

TFPPCt = PFPurchCt + CJPurchCt + CPurchCt . (5.66)

5.4.21
Transportation Costs

This item involves transportation costs among the different nodes of the chain.
Such costs essentially depend on the distance between facilities and a specific trans-
portation cost, which is discriminated as refrigerated and not refrigerated. Equa-
tions (5.67)–(5.90) define single terms and total transportation costs. From farms

to PPs:




TCfFtPPt = nrtc df,p X2f,p,v,t . (5.67)
f p v

From farms to CSFs:



TCfFtNPFSt = nrtc df,s X18f,s,v,t . (5.68)
f s v

From farms to FRSs:



TCfFtFRSt = nrtc df,r (X35f,r,v,t + X36f,r,v,t ) . (5.69)
f r v

From farms to MPs:



TCfFtMPt = nrtc df,m (X46f,m,v,t + X47f,m,v,t ) . (5.70)
f m v

From farms to CPs:



TCfFtCPt = nrtc df,cp (X65f,cp,v∈VWw1,t
f cp v

+ X66f,cp,v∈VWw1,t ) . (5.71)

From CSFs to PPs:



TCfNPFStPPt = nrtc dp,s X23s,p,v,t . (5.72)
s p v

From CSFs to FRSs:

 


TCfCSCtFRSt = nrtc ds,r X24s,r,v,t + X31s,r,v,c,t .
s r v c
(5.73)
From CSFs to MPs
 5.4 Mathematical Programming Model 141


 


TCfCSCtMPt = nrtc ds,m X26s,m,v,t + X32s,m,v,c,t .
s m v c
(5.74)
From CSFs to CPs:


 

TCfCSCtCPt = nrtc ds,cp X27s,cp,v∈w1,t
s cp v


+ X33s,cp,v∈w1,c,t . (5.75)
c

From CSFs to clients:



TCfPFStCt = rtc ds,c X29s,v,c,t . (5.76)
s c v

From PPs to CSFs: PPs have certain cold storage capacity. For the sake of formality

such facilities have been treated as non-integrated CSFs in this model. The CSFs
associated to a PP will be described by a distance value (dp,s ) of zero in the following
equation:





TCfPPtPFSt = rtc dp,s X9p,s,v,c,t . (5.77)
p s v c

From PPs to clients:



TCfPPtCt = rtc dp,c X10p,v,c,t . (5.78)
p c v

From PPs to FRSs:

 


TCfPPtFRSt = nrtc dp,r X12p,r,v,c,t
p r v c


+ X15p,r,w,t . (5.79)
w

From PPs to MPs:

 


TCfPPtMPt = nrtc dp,m X13p,m,v,c,t
p m v c


+ X16p,m,w,t . (5.80)
w

From PPs to CPs:

 


TCfPPtCPt = nrtc dp,cp X14p,cp,v∈w1,c,t
p cp v c

+ X17p,cp,w1,t . (5.81)

From FRSs to MPs:



TCfFRStMPt = nrtc dr,m X43r,m,w,t . (5.82)
r m w
142 5 Supply Chain Tactical Optimization in the Fruit Industry

From FRSs to CPs:

TCfFRStCPt = nrtc (dr,cp X44r,cp,w1,t ). (5.83)
r cp

From FRSs to CJPs: Each CJP possesses a FRS. For the sake of formality such facil-

ities have been treated as non-integrated FRS in this model. The FRS associated to
a CJP will be described by a distance value (dr,j ) of zero in the following equation:




TCfFRStCJPt = nrtc dr,j X42r,j,w,t . (5.84)
r j w

From MPs to CJPs:



TCfMPtCJPt = mftc dm,j X53m,j,w,t . (5.85)
m j w

From CPs to clients:

TCfCPtCt = nrtc (dcp,cc X72cp,cc,t ). (5.86)
cp cc

From CJPs to CJSs:



TCfCJPtCJSt = nrtc dj,sj X58j,sj,w,t . (5.87)
j sj w

From CJPs to clients:



TCfCJPtCt = nrtc dj,cj X60j,w,cj,t . (5.88)
j cj w

From CJSs to clients:



TCfCJStCt = nrtc dsj,cj X62sj,w,cj,t . (5.89)
sj cj w

Total transportation costs:

TTCt = TCfFtPPt + TCfFtNPFSt + TCfFtFRSt + TCfFtMPt + TCfFtCPt

+ TCfNPFStPPt + TCfCSCtFRSt + TCfCSCtMPt + TCfCSCtCPt
+ TCfPFStCt + TCfPPtPFSt + TCfPPtCt + TCfPPtFRSt
+ TCfPPtMPt + TCfPPtCPt + TCfFRStMPt + TCfFRStCPt
+ TCfFRStCJPt + TCfMPtCJPt + TCfCPtCt + TCfCJPtCJSt
+ TCfCJPtCt + TCfCJStCt . (5.90)

5.4.22
Cold Storage Cost (Third Party Rental)

Third party cold storage cost is described by a rental cost.

TPCSCt = ttppcscs TSs,t . (5.91)
s∈TPS
 5.4 Mathematical Programming Model 143

5.4.23
Operating Costs

Operating costs comprehend processing costs at PPs, CJPs, MPs, and CPs, opera-
tional costs at FRSs and cold storage costs at own CSFs. The first, essentially involve
labor and maintenance expenses, while cold storage costs mainly has to do with the
energy required to refrigerate the goods (fruit and juice). In all cases the operating
costs are proportional to the amount of material operated at the facility. Equations
(5.92)–(5.102) defines single and total operating cost terms. PPs’ operating costs:



OCPPt = ooccppp X5p,v,t . (5.92)
p v

FRSs’ operating costs:



OCFRSt = ooccfrsr X40r,w,t . (5.93)
r w

MPs’ operating costs:



OCMPt = ooccmpm X51m,w,t . (5.94)
m w

CPs’ operating costs:

OCCPt = (oocccpcp X70cp,t ). (5.95)
cp

CJPs’ operating costs:

vspace12pt The operating cost of CJPs is conformed by two terms: the fresh fruit
processing cost and the milled fruit processing cost.

 


OCCJPFFt = oocccjpffj X42r,j,w,t . (5.96)
j r w

 


OCCJPMFt = occjpfmj X53m,j,w,t + X55j,w,t .
j w m
(5.97)
OCCJPt = OCCJPFFt + OCCJPMFt . (5.98)
Own cold storage operating cost:




OCSCt = ooccscs TSs,t + ooccjjscsj CJSsj,w,t . (5.99)
s∈OS sj w

Own MPs storage operating cost:

OMPSOCt = ompsocm MPm,w,t . (5.100)
m w
Own CPs storage operating cost:


OCPSOCt = ocpsoccp CScp,t . (5.101)
cp

Total operating costs:

144 5 Supply Chain Tactical Optimization in the Fruit Industry

TOCt = OCPPt + OCFRSt + OCMPt + OCCPt + OCCJPt

+ OCSCt + OMPSOCt + OCPSOCt . (5.102)

5.4.24
Sales Income

A major part of the earnings of the system comes from the selling of the different
final products (packed fruit, concentrated juice, and bottled cider). Fresh fruit from
farms and waste fruit from PPs and farms can also be sold to TPCs if convenient.
Equations (5.103)–(5.112) represent single and total income terms. Packed fruit:

SIPFCt = pfcpv,c X34v,c,t . (5.103)
v c


 


SIPFTPCt = pftpcpv,c X11p,v,c,t + X30s,v,c,t .
v c p s

(5.104)

Concentrated juice:



SICJCt = cjcpw,cj X64w,cj,t . (5.105)
w cj

 


SICJTPCt = cjtpcpw X59j,w,t + X63sj,w,t . (5.106)
w j sj

Cider:



SICCt = ccpcc X72cp,cc,t . (5.107)
cc cp


SICTPCt = (tpccpX73cp,t ). (5.108)
cp

Fresh fruit:

 


SIFFt = ffpv X1f,v,t + X25s,v,t . (5.109)
v f s

Waste fruit:



SIWFt = wfpw X45r,w,t . (5.110)
w r

Milled fruit:



SIMFt = mfpw X54m,w,t . (5.111)
w m

Total sales income:

TSIt = SIPFCt + SIPFTPCt + SICJCt + SICJTPCt + SICCt + SICTPCt

+ SIFFt + SIWFt + SIMFt . (5.112)
 5.5 Results and Discussion 145

5.4.25
Cold Storage Capacity Rental Income

If the company possesses a large cold storage capacity, which exceeds its own re-
quirements, it can consider the rental of the available chambers, as defined by
Eq. (5.113)


CSCRIt = ocsprs (MaxStoreCapCSCs − TSs,t ). (5.113)
s∈OS

5.4.26
Objective Function

The objective function (OF) is the net profit as defined by Eq. (5.114):


TotalProfit = (TSIt + CSCRIt − TRMCt − TFPPCt − TTCt
t
− TPCSCt − TOCt ). (5.114)

5.4.27
Company Supply Chain Optimization Problem

The final model that sets the optimum profiles for maximum product availability,
fresh and processed fruit requirements, and cold storage, and transportation along
a business horizon equivalent to a fruit production cycle, is:

Maximize OF = Total Profit

s.t.
{Eqs. (5.1) to (5.114)} (5.115)
Equation (5.115) represents a linear programming formulation.

5.5
Results and Discussion

In order to illustrate the capabilities of the proposed model, a particular instance

representing a typical company within the described FISC is considered for analy-
sis. The studied instance takes into account eight fruit varieties, three types of
packed fruit (one for each of the three different markets), two types of concen-
trated juice (apple and pear), and bottled cider. The system comprehends five own
farms, two PPs, four fruit CSFs, two juice CSFs, two CJPs, three FRSs, two MPs,
and two CPs. A weekly discretization of the time domain was adopted for a 1-year
planning horizon corresponding to a typical pip fruit business cycle. Therefore, 52
time periods were used in the model.
For the purposes of the present analysis, it is considered that raw material (fresh
and waste fruit) is available only from TPFs and none from other TPSs, which usu-
ally can provide rather small amounts of fruit at only certain time periods during
146 5 Supply Chain Tactical Optimization in the Fruit Industry

the year. It is also considered that no final products (packed fruit, concentrated
juice, and cider) are purchased from TPSs to complement production.
The complete set of data for the described instance is provided in the appendix
to this chapter. The reported data realistically reflect a typical FISC company in-
frastructure. Model parameters were taken from annual reports published by the
Federal Investment Council [9], direct knowledge from the authors, and personal
communications with plant engineers working at different companies.
Fresh fruit related data are reported in Table 5.1. It should be noted that, in gen-
eral, the average and standard deviation values for ywv and ycv, c are time- and
farm-dependent. They should vary within the harvest period to take into account
the effect of the so-called “partial harvests.” In each partial harvest (normally two or
three) of a given variety, only selected fruit is cut. The amount of cut depends on the
pruning, training, and thinning practices, which usually vary from farm to farm.
For the sake of simplicity, single representative values of average and standard de-
viations per variety are considered in the present work. Parameter ycv, c describes
quality distribution of waste-free fruit among different clients. Since three packed
fruit clients are considered, figures for two of them are provided (the remaining is
calculated by difference).
Production data per farm per variety are reported in Table 5.2 and waste fraction
production from farms in Table 5.3. Such figures also fluctuate with time due to
partial harvesting effects as described above but are considered constants to the
purposes of the present model.
Prices of different goods within the system are reported in Tables 5.4–5.11. Sell-
ing prices of final products (packed fruit, concentrated juice, and cider), especially
those corresponding to the international market, posses a large degree of uncer-
tainty since they depend on the global fruit market situation (fruit availability,
clients preferences, etc.). In fact, the definitive figures are determined during the
negotiation instance as a result of offer–demand tradeoffs. Such figures are diffi-
cult to set and usually forecasting techniques are required for proper evaluation.
For the purposes of the present model, data corresponding to a typical business
scenario where the packed fruit “pulls” the business were adopted.
Transportation and operating costs related data are reported in Tables 5.12–5.18.
Processing costs at CSFs, CJSs CPs, FRSs, and MPs, and transportation costs, al-
though subject to certain uncertainty can be also considered fair estimations of the
real operations. Constant values have been adopted for the whole season in the
present model.
Distances between the different nodes of the chain, required to evaluate the
transportation costs, are provided in Tables 5.19–5.40. It should be noted that the
CSF associated to each PP as well as the FRS associated to each CJP are described
in the corresponding tables with a distance of zero.
Maximum processing and storage capacities of the different facilities are re-
ported in Tables 5.40–5.47. Conversion factors of fruit into juice and cider are re-
ported in Tables 5.48 and 5.49. Table 5.50 reports product delivery schedules for
different markets.
 5.5 Results and Discussion 147

Fig. 5.2 Available packed fruit for foreign market.

Fig. 5.3 Available packed fruit for regional market.

The resulting linear programming model was implemented in the GAMS lan-
guage [10] and expands 33,959 equations and 143,437 variables for the considered
horizon of 52 time periods. It takes 39 s of CPU time to solve the problem on a
Pentium IV, 2.4 GHz, 512 Mb RAM personal computer.
148 5 Supply Chain Tactical Optimization in the Fruit Industry

Fig. 5.4 Available packed fruit for local market.

In Figs. 5.2–5.4, the expected “production profiles” of available packed fruit per
variety in the corresponding delivery dates for foreign, regional, and local markets,
respectively, are shown. As already commented, the model assumes that the whole
amount of available packed fruit can leave the system in each delivery date. In
order to maximize the total profit of the system the model effectively “sells” all the
available fruit at CSFs, which on one side maximizes the sales income and on the
other minimizes the cold storage costs by empting the chambers.
It can be seen that in all the cases the fruit availability per variety in each delivery
date is related with the harvest period of that variety within the year. For example,
variety 1 which is the first harvestable is available in the first few delivery dates for
all clients, while variety 8, the last harvested, can be only delivered at the final dates
of the delivery schedules.
The “production profiles” of juice of both types (w1: pear, w2: apple) are shown
in Fig. 5.5. As for packed fruit, the optimal solution corresponds to the selling of
the whole existence in the delivery date, which corresponds to a complete empty of
the juice chambers as can be observed in Fig. 5.6. A similar behavior presents the
cider business (Fig. 5.7).
In order to achieve feasible operation of the system during the season, certain
amount of raw fruit from third party farms should be purchased. The rental of third
party storage capacity should be required as well, if not enough space is available
in own facilities. Finally, the transport logistics to move the required volumes of
fruit among different facilities of the system should be also provided for feasible
operation.
Figures 5.8 and 5.9 depicts the required profile of fresh and waste fruit per va-
riety, respectively, that should be purchased from third party farms to achieve the
product profiles reported in Figs. 5.2–5.5, and 5.7. The purchased fruit is used
to complement the production in own farms for packed fruit and juice and cider
 5.5 Results and Discussion 149

Fig. 5.5 Available concentrated juice.

Fig. 5.6 Concentrated juice inventory profile.

production. This information is required for settling arrangements of fresh fruit

purchase with third party suppliers, before the harvest season.
In Fig. 5.10, the evolution along the season of own (solid line) and third party
(dashed line) cold storage are shown. It can be observed that own cold storage
capacity saturates (maximum of 4000 tonnes) during periods 21–25 in the middle
of the season when fresh fruit income is large. In these periods, there exists a
peak in third party storage requirement as can be observed in the same figure.
Towards as the end of the harvest, when the fresh fruit income decreases, the plants
150 5 Supply Chain Tactical Optimization in the Fruit Industry

Fig. 5.7 Available bottled cider.

Fig. 5.8 Fresh fruit purchase requirement.

continue processing the stored fruit. As expected, there exist a steeper fall in third
party storage than in own storage since that fruit is processed first to minimize
third party storage which is more expensive than own storage.
In Fig. 5.11, the evolution of the own cold storage along the season is presented.
Results correspond to CSFs integrated into PPs and nonintegrated or “stand alone”
CSFs, usually placed far from PPs. It can be observed that the profile of integrated
CSFs (dashed line) are more “erratic” than that corresponding to non-integrated
ones (solid line), which shows plateaus in certain periods, especially from the mid-
dle of the year when the harvest season is over. This is an expected result since
 5.5 Results and Discussion 151

Fig. 5.9 Waste fruit purchase requirement.

Fig. 5.10 Cold storage requirement.

there exist a transportation cost to move fruit from nonintegrated CSFs to PPs and
any such movement is discouraged by the model in favor of processing first the
fruit available at the PPs. By the end of the season, nonintegrated CSFs show a
monotone decrease since nonprocessed fruit is continuously withdrawn, but the
produced packed fruit is stored at the CSFs of the PPs, rather than transported
back to the nonintegrated CSFs.
In Fig. 5.12, the transportation requirement, classified as refrigerated and nonre-
frigerated, is presented. A large nonrefrigerated transportation activity is observed
in the first part of the season (approximately until period 25) when the harvest of
152 5 Supply Chain Tactical Optimization in the Fruit Industry

Fig. 5.11 Own cold storage profile.

Fig. 5.12 Transportation requirements.

different varieties takes place, reflecting the transport of fresh fruit from farms to
plants. When the harvest finishes the transportation activity reduces since only
fruit from cold storage to processing plants is required. Regarding refrigerated
transport, a larger activity is also observed during the harvest period since a large
volume of overseas packed fruit is delivered during this period, which is no longer
demanded when the harvest is over.
It should be remarked that the solution reported in Figs. 5.2–5.12 correspond
to the best possible business scenario, when the whole production is delivered as
soon as available. In order to maximize profit, the company should try to allocate
 5.5 Results and Discussion 153

Fig. 5.13 Total processing at packaging plants.

as much as possible packed fruit, concentrated juice, and cider in different periods
according to the “production profiles” of Figs. 5.2–5.5, and 5.7.
Table 5.51 presents the main economic results for such scenario. A maximum
net profit of US$ 4,384,291 could be obtained if contracts can be negotiated to
sell the total production of the company for an amount of US$ 47,204,180. The
corresponding total cost is US$ 42,819,889, conformed by 54% of operating cost,
24% of raw material cost, 21% of transportation cost, and 1% of third party cold
storage cost.
The “selling profile” resulting from the negotiation between the company and
the clients should match the “production profile” for optimal performance. In gen-
eral, however, that will not be the case since the products cannot usually be com-
pletely allocated and the amounts of the “selling profile” are usually lower than the
corresponding to the “production profiles” for many delivery dates. Moreover, it is
expected that as a result of the negotiation, the final selling prices will be updated
with respect to the original estimation.
In such a case the model could be run again including the actual delivery con-
straints imposed by the “selling profile” and the definitive selling prices. New pro-
files of fresh and waste fruit that should be purchased to third party farms would
result as well as a new third party cold storage requirement and transportation
volumes. With this updated information, definitive contracts could be agreed with
third party suppliers.
Besides the “practical” information for the “negotiation instance” provided by
Figs. 5.2–5.5, 5.7–5.10, and 5.12, some additional insight on the operations of the
system can be gained from the model results. For example, the total processing at
packaging plants during the season can be seen in Fig. 5.13. As can be expected,
the model shows that PPs operates at the maximum possible rate in the first part
154 5 Supply Chain Tactical Optimization in the Fruit Industry

of the year, during the harvest period when the amount of fresh fruit entering
the system is large. Once the harvest is over, the PPs reduces its operation to an
allowed minimum in order to process the remaining fruit at CSFs. The maximum
processing capacity of the PPs, given by a maximum related with the number of
labor shifts and processing lines, constitutes a “bottleneck” of the system.

5.6
Conclusions and Future Work

In this work, a complete and detailed tactical mathematical programming model

of a typical pip fruit industry supply chain has been presented which takes into
account the complex interactions among its many nodes. It was devised as a de-
cision mathematical tool to aid in the negotiation instance of the business when
commercial commitments have to be arranged with clients and suppliers.
The proposed model provides the “production profiles” of packed fruit, concen-
trated juice, and cider which should be pursued to optimize the business profit,
along with the required profiles of raw fruit (fresh and waste), third party cold stor-
age, and transportation to feasibly operate throughout the fruit business cycle.
Due to the high uncertainty present in the system, especially in the availability
of fresh fruit and in the selling prices of the final products in the coming season,
the model is also considered to be a valuable tool for FISC companies’ managers to
perform “what if” analysis for various possible productive and market scenarios.
The potential of the model was demonstrated by solving a case study correspond-
ing to a typical company operating several packaging, concentrated juice, and cider
plants, storage facilities, and production farms. A set of profiles for the main eco-
nomic and operative variables was obtained, as a demonstration of the valuable
information provided by the model that company managers could have at their dis-
posal to negotiate the commercial contracts with product clients at different mar-
kets and suppliers, in advance to the beginning of the fruit business cycle.
A natural extension of the present work is to formulate an operational planning
model of the FISC company, which considers its unique features regarding other
value chains such as its “raw material fruit availability driven nature.”
Such a model would have a “demand fulfillment objective” since it is required
to satisfy as close as possible the product delivery commitments pre-established
before the season. In order to reach the required amounts of the different products
(packed fruit, juice, and cider) in the delivery periods, a “model predictive control
strategy” could be implemented to fairly account the current amounts of fruit that
comes into the system as well as possible “operation disrupting events” and “on
spot” business opportunities.
The model should consider a daily realization of the timeline and a short/medium
predictive horizon, say 15/30 days. Each day, the actual amounts of fruit entering
the system should be fed to the model together with estimations of raw fruit avail-
ability in the considered prediction horizon, including eventual TPSs availability.
Every flow of material within the system in the considered horizon can therefore
 Acknowledgment 155

be calculated in order to minimize “client dissatisfaction.” Since there exist a rather

large degree of uncertainty in the estimations of fruit availability for the future pe-
riods, only the results for the first period should be implemented. This process has
to be repeated for each day with current data and improved estimations.
The commented model should also consider the discrete nature of the flow rates
between the different nodes of the system. Since the goods are transported in
trucks, only an integer number of travels of fully charged trucks can be performed
between to facilities each day. This feature would require a MILP formulation for
the underlying mass balance model of the model predictive control strategy.
In other level of the decision-making hierarchy, future work also remains in the
“strategic planning” of the FISC.
For example, it has been shown that the maximum processing capacity of PPs
constitutes a bottleneck within the system. Increasing such capacities should ren-
der in larger productions of packed fruit and therefore in a larger business volume.
More fresh fruit from third party farms would be required for processing at PPs but
less waste fruit will be demanded from third party suppliers since larger amounts
of waste fruit will be produced at own PPs for processing at CJPs and CPs. For
an improved PPs processing capacity, it might be convenient to invest in enlarg-
ing own fruit production (by purchasing new farms) and possibly in increasing the
production capacity of concentrated juice and cider of the company.
Future contributions will address both, the operational and the strategic planning
of a typical FISC company.

Acknowledgment

The authors want to acknowledge to the Universidad Nacional del Sur, the Consejo
Nacional de Investigaciones Científicas y Técnicas, and the Agencia Nacional de
Promoción Científica y Técnica of Argentina for financial support. Eng. Masini
also wants to thank to the Engineering Faculty of the Universidad Nacional del
Comahue (Neuquén, Argentina).

References

1 Perea Lopez, E., Ydstie, B. E., Gross- alternative models for farm manage-
mann, I., A model predictive control ment: discrete versus continuous time
strategy for supply chain optimiza- horizon, Eur. J. Oper. Res. 144 (2003),
tion, Comp. Chem. Eng. 27 (2003), pp. pp. 613–628.
1201–1218. 4 Hester, S. M., Cacho, O., Modeling
2 Kearney, M., An Intertemporal Linear apple orchard systems, Agric. Syst.
Programming Model for Pipfruit Or- 77(2) (2003), pp. 137–154 (18).
chard Replacement Decisions; MAF 5 Broekmeullen, R., Operations man-
Policy Technical Paper 94/6, 1994. agement of distribution centers for
3 Vitoriano, B., Ortuño, M. T., Recio, vegetables and fruits, Int. Trans. Opl.
B., Rubio, F., Alonso-Ayuso, A., Two Res. 5(6) (1998), pp. 501–508.
156 5 Supply Chain Tactical Optimization in the Fruit Industry

6 Blanco, A. M., Masini, G., Petracci, 8 Ortmann, F. G., Modelling the South
N., Bandoni, J. A., Operations man- African Fresh Fruit Export Supply
agement of a packaging plant in the Chain, M.Sc. Thesis, 2005.
fruit industry, J. Food Eng. 70 (2005), 9 Federal Investment Council, BNRHV
pp. 299–307. Fruit Industry Overview, Giacinti,
7 Masini, G., Petracci, N., Bandoni, M. A. (ed.), Neuquén, Argentina,
A., Supply chain planning in the fruit 2000.
industry. Presented at FOCAPO 2003, 10 Brooke, A., Kendrick, D., Meeraus,
Coral Springs, Miami, USA, 2003. A., GAMS-A User’s Guide, GAMS De-
velopment Corporation, Washington,
DC, 1996.

Appendix

Nomenclature
Abreviations

Acronyms
CJPs Concentrated Juice Plants
CPs Cider Plants
CSFs Cold Storage Facilities
FRSs Fruit Reception Sites
MPs Milling Plants
PPs Packaging Plants
TPCs Third Party Customers
TPFs Third Party Farms
TPSs Third Party Suppliers

Indices
C Packed fruit clients
Cc Cider clients
Cj Concentrated juice clients
Cp Cider plants
F Farms
J Concentrated juice plants
M Milling plants
P Packaging plants
R Fruit reception sites
S Fresh fruit cold storage facilities
Sj Concentrated juice cold storage facilities
T Time period
V Fruit varieties
W Concentrated juice varieties
 Abreviations 157

Sets
BHPv Beginning of harvesting period of fruit variety
BHPw Beginning of harvesting period of fruit variety for juice production
DSc Delivery schedule to packed fruit client
DScc Delivery schedule to cider client
DScj Delivery schedule to concentrated juice client
HPv Harvesting period of fruit variety
OS Own storage
TPS Third party storage
VWw Fruit variety assigned to concentrated juice variety

Variables

Mass Flow Variables

X1f,v,t Fresh fruit from farms to TPCs
X2f,p,v,t Fresh fruit from farms to PPs
X3p,v,t Third party farms fresh fruit supply to PPs
X4p,v,t TPS fresh fruit to PPs
X5p,v,t Total fresh fruit to PPs
X6p,v,c,t Processed fruit from PPs
X7p,w,t Waste fruit from PPs
X8s,v,c,t Packed fruit from TPSs
X9p,s,v,c,t Packed fruit from PPs and TPSs to PFSs
X10p,v,c,t Packed fruit from PPs to clients
X11p,v,c,t Packed fruit from PPs to TPCs
X12p,r,v,c,t Packed fruit from PPs to FRSs
X13p,m,v,c,t Packed fruit from PPs to MPs
X14p,cp,v,c,t Packed fruit from PPs to CPs
X15p,r,w,t Waste fruit from PPs to FRSs
X16p,m,w,t Waste fruit from PPs to MPs
X17p,cp,w1,t Waste fruit from PPs to CPs
X18f,s,v,t Fresh fruit from farms to CSFs
X19s,v,t Fresh fruit from third party farms to CSFs
X20s,v,t Fresh fruit from TPSs to CSFs
X21s,v,t Total fresh fruit to CSFs
X22s,v,t Cold fresh fruit from CSFs
X23s,p,v,t Cold fresh fruit from CSFs to PPs
X24s,r,v,t Cold fresh fruit from CSFs to FRSs
X25s,v,t Cold fresh fruit from CSFs to TPCs
X26s,m,v,t Cold fresh fruit from CSFs to MPs
X27s,cp,v,t Cold fresh fruit from CSFs to CPs
X28s,v,c,t Cold-packed fruit from CSFs
X29s,v,c,t Cold-packed fruit from CSFs to clients
X30s,v,c,t Cold-packed fruit from CSFs to TPCs
158 5 Supply Chain Tactical Optimization in the Fruit Industry

X31s,r,v,c,t Cold-processed fruit from CSFs to FRSs

X32s,m,v,c,t Cold-processed fruit from CSFs to MPs
X33s,cp,v,c,t Cold-processed fruit from CSFs to CPs
X34v,c,t Packed fruit to clients
X35f,r,v,t Fresh fruit from farms to FRSs
X36f,r,v,t Waste fruit from farms to FRSs
X37r,v,t Fresh fruit from TPFs to FRSs
X38r,w,t Waste fruit from TPFs to FRSs
X39r,v,t Fresh fruit from TPSs to FRSs
X39r,w,t Waste fruit from TPSs to FRSs
X40r,w,t Total fruit to FRSs
X41r,w,t Total fruit from FRSs
X42r,j,w,t Fruit from FRSs to CJPs
X43r,m,w,t Fruit from FRSs to MPs
X44r,cp,w1,t Fruit from FRSs to CPs
X45r,w,t Fruit from FRSs to TPCs
X46f,m,v,t Fresh fruit from farms to MPs
X47f,m,v,t Waste fruit from farms to MPs
X4848m,v,t Fresh fruit from TPFs to MPs
X49m,w,t Waste fruit from TPFs to MPs
X50m,v,t Fresh fruit from TPSs to MPs
X50m,w,t Waste fruit from TPSs to MPs
X51m,w,t Total fruit to MPs
X52m,w,t Total milled fruit from MPs
X53m,j,w,t Milled fruit from MPs to CJPs
X54m,w,t Milled fruit from MPs to TPCs
X55j,w,t Milled fruit from TPSs
X56j,w,t Concentrated juice from CJPs
X57sj,w,t Concentrated juice from TPSs to CJSs
X58j,sj,w,t Concentrated juice from CJP to CJS
X59j,w,t Concentrated juice from CJP to TPC
X60j,w,cj,t Concentrated juice from CJP to clients
X61sj,w,t Total concentrated juice from CJSs
X62sj,w,cj,t Concentrated juice from CJSs to clients
X63sj,w,t Concentrated juice from CJSs to TPCs
X64w,cj,t Total concentrated juice to clients
X65f,cp,v,t Fresh fruit from farms to CPs
X66f,cp,v  ,t Waste fruit from farms to CPs
X67cp,v,t Fresh fruit from TPFs to CPs
X68cp,w1,t Waste fruit from TPFs to CPs
X69cp,v,t Fresh fruit from TPSs to CPs
X69cp,w1,t Waste fruit from TPSs to CPs
X70cp,t Total fruit to CPs
X71cp,t Total cider from CPs
X72cp,cc,t Cider from CPs to clients
 Abreviations 159

X73cp,t Cider from CPs to TPCs

X74cp,t Cider from TPSs

Inventory Variables
CJSsj,w,t Concentrated juice at CJSs
FRSr,w,t Total fruit at FRSs
MPm,w,t Total milled fruit at MPs
CScp,t Total cider at CPs
NPFSs,v,t Nonprocessed fruit at CSFs
PFSs,v,c,t Packed fruit at CSFs
TSs,t Total fruit at CSFs

Cost Variables

Raw Material
CJPurchCt Concentrated juice purchase costs
CPurchCt TPSs cider purchase costs
OFFFProdCt Production costs at farms
PFPurchCt Packed fruit purchase costs
TFPPCt Total final products purchase costs
TPFFFPurchCt TPFs fresh fruit purchase costs
TPFWFPurchCt TPFs waste fruit purchase costs
TPRSFFPurchCt TPSs fresh fruit purchase costs
TPRSWFPurchCt TPSs waste fruit purchase costs
TPRSMFPurchCt TPSs milled fruit purchase costs
TRMCt Total raw material costs

Transportation
TCfCJPtCJSt From CJPs to CSCs
TCfCJPtCt From CJPs to clients
TCfCJStCt From CJSs to clients
TCfCPtCt From CPs to clients
TCfCSCtCPt From CSFs to CPs
TCfCSCtFRSt From CSFs to FRSs
TCfCSCtMPt From CSFs to MPs
TCfFRStCJPt From FRSs to CJPs
TCfFRStCPt From FRSs to CPs
TCfFRStMPt From FRSs to MPs
TCfFtCPt From farms to CPs
TCfFtFRSt From farms to FRSs
TCfFtMPt From farms to MPs
TCfFtNPFSt From farms to CSFs
TCfFtPPt From farms to PPs
160 5 Supply Chain Tactical Optimization in the Fruit Industry

TCfMPtCJPt From MPs to CJPs

TCfNPFStPPt From CSFs to PPs
TCfPFStCt From CSFs to clients
TCfPPtCPt From PPs to CPs
TCfPPtCt From PPs to clients
TCfPPtFRSt From PPs to FRSs
TCfPPtMPt From PPs to MPs
TCfPPtPFSt From PPs to PFSs
TTCt Total transportation costs
TPCSCt Third party cold storage costs

Operating
OCCJPFFt CJPs (fresh fruit)
OCCJPMFt CJPs (milled fruit)
OCCJPt Total CJP
OCCPt CPs
OCFRSt FRSs
OCMPt MPs
OCPPt PPs
OCSCt Own CSFs
OMPSOCt MPs storage
OCPSOCt CPs storage
TOCt Total Operating costs

Income

Sales
SICCt Cider to clients
SICJCt Concentrated juice to clients
SICJTPCt Concentrated juice to TPCs
SICTPCt Cider to TPCs
SIFFt Fresh fruit to TPCs
SIMFt Milled fruit to TPCs
SIPFCt Packed fruit to clients
SIPFTPCt Packed fruit to TPCs
SIWFt Waste fruit to TPCs
TSIt Total sales income
CSCRIt Own CSF rental
TotalProfit Objective function

Parameters
APf,v Production average per farm and variety
Aycv,c Fraction average per variety and client of packed fruit
 Abreviations 161

Aywff,v Waste average in farms per farm and variety

Aywv Waste average in PPs
ccpcc Cider specific price per client
cjcpw,cj Concentrated juice specific price per client
cjpucw Concentrated juice purchase specific cost
cjtpcpw Concentrated juice TPCs specific price
ocsprs Cold storage specific rental price
dcp,cc Distances between CPs and clients
df,cp Distances between farms and CPs
df,m Distances between farms and MPs
df,p Distances between farms and PPs
df,r Distances between farms and FRSs
df,s Distances between farms and CSFs
dj,cj Distances between CJPs and juice clients
dj,sj Distances between CJPs and CJSs
dm,j Distances between MPs and CJPs
dp,c Distances between PPs and clients
dp,cp Distances between PPs and CPs
dp,m Distances between PPs and MPs
dp,r Distances between PPs and FRSs
dp,s Distances between PPs and CSFs
dr,cp Distances between FRSs and CPs
dr,j Distances between FRSs and CJPs
dr,m Distances between FRSs and MPs
ds,c Distances between CSFs and clients
ds,cp Distances between CSFs and CPs
ds,m Distances between CSFs and MPs
ds,r Distances between CSFs and FRSs
dsj,cj Distances between CJSs and juice clients
ffprcf,v Fresh fruit production cost in farms per farm per variety
ffpv Fresh fruit selling specific price
MaxCJTPJSw Maximum availability of concentrated juice from TPSs
MaxCTPCS Maximum availability of cider from TPSs
MaxFFTPFv,t Maximum availability of fresh fruit from TPFs
MaxFFTPRSv Maximum availability of fresh fruit from TPSs
MaxMFTPMSw Maximum availability of milled fruit from TPSs
MaxPFTPPSv Maximum availability of packed fruit from TPSs
MaxProcCapCJPj Maximum processing capacity of CJPs
MaxProcCapCPcp Maximum processing capacity of CPs
MaxProcCapCSCs Maximum processing capacity of CSCs
MaxProcCapFRSr Maximum processing capacity of FRSs
MaxProcCapMPm Maximum processing capacity of MPs
MaxProcCapPPp Maximum processing capacity of PPs
MaxStoreCapCJSsj Maximum storage capacity of CJSs
MaxStoreCapCScp Maximum storage capacity of CPs
162 5 Supply Chain Tactical Optimization in the Fruit Industry

MaxStoreCapCSCs Maximum storage capacity of CSFs

MaxStoreCapFRSr Maximum storage capacity of FRSs
MaxStoreCapMPm Maximum storage capacity of MPs
MaxWFTPFw,t Maximum availability of waste fruit from TPFs
MaxWFTPRSw Maximum availability of waste fruit from TPSs
mfpw Milled fruit specific selling price
Mftc Milled fruit specific transportation cost
MinProcCapPPp Minimum processing capacity of PPs
Nrtc Nonrefrigerated specific transportation cost
oocccjpffj Specific operating cost of CJPs (fruit)
occjpfmj Specific operating cost of CJPs (pulp)
oocccpcp Specific operating cost of CPs
ooccfrsr Specific operating cost of FRSs
ooccmpm Specific operating cost of MPs
ooccppp Specific operating cost PPs
ooccscs Specific operating cost of own CSFs
ooccjjscsj Specific operating cost of own juice CSFs
ompsocm Specific MPs storage operating cost
ocpsoccp Specific CPs storage operating cost
Pf,v,t Fresh fruit production in farms
pfcpv,c Specific price of packed fruit per client
pfpucv,c Specific purchase cost of packed fruit
pftpcpv,c Specific cost of packed fruit to TPCs
Rtc Specific refrigerated transportation cost
SDPf,v Standard deviation of production per farm per variety
SDycv,c Standard deviation of fraction of product per variety per client
SDywff,v Standard deviation of waste in production per farm per variety
SDywv Standard deviation of waste production in PPs per variety
Tpccp Specific selling price of cider to TPCs
tpcscpuc Specific purchase price of cider from TPSs
ttppcscs Specific purchase price of third party CSFs rental
tpfffpucv Specific purchase price of fresh fruit from TPFs
tpfwfpucw Specific purchase price of waste fruit from TPSs
tprsffpucv Specific purchase price of fresh fruit from TPSs
tprsmfpucw Specific purchase price of milled fruit from TPSs
tprswfpucw Specific purchase price of waste fruit from TPSs
wfpw Specific selling price of waste fruit
ycjw,j Conversion factor of fruit or milled fruit into juice per juice variety
ycv,c,t Fraction of fresh fruit per variety for each client
ywff,v,t Fraction of waste fruit production in farms
ywv,t Fraction of waste fruit production in PPs
 Table Data Collection 163

Table Data Collection

Table 5.1 Fresh fruit harvest and quality data.

V Variety HPv Aywv (%) SDywv (%) Aycv, c (%) SDycv, c (%)
C1 C2 C1 C2

V1 Williams pear 2–5 30 3 41 27 2.1 2

V2 Beurre D´Anjou pear 6–10 25 3 32 34 2.3 2.1
V3 Beurre Bosc pear 10–14 30 3.3 28 29 3.2 2.1
V4 Improved red apple 11–15 32 4.8 31 45 3.4 2.3
V5 Packams Triumph pear 13–16 36 5.8 37 41 4.1 2.6
V6 Red delicious apple 14–17 31 4 42 23 2.4 3.1
V7 Red apple 18–21 29 2.9 44 29 3.2 4.1
V8 Granny Smith apple 19–22 24 2.2 33 36 1.8 2.8

Table 5.2 Farm production data (APf, v and SDPf, v ) (105 kg/week).

V1 V2 V3 V4 V5 V6 V7 V8
A SD A SD A SD A SD A SD A SD A SD A SD

F1 1.71 0.07 1.76 0.08 1.16 0.04 1.76 0.04 0.52 0.02 1.27 0.09 0.41 0.02 0.73 0.04
F2 0.33 0.02 0.44 0.01 0.92 0.05 1.78 0.09 1.28 0.05 0.39 0.01 1.73 0.11 0.88 0.06
F3 1.28 0.05 1.42 0.05 1.20 0.05 1.18 0.05 0.91 0.05 1.87 0.11 1.68 0.04 0.62 0.01
F4 0.73 0.02 0.96 0.02 1.08 0.04 0.97 0.04 1.52 0.07 1.63 0.05 1.78 0.10 1.20 0.06
F5 0.96 0.04 1.20 0.04 1.54 0.09 1.68 0.07 0.90 0.02 1.20 0.03 1.87 0.04 0.96 0.03

Table 5.3 Waste fraction from farms (Aywff,v and SDywff,v ) (%).

V1 V2 V3 V4 V5 V6 V7 V8
A SD A SD A SD A SD A SD A SD A SD A SD

F1 18 1.7 19 1.0 34 3.1 23 1.9 22 2.2 32 3.1 33 2.0 16 1.5

F2 26 2.5 15 1.2 30 1.6 17 1.3 32 2.9 26 2.6 16 1.2 16 1.3
F3 32 2.4 31 2.9 22 1.6 27 2.0 17 1.4 26 1.5 16 1.4 17 1.5
F4 33 2.0 33 2.7 27 1.6 30 2.0 26 1.4 18 1.5 30 2.6 24 1.3
F5 30 2.4 20 1.2 32 2.2 29 1.7 34 2.1 28 1.5 18 1.3 15 1.5
164 5 Supply Chain Tactical Optimization in the Fruit Industry
Table 5.4 Fresh fruit production cost in farms ($/kg).

ffprcf , v V1 V2 V3 V4 V5 V6 V7 V8

F1 0.205 0.214 0.218 0.223 0.209 0.242 0.269 0.253

F2 0.287 0.24 0.29 0.273 0.264 0.237 0.293 0.274
F3 0.288 0.277 0.23 0.231 0.21 0.298 0.276 0.255
F4 0.287 0.238 0.207 0.285 0.253 0.288 0.228 0.28
F5 0.228 0.253 0.207 0.253 0.251 0.255 0.268 0.234

Table 5.5 Fresh fruit prices ($/kg).

V1 V2 V3 V4 V5 V6 V7 V8

TPF purchase tpfffpucv

0.35 0.35 0.35 0.35 0.35 0.35 0.35 0.35
TPS purchase tprsffpucv
0.40 0.37 0.35 0.36 0.35 0.35 0.36 0.35
TPC selling ffpv
0.27 0.28 0.25 0.29 0.31 0.29 0.33 0.27

Table 5.6 Waste and milled fruit prices ($/kg).

w1 w2

Waste TPF purchase tpfwfpucw 0.072 0.07

Waste TPS purchase tprswfpucw 0.095 0.086
Waste selling wfpw 0.095 0.086
Milled fruit selling mfpw 0.038 0.044

Table 5.7 Concentrated juice ($/gallon).

Clients selling price cjcpw, cj w1 w2

cj1 5.5 5.45
TPS purchase price cjpucw 5.5 5.45
TPC selling price cjtpcpw 5.5 5.45

Table 5.8 Cider ($/l).

Clients selling price TPC selling price TPS purchase price

ccpcc tpccp tpcscpuc
cc1
1.4 1.4 1.28
 Table Data Collection 165

Table 5.9 Packed fruit selling price to clients ($/kg).

pfcpv, c V1 V2 V3 V4 V5 V6 V7 V8

c1 1.45 1.49 1.61 1.50 1.63 1.56 1.62 1.52

c2 1.31 1.34 1.45 1.35 1.47 1.41 1.46 1.37
c3 1.24 1.27 1.37 1.28 1.39 1.33 1.37 1.29

Table 5.10 Packed fruit selling price to TPCs ($/kg).

pftpcpv, c V1 V2 V3 V4 V5 V6 V7 V8

c1 1.70 1.54 1.31 1.20 1.76 1.04 0.91 1.29

c2 1.08 1.30 1.45 0.95 1.41 1.10 0.83 1.12
c3 0.87 1.05 1.09 0.76 0.96 0.90 0.65 1.23

Table 5.11 Packed fruit purchase price ($/kg).

pfpucv, c V1 V2 V3 V4 V5 V6 V7 V8

c1 1.70 1.54 1.31 1.20 1.64 1.04 0.91 1.29

c2 1.08 1.30 1.45 0.95 1.41 1.10 0.83 1.12
c3 0.87 1.05 1.09 0.76 0.96 0.90 0.85 1.23

Table 5.12 Transportation costs ($/km/tonnes).

Milled fruit mftc 0.12

Nonrefrigerated nrtc 0.15
Refrigerated rtc 0.212

Table 5.13 Cold storage.

Own Third party

s1 s2 s3 s4 sj1 sj2 s5

Operating cost Operating cost Rental cost

($/Tonnes-week) ($/Gallon-week) ($/Tonnes-week)
ooccscs ooccjjscsj ttppcscs
0.84 0.841 0.839 0.84 1.367e-3 1.36e-3 1.26
Rental price
ocsprs
1.00 1.00 1.00 1.00
166 5 Supply Chain Tactical Optimization in the Fruit Industry
Table 5.14 CJPs operating cost ($/Tonnes).

j1 j2

Operating cost
oocccjpffj
90.00 91.00
occjpfmj
65.00 66.00

Table 5.15 CPs operating and storage cost ($/Tonnes).

cp1 cp2

oocccpcp
107.70 107.90
ocpsoccp
8.00 8.10

Table 5.16 FRSs operating cost ($/Tonnes).

ooccfrsr
r1 r2 r3

4.30 4.20 4.10

Table 5.17 MPs operating and storage cost ($/Tonnes).

m1 m2

ooccmpm
8.20 8.50
ompsocm
3.20 2.50

Table 5.18 PPs operating cost ($/kg).

ooccppp
p1 p2

0.50 0.49
 Table Data Collection 167

Table 5.19 Distances between FRSs and CPs (km).

d r,cp r1 r2 r3

cp1 117 83 65
cp2 64 30 7

Table 5.20 Distances between FRSs and CJPs (km).

d r,j r1 r2 r3

j1 97 63 43
j2 129 95 78

Table 5.21 Distances between FRSs and MPs (km).

d r,m r1 r2 r3

m1 118 84 66
m2 195 162 147

Table 5.22 Distances between CPs and clients (km).

d cp,cc cp1 cp2

cc1 65 7

Table 5.23 Distances between farms and FRSs (km).

d f ,r r1 r2 r3

F1 9 44 67
F2 10 24 48
F3 30 5 29
F4 21 5 25
F5 35 11 13

Table 5.24 Distances between farms and CPs (km).

d f ,cp F1 F2 F3 F4 F5

cp1 126 107 88 66 54

cp2 73 54 34 9 5
168 5 Supply Chain Tactical Optimization in the Fruit Industry
Table 5.25 Distances between farms and MPs (km).

d f ,m F1 F2 F3 F4 F5

m1 128 108 89 67 55
m2 204 185 167 147 137

Table 5.26 Distances between farms and PPs (km).

d f ,p F1 F2 F3 F4 F5

p1 55 36 17 9 23
p2 89 70 50 26 13

Table 5.27 Distances between farms and CSFs (km).

d f ,s s1 s2 s3 s4 s5

F1 55 89 138 248 363

F2 36 70 119 230 345
F3 17 50 100 212 326
F4 9 26 79 192 304
F5 23 13 68 182 292

Table 5.28 Distances between CJPs and juice clients (km).

d j ,cj j1 j2

cj1 264 235

Table 5.29 Distances between CJPs and juice CJSFs (km).

d j ,sj sj1 sj2

j1 0 34
j2 34 0

Table 5.30 Distances between MPs and CJPs (km).

d m,j m1 m2

j1 22 104
j2 13 70
 Table Data Collection 169

Table 5.31 Distances between PPs and FRSs (km).

d p,r r1 r2 r3

p1 46 12 13
p2 80 45 24

Table 5.32 Distances between PPs and clients (km).

d p,c c1 c2 c3

p1 316 954 2822

p2 282 957 2828

Table 5.33 Distances between PPs and CPs (km).

d p,cp cp1 cp2

p1 72 18
p2 41 17

Table 5.34 Distances between PPs and MPs (km).

d p,m m1 m2

p1 74 153
p2 42 124

Table 5.35 Distances between CSCs and FRSs (km).

d s,r s1 s2 s3 s4 s5

r1 46 80 129 239 354

r2 12 45 95 207 321
r3 13 24 78 192 303

Table 5.36 Distances between CSCs and clients (km).

d s,c s1 s2 s3 s4 s5

c1 81 282 235 158 316

c2 922 957 934 889 954
c3 2795 2828 2810 2769 2822
170 5 Supply Chain Tactical Optimization in the Fruit Industry
Table 5.37 Distances between CSCs and CPs (km).

d s,cp s1 s2 s3 s4 s5

cp1 72 41 13 128 238

cp2 18 17 71 186 296

Table 5.38 Distances between CSCs and MPs (km).

d s,m s1 s2 s3 s4 s5

m1 74 42 13 127 237
m2 153 124 70 45 161

Table 5.39 Distances between PPs and CSCs (km).

d p,s s1 s2 s3 s4 s5

p1 0 34 85 198 311
p2 34 0 54 169 279

Table 5.40 Distances between CJCSFs and juice clients (km).

d sj ,cj sj1 sj2

cj1 264 235

Table 5.41 CPs maximum processing capacity (Tonnes/week).

MaxProcCapCPcp
cp1 cp2

52 27

Table 5.42 CPJs maximum processing capacity (Tonnes/week).

MaxProcCapCJPj
j1 j2

220 190
 Table Data Collection 171

Table 5.43 PPs maximum processing capacity (Tonnes/week).

MaxProcCapPPp
p1 p2

720 690

Table 5.44 Maximum storage capacity of CJCSFs (103 gallon).

MaxProcCapCJSij
sj1 sj2

829 795

Table 5.45 Maximum storage capacity of CSFs (106 kg).

MaxStoreCapCSCs
s1 s2 s3 s4 s5

1.0 1.0 1.0 1.0 25.0

Table 5.46 Maximum storage capacity of FRSs (106 kg).

MaxStoreCapFRSr
s1 s2 s3

2.23 2.38 1.80

Table 5.47 Maximum storage capacity of MPs (Tones).

MaxStoreCapMPm
m1 m2

500 500

Table 5.48 Conversion factors of fruit into juice per juice variety (gallon/kg).

ycjw, j j1 j2

w1 0.0323 0.0317
w2 0.0313 0.0303
172 5 Supply Chain Tactical Optimization in the Fruit Industry
Table 5.49 Conversion factors of apple into cider (l/kg).

ycjcp

cp1 0.74
cp2 0.76

Table 5.50 Delivery dates.

Delivery number
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Packed fruit
c1 2 5 6 8 10 12 15 17 18 19 21 22 23 24 26
c2 2 10 15 20 25 30 35 40 45 50 52
c3 1 4 8 11 15 18 22 25 29 32 36 39 43 47 51
Concentrate juice
cj1 4 15 20 25 30 35 40 50
Cider
Cc1 2 5 12 16 21 24 28 32 38 44 48 52

Table 5.51 Main business economic results.

Annual Total sale Total Total raw Total Third party

net profit income operating material cost transportation cold storage
(US$) (US$) cost (US$) (US$) cost (US$) cost (US$)

4,384,291 47,204,180 23,163,080 10,253,900 9,137,581 265,328

 173

6
Short-Term Scheduling of Batch and Continuous Processes
Munawar A. Shaik, Christodoulos A. Floudas

In this chapter, an overview of the exciting developments in scheduling of multi-

product, multipurpose batch, and continuous processes is presented. Scheduling
is a decision-making process to determine when, where, how to produce a set of
products given requirements in a specific time horizon, a set of limited resources,
and processing recipes. The research area of batch and continuous process schedul-
ing has received great attention from both academia and industry in the past two
decades. This is motivated by the increasing pressure to improve efficiency and
reduce costs and by the significant advances in relevant modeling and solution
techniques and the rapidly growing computational power. In multiproduct and
multipurpose batch, semicontinuous, and continuous plants, different products
are manufactured via the same or similar sequence of operations by sharing avail-
able pieces of equipment, intermediate materials, and other production resources.
The inherent operational flexibility of these plants provides the platform for great
savings reflected in good production schedules.

6.1
Classification of Scheduling Formulations

6.1.1
Time Representation

One of the key issues concerning process scheduling problems is the time repre-
sentation. Broadly, there are two approaches: discrete-time models and continuous-
time models. Early efforts in modeling the process scheduling problems relied on
the discrete-time approach, in which the time horizon is uniformly divided into
equal time intervals and events such as the beginning and ending of a task are at-
tached with the boundaries of these time intervals. For a precise approximation of
the original problem, it requires the usage of a huge number of small time inter-
vals leading to very large combinatorial problems of intractable size for real world
problems. Due to these limitations, continuous-time models are widely used in

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
174 6 Short-Term Scheduling of Batch and Continuous Processes

Fig. 6.1 Different time representations.

the literature, in which the events can take place at any point in the continuous-
time domain, resulting in mathematical models of smaller size and requiring less
computational effort for their solution. Recently, Floudas and Lin [1, 2] presented
state-of-the-art reviews comparing various discrete- and continuous-time-based for-
mulations.
The different continuous-time models can be broadly classified into three dis-
tinct categories: slot-based, global event-based, and unit-specific event-based for-
mulations. One of the first methods used to formulate continuous-time models
is based on the concept of time slots [3–10], in which the time horizon is repre-
sented as ordered blocks of unknown, variable lengths, or slots. Global event-based
models [11–28] use a set of events that are common across all units and the events
are defined for either the beginning or end (or both) of each task in each unit.
Unit-specific event-based models [29–37] define events on a unit basis, allowing
tasks corresponding to the same event but in different units to take place at dif-
ferent times. The different time representations [38] are summarized in Fig. 6.1.
The uniform-time discretization is depicted in part (a). The different variations of
the continuous-time representation are illustrated in parts (b)–(d). For the specific
instance of the four tasks considered on three units in parts (b)–(d) of Fig. 6.1,
the slot-based representation requires five slots (or six events), the global event-
based representation requires four events, while the unit-specific event-based rep-
resentation requires consideration of only two events. A detailed comparison of
 6.1 Classification of Scheduling Formulations 175

various continuous-time models for short-term scheduling of batch plants was per-
formed recently by Shaik et al. [38]. They concluded that, due to heterogeneous
locations of event points used, the unit-specific event-based models always require
less event points and exhibit favorable computational performance compared to
both slot-based and global event-based models. A detailed examination of the ex-
isting continuous-time formulations for batch and continuous process scheduling
problems is provided subsequently.

6.1.2
Characteristics of Process Scheduling Problems

Although most process scheduling problems have certain common features such
as equipment-task assignment, sequencing, and timing of activities, different prob-
lems may also vary considerably in the following aspects, which represent different
provisions or extent of difficulty for the modeling of these processes.

6.1.2.1 Processing Sequences

Based on the complexity of the processing sequences used for production, the
multiproduct/multipurpose plants can be classified as either (i) sequential or (ii)
network-oriented. In sequential processes, different products are produced in sin-
gle or multistages with either only one unit or parallel units per stage and same
processing sequence is followed for each product. In this type of processes, no
explicit mass balances are required and batches are used to represent produc-
tion. Processing networks such as state-task network (STN), resource-task network
(RTN), and recipe diagrams are used to represent more complex production recipes
in which different products follow different production routes. In these types of
processes, explicit mass balances are required because batches can merge and/or
split. In the STN process representation [39], raw materials, intermediates, and
final products are represented as states denoted by circles, and an operation is rep-
resented as a task denoted by rectangle box. For example, Fig. 6.2 shows the STN
representation of a widely studied process involving merging/splitting of materials
and recycle.
In the RTN process representation [40], processing and storage equipment and
utilities are uniformly represented as resources. The RTN representation of the
same process as in the STN example is given in Fig. 6.3, in which the processing
equipment, denoted by ellipses, are also additionally shown. Here, tasks taking
place in different units are considered as different tasks.
176 6 Short-Term Scheduling of Batch and Continuous Processes

Fig. 6.2 Example of state-task network.

S8 Product 1
Heater
J1
Product 2
Feed A Hot A Reaction 2 S9
Heating
S4 0.4 i=4 0.4
S1
i=1 Reaction 2 0.6
0.6 0.9
i=5 IntAB
0.1 Separation Separator
S5
Reactor1 i=8 J4
J2

Feed B IntBC Reaction 3

Reaction 1 0.8 i=6
S2 0.5 i=2 S6 S7
Reaction 1 Reaction 3
0.5 i=7
i=3 Reactor2 Impure E
J3
0.2
Feed C S3

Fig. 6.3 Example of resource-task network.

6.1.2.2 Processing Modes of Operation

The processing tasks can be either (i) batch or (ii) continuous tasks. In a batch task,
raw material is provided at the start of the task and products are produced at the
end of the task after a pre-specified period of time. In a continuous task, materials
are fed and/or products are continuously withdrawn with a processing rate which
is either fixed or within a certain range.

6.1.2.3 Intermediate Storage Policies

The storage capacity available for an intermediate could be either finite, denoted as
(i) finite intermediate storage (FIS), or (ii) infinite, denoted as unlimited interme-
diate storage (UIS). If no storage space is available, denoted as (iii) no intermediate
storage (NIS), then the material produced can be held in the same batch processing
 6.2 Short-Term Scheduling of Batch Processes 177

unit until it is consumed by the next processing task. For certain unstable materi-
als which require immediate consumption, (iv) a zero-wait (ZW) policy needs to be
enforced, which requires special timings constraints.

6.1.2.4 Demand Patterns

There are two different ways of specifying demands of products. (i) The demands
are due at the end of the horizon under consideration typical to cyclic scheduling
problems, (ii) the demands are due at specified intermediate due dates within the
time horizon, which is a general case of short-term scheduling.

6.1.2.5 Resource Considerations

Utilities such as manpower, steam, cooling water, and electricity are considered
as renewable resources which are completely recoverable at the end of the task
that requires them. These resources have either (i) no restrictions or (ii) there is a
limitation on the maximum availability at any time during the production.

6.1.2.6 Changeovers
For two different tasks taking place in the same unit, the different possible
changeover requirements are: (i) no setup time required, (ii) changeover time is
required but is independent of the sequence in which the two consecutive tasks
take place, and (iii) sequence-dependent changeovers.

6.1.2.7 Performance Criteria

The different performance criteria or objectives for a process scheduling problem
are: (i) maximization profit, in which there is revenue from the sales of products,
(ii) minimization of costs, in which the overall cost of production, inventory, tran-
sitions, transportation, earliness, and tardiness is minimized. Another objective
often used in scheduling is (iii) minimization of makespan, which represents the
minimum completion time of the whole process for a specified production require-
ment.

6.2
Short-Term Scheduling of Batch Processes

The problem of short-term scheduling of multiproduct and multipurpose batch

plants has received significant attention and numerous formulations have been
proposed in the literature using different continuous-time representations. Mod-
els from each of the categories of slot-based, global event-based, and unit-specific
event-based continuous-time representations are discussed below, in chronological
order.
178 6 Short-Term Scheduling of Batch and Continuous Processes

6.2.1
Unit-Specific Event-Based Model of Ierapetritou and Floudas [29] (I&F)

The authors presented the original concept of event points which correspond to a
sequence of time instances located along the time axis of each unit, each represent-
ing the beginning of a task or the utilization of the unit. The location of event points
is different for each unit, allowing different tasks to start at different times in each
unit for the same event point. The timings of tasks are accounted through special
sequencing constraints involving big-M constraints. No resources other than ma-
terials and equipment are considered. Although the model originally claimed its
superiority due to both decoupling of task and unit events and nonuniform-time
grid, later it became evident that it is primarily the introduction of the unit-specific
events that gives the model the resulting cutting edge and makes it a class apart
from all other models proposed in the literature. The resulting model requires
less event points compared to the corresponding global event or slot-based mod-
els, thus yielding better computational results [38], although big-M constraints are
used. This model was later extended by Janak et al. [35, 36] allowing tasks to spread
over multiple events to accurately account for the utilization of different resources
and storage policies.
Giannelos and Georgiadis [37] also proposed an STN represented, unit-specific
event-based formulation for short-term scheduling of multipurpose batch plants.
This is a slight variation of the I&F model, wherein the authors relaxed the task du-
rations using buffer times and implicitly eliminated the various big-M constraints
of the I&F model. However, the authors introduced special duration and sequenc-
ing constraints that effectively transformed the nonuniform-time grid to a uniform
one (global events) for the purposes of material balance and storage constraints.
The start times (end times) of the tasks producing/consuming the same state were,
respectively, forced to be the same, leading to suboptimal solutions [10, 38].
An improved and modified version of the I&F model, as presented in the com-
parative study of Shaik et al. [38] is given below.
For the objective of maximization of profit:




Max Profit = prices ST (s, N) + ρsi b(i, j, N) , (6.1)
s i∈ρsi >0 j ∈suitij


w(i, j, n)
1, ∀j, n, (6.2)
i∈suitij

w(i, j, n)Bijmin
b(i, j, n)
w(i, j, n)Bijmax , ∀i, j ∈ suitij , ∀n, (6.3)

ST(s, n) = ST(s, n − 1) + ρsi b(i, j, n − 1)
i∈ρsi >0 j ∈suitij



+ ρsi b(i, j, n), ∀s, n, (6.4)
i∈ρsi <0 j ∈suitij
 6.2 Short-Term Scheduling of Batch Processes 179
     
ts(i, j, n + 1)  ts i  , j, n + αi  j w i  , j, n + βi  j b i  , j, n ,
∀i, i  , j ∈ suitij , suiti  j , ∀n < N, (6.5)
     
ts(i, j, n + 1)  ts i  , j  , n + αi  j  w i  , j  , n + βi  j  b i  , j  , n
 
− H 1 − w(i  , j  , n) , ∀s, i, i  , j, j  ∈ suitij , suiti  j  ,
i = i  , j = j  , ρsi < 0, ρsi  > 0, ∀n < N, (6.6)
ts(i, j, N) + αij w(i, j, N) + βij b(i, j, N)
H, ∀i, j ∈ suitij , (6.7)
ts(i, j, n)
H, ∀i, j ∈ suitij , ∀n, (6.8)
ST(s, n)
STmax
s , ∀s ∈ FIS, ∀n, (6.9)
w(i, j, n) = b(i, j, n) = ts(i, j, n) = 0, ∀i, j ∈ suitij = 0. (6.10)
For the objective of minimization of makespan:
Min MS, (6.11)



ST(s, N) + ρsi b(i, j, N)  Demands , (6.12)
i∈ρsi >0 j ∈suitij

ts(i, j, N) + αij w(i, j, N) + βij b(i, j, N)
MS, ∀i, j ∈ suitij . (6.13)

6.2.2
Global Event-Based Model of Castro and coworkers [21, 22]

Castro et al. [21] proposed a formulation using RTN representation for short-term
scheduling of batch plants. The time horizon is divided into several global events
that are common across all units. Binary variables are defined for assigning both
start and end times of different tasks to the corresponding global events. Because
of the unified treatment of various resources in the RTN framework, no special
sequencing constraints are required. All the balances are written in terms of a
single excess resource constraint, which implicitly includes the balances on the
status and batch amounts of each unit. This model has no big-M constraints ex-
cept for those that relate the extents of each task to the corresponding binary vari-
ables. Because of the provision for end times of tasks to be before the end times of
the corresponding time slots, the processing time of each task on a given unit is
not exactly represented but has an additional waiting period. Although the authors
claimed superiority over the STN based event-driven formulation of Ierapetritou
and Floudas [29], it was established later [31] that the claims were based on incor-
rect data obtained from rounding off the parameter values used. Later, Castro et
al. [22] (CBMN) proposed an improved model by eliminating some of the redun-
dant binary and continuous variables and proposed new timing constraints that
result in compact problem statistics and improved relaxed solutions. They com-
pared the results for two different models (MN and MO) with the new and old
timing constraints, respectively. It should be noted that, in the model of Castro et
al. [22], there is an additional parameter (t) that defines a limit on the maximum
180 6 Short-Term Scheduling of Batch and Continuous Processes

number of events over which a task can occur, and it has a significant impact on
the solution obtained, the computational time, and the problem statistics. At each
event point, we need to iterate over this parameter to get the global optimal solu-
tion. The MN model has fewer constraints and gives better LP relaxed solution over
the MO model and is given below.
For the objective of maximization of profit:


Max Profit = pricer R(r, t = |T |), (6.14)
r

 
T (t  ) − T (t)  µri αi N (i, t, t  ) + βi ξ (i, t, t  ) ,
i

∀r ∈ R EQ , t, t  , t < t 
t + t, t = |T |, (6.15)

Vimin N(i, t, t  )
ξ (i, t, t  )
Vimax N (i, t, t  ),

∀i, t, t  , t < t 
t + t, t = |T |, (6.16)

 


R(r, t) = Rr0 t=1 + R(r, t − 1)t>1 + µri N (i, t, t  )
i t<t 
t+t


 

+ vri ξ (i, t, t ) + µri N(i, t  , t) + vri ξ (i, t  , t) ,
i t−t
t  <t
∀r, t, (6.17)

Rrmin
R(r, t)
Rimax , ∀r, t, (6.18)

T (t)
H, ∀t, (6.19)
T (t) = 0, ∀t = 1, (6.20)
   
N(i, t, t ) = ξ (i, t, t ) = 0, ∀t = 1 or t
t or ∀t = |T |. (6.21)

For the objective of minimization of makespan:

Min MS, (6.22)

R(r, t)  Demandr , t = |T | , (6.23)
T (t)
MS, t = |T | . (6.24)

The model for makespan minimization is composed of constraints (6.15)–(6.18)

and (6.20)–(6.23). For zero-wait tasks, the following constraint needs to be added:



T (t  ) − T (t)
H 1 − µri N (i, t, t  ) + µri αi N (i, t, t  )
i∈I ZW i∈I ZW


+βi ξ (i, t, t ) ,
∀r ∈ R EQ , t, t  , t < t 
t + t, t = |T |. (6.25)
 6.2 Short-Term Scheduling of Batch Processes 181

6.2.3
Global Event-Based Model of Maravelias and Grossmann [28] (M&G)

This is a global event-based model using STN process representation. The model
accounts for resource constraints (other than equipment) such as utilities, vari-
ous storage policies (unlimited, finite, zero wait, and no intermediate storage), and
sequence-dependent changeover times and allows for batch mixing/splitting. This
model reduces to the case of no resources, and it was used as such for comparison
to other approaches [28]. Global event points are used that are common across all
units, and tasks are allowed to be processed over multiple events. Different binary
variables are used to denote if a task starts, or continues over multiple events, or if
it finishes processing a batch at a given event point. Also, a new class of tightening
inequalities is proposed for tightening the relaxed LP solutions. The following is
the model of Maravelias and Grossmann [28].
For the objective of maximization of profit:


Max Profit = prices ST (s, N), (6.26)
s


Ws(i, n)
1, ∀j, n, (6.27)
i∈suitij


Wf(i, n)
1, ∀j, n, (6.28)
i∈suitij



Ws(i, n) = Wf(i, n), ∀i, (6.29)
n n



(Ws(i, n ) − Wf(i, n ))
1, ∀j, n, (6.30)
i∈suitij n
n

D(i, n) = αi Ws(i, n) + βi Bs(i, n), ∀i, n, (6.31)

Tf(i, n)
Ts(i, n) + D(i, n) + H (1 − Ws(i, n)), ∀i, n, (6.32)
Tf(i, n)  Ts(i, n) + D(i, n) − H (1 − Ws(i, n)), ∀i, n, (6.33)
Tf(i, n) − Tf(i, n − 1)
H Ws(i, n), ∀i, n > 1, (6.34)
Tf(i, n) − Tf(i, n − 1)  D(i, n), ∀i, n > 1, (6.35)
Ts(i, n) = T (n), ∀i, n, (6.36)
Tf(i, n − 1)
T (n) + H (1 − Wf(i, n)), ∀i, n > 1, (6.37)
Ws(i, n)Bimin
Bs(i, n)
Ws(i, n)Bimax , ∀i, n, (6.38)
Wf(i, n)Bimin
Bf(i, n)
Wf(i, n)Bimax , ∀i, n, (6.39)
 



Bimin  Ws(i, n ) − Wf(i, n )
Bp(i, n)
n <n n ≤n
 




Bimax  Ws(i, n ) − Wf(i, n ) , ∀i, n, (6.40)
n <n n
n
182 6 Short-Term Scheduling of Batch and Continuous Processes

Bs(i, n − 1) + Bp(i, n − 1) = Bp(i, n) + Bf(i, n), ∀i, n > 1, (6.41)

B I (i, s, n) = ρsi Bs(i, n), ∀i, n, ∀s ∈ SI(i), (6.42)

I
B (i, s, n)
Bimax ρsi Ws(i, n), ∀i, n, ∀s ∈ SI(i), (6.43)

B O (i, s, n) = ρsi Bf(i, n), ∀i, n, ∀s ∈ SO(i), (6.44)

B O (i, s, n)
Bimax ρsi Wf(i, n), ∀i, n, ∀s ∈ SO(i), (6.45)

ST (s, n) = ST (s, n − 1) + B O (i, s, n) − B I (i, s, n),
i∈O(s) i∈I (s)
∀s, n > 1, (6.46)

T (n + 1)  T (n), ∀n < N, (6.47)

D(i, n)
H, ∀j, (6.48)
i∈suitij n



D(i, n )
H − T (n), ∀j, n, (6.49)
i∈suitij n n


 
αi Wf(i, n ) + βi Bf(i, n )
T (n), ∀j, n, (6.50)
i∈suitij n
n

Ts(i, n)
H, ∀i, n, (6.51)

Tf(i, n)
H, ∀i, n, (6.52)

ST(s, n)
STsmax , ∀s ∈ FIS, ∀n, (6.53)

O
T (n) = Wf(i, n) = Bf(i, n) = B (i, s, n) = 0, ∀n = 1, (6.54)

Ws(i, n) = Bs(i, n) = D(i, n) = Bp(i, n) = B I (i, s, n) = 0,

∀n = N, (6.55)

T (N) = H. (6.56)

For the objective of minimization of makespan:

Min MS, (6.57)

ST(s, N)  Demands , ∀s, (6.58)

T (N) = MS, (6.59)

D(i, n)
MS, ∀j, (6.60)
i∈suitij n



D(i, n )
MS − T (n), ∀j, n. (6.61)
i∈suitij n n
 6.2 Short-Term Scheduling of Batch Processes 183

The model for makespan minimization uses constraints (6.27)–(6.47), (6.50)–

(6.55), and (6.57)–(6.61). For zero-wait tasks, the following constraints are added:
Tf(i, n − 1)  T (n) − H (1 − Wf(i, n)), ∀i ∈ I ZW , n > 1, (6.62)
When utility requirements are considered, the following constraints are added:
R I (i, r, n) = γir Ws(i, n) + δir Bs(i, n), ∀i, r, n, (6.63)
R O (i, r, n) = γir Wf(i, n) + δir Bf(i, n), ∀i, r, n, (6.64)



R(r, n) = R(r, n − 1) − R O (i, r, n − 1) + R I (i, r, n), ∀r, n, (6.65)
i i

R(r, n)
Rrmax , ∀r, n. (6.66)

6.2.4
Slot-Based Model of Sundaramoorthy and Karimi [10] (S&K)

Among the various slot-based formulations proposed in the literature, this is a re-
cent model available for the short-term scheduling of multipurpose batch plants.
The authors claim superior performance for the models they compared with, in-
cluding those based on global events and that of Giannelos and Georgiadis [37].
They use generalized recipe diagrams for process representation, wherein a stor-
age task is used to model the mixing and splitting of the same material streams.
No resources other than materials and equipment are considered, and transfer and
setup times are lumped into the batch processing times of tasks. The time horizon
is divided into multiple time slots of varying lengths, and tasks are allowed to con-
tinue processing over multiple time slots. For each unit, binary variables are used
to assign the beginning of each task to various time slots, and [0, 1] continuous
variables are used to denote tasks that continue over multiple slots and to denote
tasks that release their batch amount at the end of a slot. An additional zero task
is defined for modeling idling of units and to occupy extra redundant slots. Even
though this model is categorized as slot-based, tasks are allowed to finish before the
end of the time slot, making the model inherently similar to the global event-based
models, except for the differences in accounting the various balances. Several bal-
ances are proposed based on status of each unit, material and storage constraints,
and a new way of writing the balance on the time remaining on each unit, leading
to a compact model. Some of the examples reported in their paper are not solved
to zero integrality gap. The readers can easily verify that there are typographical
mistakes in the balances for the batch amount in a unit (constraints (11) and (12)
of the original paper) which, if corrected, are similar to big-M constraints. Except
for these constraints, the resulting model has no other big-M constraints. The fol-
lowing is the model of Sundaramoorthy and Karimi [10]. Here, the set of tasks (I)
also includes an idle task ‘i0’ that is suitable on all units.
For the objective of maximization of profit:


Max Profit = prices ST(s, K), (6.67)
s
184 6 Short-Term Scheduling of Batch and Continuous Processes


SL(k)
H, (6.68)
k

Z(j, k) = Y (i, j, k), ∀j, 0
k < K, (6.69)
i∈suitij

Y (i, j, k)Bijmin
B(i, j, k)
Y (i, j, k)Bijmax ,

∀i > 0, j ∈ suitij , 0
k < K, (6.70)

y(i, j, k) = y(i, j, k − 1) + Y (i, j, k − 1) − YE(i, j, k),

∀i, j ∈ suitij , 0 < k < K, (6.71)


Z(j, k) = YE(i, j, k), ∀j, 0 < k < K, (6.72)
i∈suitij

 
t (j, k + 1)  t (j, k) + αij Y (i, j, k) + βij B(i, j, k) − SL(k + 1),
i∈suitij
∀j, k < K, (6.73)

b(i, j, k) = b(i, j, k − 1) + B(i, j, k − 1) − BE(i, j, k),

∀i > 0, j ∈ suitij , k > 0, (6.74)

y(i, j, k)Bijmin
b(i, j, k)
y(i, j, k)Bijmax ,

∀i > 0, j ∈ suitij , 0 < k < K, (6.75)

YE(i, j, k)Bijmin
BE(i, j, k)
YE(i, j, k)Bijmax ,

∀i > 0, j ∈ suitij , 0 < k
K, (6.76)


t (j, k)
αij y(i, j, k) + βij b(i, j, k), ∀j, 0 < k < K, (6.77)
i∈suitij

ST(s, k) = ST(s, k − 1) + ρsi BE(i, j, k)
j i∈suitij ,i=0,ρsi >0



+ ρsi B(i, j, k), ∀s, k, (6.78)
j i∈suitij ,i=0,ρsi <0

ST(s, k)
STmax
s , ∀s ∈ FIS, ∀k, (6.79)

SL(k)
max max (αij + βij Bijmax ) , ∀k > 0, (6.80)

j i∈suitij

t (j, k)
max (αij + βij Bijmax ), ∀j, k, (6.81)

i∈suitij
 6.2 Short-Term Scheduling of Batch Processes 185

Y (i, j, k) = y(i, j, k) = b(i, j, k) = B(i, j, k) = 0,

∀i, j ∈ suitij = 0 or k = K, (6.82)

YE(i, j, k) = y(i, j, k) = b(i, j, k) = BE(i, j, k) = 0,

∀i, j ∈ suitij = 0 or k = 0, (6.83)

Z(j, k) = t (j, k) = 0, ∀j, k = K, (6.84)

t (j, k) = 0; SL(k) = 0, ∀k = 0, (6.85)

0
y(i, j, k), YE(i, j, k), Z(j, k)
1. (6.86)

For the objective of minimization of makespan:

K
Min MS = SL(k), (6.87)
k=1

ST(s, K)  Demands , ∀s. (6.88)

The model for makespan minimization consists of constraints (6.69)–(6.88). The
constraints (6.75) and (6.76) are misprinted in the original paper (constraints (11)
and (12) of Sundaramoorthy and Karimi [10]), in which they were written as fol-
lows:

Bijmin
b(i, j, k)
Bijmax − y(i, j, k),

∀i > 0, j ∈ suitij , 0 < k < K, (6.89)

YE(i, j, k)Bijmin
BE(i, j, k)
Bijmax − YE(i, j, k),

∀i > 0, j ∈ suitij , 0 < k
K. (6.90)

6.2.5
Computational Studies

6.2.5.1 Example 1
Consider the same example illustrated in Fig. 6.1. This is a standard example for
short-term scheduling of multipurpose batch plants and has been studied com-
prehensively by several authors. Two different products are produced through five
processing stages: heating, reactions 1, 2, and 3, and separation, as shown in the
STN representation of the plant flow sheet in Fig. 6.1. Since each of the reaction
tasks can take place in two different reactors, each reaction is represented by two
separate tasks. The processing time of task i on unit j is assumed to be a linear
function, αij + βij B, of its batch size, B. For simplicity, the relevant data [10, 28, 38]
of the constant (αij ) and linear (βij ) coefficients for processing times of different
tasks (i), the suitable units (j ), and their minimum (Bijmin ) and maximum (Bijmax )
186 6 Short-Term Scheduling of Batch and Continuous Processes

batch sizes are shown in Table 6.1 using the STN representation. The initial stock
level for all intermediates is assumed to be zero and unlimited storage capacity is
assumed for all states. The prices of products 1 and 2 are $10/mu.
For the objective of maximization of profit and a time horizon of 12 h, this
example is solved using the unit-specific event-based model of Ierapetritou and
Floudas [29] (I&F), the global event-based models of Castro et al. [22] (CBMN),
and Maravelias and Grossmann [28] (M&G), and using the slot-based model of
Sundaramoorthy and Karimi [10] (S&K). All the resulting mixed-integer linear pro-
gramming (MILP) models are solved in GAMS [41] distribution 21.1 using CPLEX
8.1.0 on the same computer (3 GHz Pentium 4 with 2 GB RAM). Table 6.2 pro-
vides a comparative study of these models in terms of the problem statistics such
as the number of binary and continuous variables, number of constraints, CPU
time taken to solve to the specified integrality gap, the number of nodes taken to
reach the optimal solution, the objective function at the relaxed node, and so forth.
It should be noted that for the S&K model, n event points are shown to represent
n − 1 slots for a valid comparison with the other global event and unit-specific
event-based models.
For this case, the slot-based/global event-based models require at least 11 event
points compared to the unit-specific event-based I&F model, which requires only
seven events. The slot-based/global event-based models are not solved until zero in-
tegrality gap as they take excessive computational time and because the I&F model
solves to the global optimal solution in just 6.19 s. The slot-based/global event-
based models take excessive computational time, and only the M&G model is able
to solve to the global optimal solution in the specified CPU time. The Gantt charts
for this case are shown in Figs. 6.4 and 6.5 for the I&F and M&G models, respec-
tively.
In the Gantt chart of Fig. 6.5 for the M&G model, it was shown [38] that, it cor-
responds to the requirement of a very tiny slot of duration 0.087 h (in the second
slot) for the slot-based/global event-based models to find the reported global opti-
mal solution. This is evidenced by the excessive CPU time taken by the slot-based
S&K model for which the global optimal solution is not obtained. However, in the

Table 6.1 Data of coefficients of processing times of tasks,

limits on batch sizes of units for Example 1.

Task (i) Unit (j ) α ij β ij B min

ij (mu) B max
ij (mu)

Heating (i = 1) Heater 0.667 0.00667 – 100

Reaction1 (i = 2) Reactor1 1.334 0.02664 – 50
(i = 3) Reactor2 1.334 0.01665 – 80
Reaction2 (i = 4) Reactor1 1.334 0.02664 – 50
(i = 5) Reactor2 1.334 0.01665 – 80
Reaction3 (i = 6) Reactor1 0.667 0.01332 – 50
(i = 7) Reactor2 0.667 0.008325 – 80
Separation (i = 8) Separator 1.3342 0.00666 – 200
Table 6.2 Model statistics and computational results for Example 1 under maximization of profit.

Model Events CPU time Nodes RMILP MILP Binary Continuous Constraints Nonzeros
(s) ($) ($) variables variables

Example 1 (H = 12)

S&K 9 561.58 288574 3265.2 2646.8 96 499 525 1867

10 10889.61 3438353 3315.8 2646.8 108 565 594 2119
11 >67000b 17270000 3343.4 2646.8a 120 631 663 2371
M&G 9 2317.38 611206 3265.2 2646.8 128 692 1594 5805
10 >67000c 10737753 3315.8 2646.8 144 775 1786 6782
11 >67000d 9060850 3343.4 2658.5 160 858 1978 7815
CBMN (t = 2) 9 331.72 593182 3730.5 2646.8 120 248 298 1426
10 4366.09 6018234 4070.0 2646.8 136 278 335 1614
11 >67000e 80602289 4409.5 2646.8a 152 308 372 1802
I&F 7 6.19 14962 3788.3 2658.5 42 165 318 1046

a Suboptimal solution.
b Relative gap = 1.59%.
c Relative gap = 3.16%.
d Relative gap = 5.12%.
e Relative gap = 2.58%.
6.2 Short-Term Scheduling of Batch Processes
187
188 6 Short-Term Scheduling of Batch and Continuous Processes

Fig. 6.4 Gantt chart for Example 1 (seven events) using I&F
model under maximization of profit.

Fig. 6.5 Gantt chart for Example 1 (11 events) using M&G model under maximization of profit.

Gantt chart of Fig. 6.4 for the I&F model, it is evident that such a slot would not be
necessary as they use a unit-specific event-based model. Hence, this case empha-
sizes the important difference between the slot-based/global event-based models
and the unit-specific event-based models. Because of the different alignment of the
start times of different units, sometimes the slot-based/global event-based mod-
els may require very small slots which can result in a very large number of event
points, and may prohibit the realization of global optimal solution in reasonable
CPU time compared to the unit-specific event-based models. The I&F model takes
only 6.19 s compared to the M&G model, which takes more than 67,000 s for ob-
taining the same global optimal solution. The S&K model, although has no big-M
constraints, is not able to find the global optimal solution in reasonable CPU time.

6.2.5.2 Example 2
Finding optimal solutions for problems where the minimization of the makespan
is the objective function is reported in the literature to be the most difficult schedul-
ing problem to solve, even for simple examples. For instance, consider a simple
example involving a multipurpose batch plant that requires one raw material and
produces two intermediates and one final product. The raw material is processed
in three sequential tasks, where the first task is suitable in two units (J1 and J2), the
second task is suitable in one unit (J3), and the third task is suitable in two units
(J4 and J5). The STN for this example is shown in Fig. 6.6.
 6.2 Short-Term Scheduling of Batch Processes 189

Fig. 6.6 STN for Example 2.

Table 6.3 Data of coefficients of processing times of tasks,

limits on batch sizes of units for Example 2.

Task (i) Unit (j ) α ij β ij B min

ij (mu) B max
ij (mu)

Task1 (i = 1) Unit1 1.333 0.01333 – 100

(i = 2) Unit2 1.333 0.01333 – 150
Task2 (i = 3) Unit3 1.000 0.00500 – 200
Task3 (i = 4) Unit4 0.667 0.00445 – 150
(i = 5) Unit5 0.667 0.00445 – 150

A task which can be performed in different units is considered as multiple, sep-

arate tasks, thus leading to five separate tasks (i = 1, . . . , 5), each suitable in one
unit. The relevant data [10, 38] is shown in Table 6.3. The initial stock level for all
intermediates is assumed to be zero and unlimited storage capacity is assumed for
all states.
Consider a demand for the final product S4 (D4 = 2000 mu). For all the models
that have big-M constraints (M&G and I&F), we need to specify a horizon time as
well (H = 50 h is used). The problem statistics are shown in Table 6.4. It can be
observed that the optimal solution obtained by the I&F model for 12 events (28.439
h) is better than the best solutions obtained by all other models, even at higher
event points. For the S&K and M&G models using 16 events, it takes relatively
longer time compared to the CBMN model, even though better solutions are not
obtained. For 17 events, using the S&K and CBMN (for t = 2) models the optimal
solution found is 28.773 h, for which both the models take more than 80,000 s. The
I&F model outperforms all the other models in terms of computational time and
problem size, and finds the global optimal solution of 27.881 h using 14 events
within a CPU time of 41.89 s.

6.2.5.3 Example 3
An additional example involving resource constraints such as utility requirements
and mixed storage policies is considered. The STN for this example is shown in
Fig. 6.7, and the corresponding data [28, 35, 38] is given in Tables 6.5 and 6.6.
There are two types of reactors available for the process (types I and II), with two
reactors of type I (R1 and R2) and one reactor of type II (R3) with four reactions
suitable in them. Reactions T1 and T2 require a type I reactor, whereas reactions T3
and T4 require a type II reactor. Additionally, reactions T1 and T3 are endothermic,
where the required heat is provided by steam (HS) available in limited amounts.
Reactions T2 and T4 are exothermic, and the required cooling water (CW) is also
available in limited amounts. Each reactor allows variable batch sizes, where the
 190

Table 6.4 Model statistics and computational results for Example 2 under minimization of Makespan.

Model Events H CPU Nodes time RMILP MILP Binary Continuous Constraints Nonzeros
time (s) (h) (h) variables variables

S&K 13 – 1.18 362 27.126 29.772 120 615 624 2074

14 – 31.54 15622 25.702 29.772 130 665 678 2253
15 – 728.05 400789 25.142 29.439 140 715 732 2432
16 – 37877.69 12064418 24.871 29.106 150 765 786 2611
17 – >80000b 17279722 24.716 28.773a 160 815 840 2790
M&G 13 50 2.19 394 27.126 29.772 120 556 1485 6056
14 645.06 139488 25.335 29.772 130 601 1605 6786
15 25253.81 5273904 25.024 29.439 140 646 1725 7551
16 >90000c 11258561 24.834 29106a 150 691 1845 8351
CBMN (t = 1) 13 – 0.02 0 27.126 29.772 60 191 239 824
(t = 1) 14 – 0.11 65 25.824 29.772 65 206 258 892
(t = 1) 15 – 0.28 417 25.358 29.772 70 221 277 960
(t = 2) 15 – 235.90 236250 19.049 29.439 135 286 407 1605
6 Short-Term Scheduling of Batch and Continuous Processes

(t = 2) 16 – 27994.64 23426601 17.049 29.106 145 306 436 1723

(t = 2) 17 – >80000d 80105289 15.049 28.773a 155 326 465 1841
I&F 12 50 0.12 208 24.236 28.439 50 160 253 732
13 2.26 7863 24.236 27.903 55 174 276 801
14 41.89 134961 24.236 27.881 60 188 299 870
15 950.64 2693556 24.236 27.881 65 202 322 939

a Suboptimal solution.
b Relative gap = 4.22%.
c Relative gap = 7.38%.
d Relative gap = 0.12%.
 6.2 Short-Term Scheduling of Batch Processes 191

Fig. 6.7 STN for Example 3.

Table 6.5 State related data for Example 3.

F1 F2 I1 I2 I3 P1 P2

STsmax (kg) 1000 1000 200 100 500 1000 1000

STs0 (kg) 400 400 0 0 0 0 0
prices ($/kg) 0 0 0 0 0 30 40

minimum batch size is half the capacity of the reactor. The processing times and
the utility requirements include a fixed time and a variable term that is propor-
tional to the batch size. The processing times are set so that the minimum batch
size is processed in 60% of the time needed for the maximum batch size. For the
raw materials and final products, unlimited storage is available, while for the inter-
mediates, finite storage is available. Two different cases of this example studied in
the literature [28, 35, 38] are considered that differ in the resource availability. In
the first case (Example 3a), we assume that the availability of both HS and CW is
40 kg/min, and in the second case (Example 3b), it is 30 kg/min. Also, two differ-
ent objective functions, maximization of profit and minimization of makespan, are
considered. For the global event-based formulations, the M&G and CBMN models,
and for the unit-specific event-based formulations, the model of Janak et al. [35, 36]
(JLF) are employed.

Maximization of Profit For the objective of maximization of profit and a time hori-
zon of 8 h, the optimal solution is $5904.0 in the first case (Example 3a) and
$5227.778 in the second case (Example 3b). The computational results in terms
of the model statistics and the CPU times are reported in Table 6.7 for the models
of M&G, JLF, and CBMN.

Minimization of Makespan For the objective of minimization of makespan, the op-

timal solution is 8.5 h in the first case (Example 3a) and 9.025 h in the second case
(Example 3b). The computational results in terms of the model statistics and the
CPU times are reported in Table 6.8. For the models involving big-M constraints
[28, 35] a common value of M = 10 is used.
 192

Table 6.6 Task related data for Example 3.a

T1 T2 T3 T4 T1 T2 T3 T4
capmin capmax α β α β α β α β γ iH S δ iH S γ iCW δ iCW γ iH S δ iH S γ iCW δ iCW

R1 40 80 0.5 0.025 0.75 0.0375 6 0.25 4 0.3

R2 25 50 0.5 0.4 0.75 0.06 4 0.25 3 0.3
R3 40 80 0.25 0.0125 0.5 0.025 8 0.4 4 0.5
6 Short-Term Scheduling of Batch and Continuous Processes

a capmin /capmax in kg, α in h, β in h/kg, γ in kg/min, and δ in

kg/min per kg of batch.
Table 6.7 Model statistics and computational results for Example 3 under maximization of profit.

Model Events CPU Nodes RMILP MILP Binary Continuous Constraints Nonzeros
time (s) ($) ($) variables variables

Example 3a
M&G 7 1.22 680 8870.5 5904.0 72 545 1082 4184
CBMN (t = 1) 7 0.30 376 8875.4 5482.04a 36 140 175 741
(t = 2) 7 1.07 1312 10396.7 5904.0 66 170 250 1216
JLF 6 1.03 294 10981.8 5904.0 45 273 1304 4606
Example 3b
M&G 6 0.27 67 7267.1 5227.8 60 470 925 3411
CBMN (t = 1) 6 0.09 96 7685.7 5000.0a 30 121 148 622
(t = 2) 6 0.16 112 8360.3 5227.8 54 145 208 1002
JLF 5 0.15 26 6414.7 5227.8 33 220 1028 3265

a Suboptimal solution.
6.2 Short-Term Scheduling of Batch Processes
193
 194

Table 6.8 Model statistics and computational results for Example 3 under minimization of Makespan.

Model Events CPU Nodes RMILP MILP Binary Continuous Constraints Nonzeros
time (s) (h) (h) variables variables

Example 3a
M&G 8 9.20 3331 5.48 8.5 84 620 1241 5039
CBMN (t = 1) 8 0.77 995 5.47 9.25a 42 159 204 861
(t = 2) 8 5.89 5641 2.37 8.5 78 195 294 1431
JLF 7 1.95 180 6.27 8.5 57 326 1601 6234
Example 3b
6 Short-Term Scheduling of Batch and Continuous Processes

M&G 7 0.74 197 5.85 9.025 72 545 1084 4209

CBMN (t = 1) 7 0.24 229 5.81 9.25a 36 140 177 742
(t = 2) 7 0.72 491 3.05 9.025 66 170 252 1217
JLF 6 0.59 10 6.35 9.025 45 273 1317 4693

a Suboptimal solution.
 6.3 Short-Term Scheduling of Continuous Processes 195

For both the objective functions, the unit-specific event-based JLF model requires
one event point less and has the least number of binary variables compared to the
global event-based models of M&G and CBMN. The CBMN model (using t = 2)
requires the least number of continuous variables, constraints and nonzeros. It
should be noted that the CBMN model yields suboptimal solution for t = 1 in
both the cases, and hence, the overall CPU time and the number of nodes (for both
t = 1 and t = 2) should be considered for a fair comparison.

6.3
Short-Term Scheduling of Continuous Processes

The problem of short-term scheduling of continuous plants has received relatively

less attention in the literature despite its practical importance in the chemical
process industries that produce a variety of products using both batch and con-
tinuous production modes. Scheduling of a continuous plant typically involves
handling continuous production, management of transitions, and accommodat-
ing inventory constraints, while meeting demands of final products with specified
due dates. One of the key differences between scheduling of batch processes ver-
sus scheduling continuous plants is in handling the processing times. In batch
plants, the processing times are typically fixed and known a priori, and the produc-
tion amount depends on the capacity of the batch unit. In continuous plants, the
processing times are a function of unit-dependent processing rates, final product
demand, and storage limitations. Additionally, in continuous plants, the produc-
tion amount is available continuously while it is being produced, unlike in batch
plants, where the amount is available only after the end time of the batch that
is processing. Due to these differences, the problem of scheduling of continuous
plants has drawn separate attention.
There are two types of demand patterns for scheduling of continuous processes
that have been addressed in the literature: cyclic and short-term. In cyclic schedul-
ing, a cyclic mode is assumed and the product demands are specified in terms
of constant demand rates at the end of a specified time horizon, while short-
term scheduling deals with a general problem where the product demands have
different sets of due dates. Sahinidis and Grossmann [42] proposed a slot-based
mixed-integer nonlinear programming (MINLP) model for cyclic scheduling of
single-stage continuous parallel production lines that do not share any common
resources. Pinto and Grossmann [43] modeled cyclic scheduling for sequential
operation of a multistage, multiproduct continuous plant based on a slot-based
continuous-time representation, leading to an MINLP model with explicit inven-
tory breakpoints for representing intermediate storage. Munawar et al. [44] ex-
tended the slot-based cyclic scheduling formulation of Pinto and Grossmann [43]
to hybrid flowshops consisting of serial and parallel configurations of processing
and storage units. They proposed a modified definition of the time slot to account
for feed losses during product transitions. Zhang and Sargent [11, 12] proposed a
global event-based MINLP model using RTN for the optimal operation of a mixed-
196 6 Short-Term Scheduling of Batch and Continuous Processes

production facility consisting of batch and continuous processes. Schilling and

Pantelides [16] also used the RTN framework for their slot-based continuous-time
formulation for short-term scheduling of batch and continuous processes and pro-
posed a special branch and bound solution that branches both on discrete and con-
tinuous variables. Karimi and McDonald [7, 45] developed two slot-based models
for production planning and short-term scheduling, which differ in the preassign-
ment of slots to the time periods for a single-stage multiproduct facility consisting
of parallel semicontinuous processors. Floudas and coworkers [29–36, 46–49] de-
veloped unit-specific event-based models for a variety of problems involving design,
synthesis, short-term, medium-term, reactive scheduling, and scheduling under
uncertainty. Ierapetritou and Floudas [29, 30] proposed a novel short-term schedul-
ing model for batch and continuous processes using STN framework deploying
unit-specific event-based continuous-time representation. They used an approxi-
mation of the storage task timings for handling different storage requirements in
their model for continuous processes [30]. Mockus and Reklaitis [14, 15] proposed a
global event-based MINLP model for short-term scheduling of batch and continu-
ous processes in which both the start and end times of tasks occur at events that are
common across all units. Their formulation can handle resource constraints such
as limited availability of utilities and manpower. Giannelos and Georgiadis [37, 50]
proposed unit-specific event-based formulation for short-term scheduling of batch
and continuous processes using STN representation. The start and end times of
the tasks producing/consuming the same state were respectively enforced to be
the same. While this feature is essential for continuous processes when bypass-
ing of storage is allowed (reasons discussed later), for all other cases, it will be very
restrictive and may lead to suboptimal solutions. Giannelos and Georgiadis [37] en-
forced these restrictions in their model for short-term scheduling of batch plants
as well, leading to suboptimal solutions as already mentioned. Mendez and Cerda
[51] proposed a production campaign-based algorithmic approach for short-term
scheduling of continuous processes leading to compact models for the case when
storage bypassing is allowed, with the assumptions that only one state is produced
by each task and no initial inventories exist for intermediate states. Castro et al. [22]
developed global event-based formulation for short-term scheduling of batch and
continuous processes using RTN representation, where changeovers are treated as
additional batch tasks. Most of the above models [30, 50–52, 22] can handle dif-
ferent storage requirements such as unlimited, finite, flexible, dedicated, and no
intermediate storage policies. However, for an industrial case study of consumer
goods manufacturing involving making, storage, and packing tasks, Ierapetritou
and Floudas [30] solved the cases of unlimited and finite intermediate storage and
reported approximate/suboptimal solutions. Mendez and Cerda [51] and Gianne-
los and Georgiadis [50] also reported suboptimal solutions for the case of finite
intermediate storage with no maximum demand limits. Castro et al. [22, 52] could
not find the global optimal solution for the no-intermediate-storage case even at
higher event points and large computational times. They classified the problem as
intractable and used a decomposition strategy [52] for improving the computational
performance.
 6.3 Short-Term Scheduling of Continuous Processes 197

Recently, Shaik and Floudas [53] (S&F) proposed an improved model for short-
term scheduling of continuous processes using unit-specific event-based continu-
ous-time representation. They modified and extended the formulation of Ierapetri-
tou and Floudas [30] and presented improved sequencing constraints to rigorously
address the different storage requirements. Their formulation is based on the STN
representation resulting in an MILP model that accurately accounts for various
storage requirements such as dedicated, finite, unlimited, and no intermediate
storage policies. The formulation allows for unit-dependent variable processing
rates, sequence-dependent changeovers, and with/without the option of bypassing
of storage. The S&F model is described below for different storage policies.

6.3.1
Unlimited Intermediate Storage

For this case, there is no need to model the storage tasks explicitly, and hence,
the model is described only using the continuous processing tasks ip . There is
no difference in the model due to whether bypassing of storage is allowed or not,
because of the availability of unlimited intermediate storage. The mathematical
model consists of the following allocation constraints, capacity constraints, mater-
ial balances for raw materials and intermediates, demand, duration, and sequenc-
ing constraints.

6.3.1.1 Allocation Constraints

w(i, n)
1, ∀j ∈ J, n ∈ N. (6.91)
i∈Ij

In each unit, only one task or no task takes place at any event as represented by
constraint (6.91).

6.3.1.2 Capacity Constraints for Processing Tasks

Rimin
p
(T f (ip , n) − T s (ip , n))
b(ip , n)
Rimax
p
(T f (ip , n) − T s (ip , n)),
∀ip ∈ Ip , n ∈ N, (6.92)
f s
b(ip , n) = Rip (T (ip , n) − T (ip , n)), ∀ip ∈ Ip , n ∈ N. (6.93)
The amount of material processed by a continuous processing task is constrained
by lower and upper bounds in (6.92), which are a function of the duration of the
corresponding task ip , which is represented by the difference between the end time
and start time of the task at event n, (T f (ip , n) − T s (ip , n)), and the minimum and
maximum processing rate of the task ip . For the case of constant processing rates,
the amount of production is related as described by constraint (6.93).

6.3.1.3 Material Balances

(A) Raw Materials:


c
ST0 (s, n) + ρsi p
b(ip , n) = 0, ∀s ∈ S R , n ∈ N. (6.94)
ip ∈Is
198 6 Short-Term Scheduling of Batch and Continuous Processes

In constraint (6.94), the amount of raw material, as and when required from the
external resources is related to the amounts consumed at the corresponding event
n. On the other hand, if the entire raw material requirement is supplied at the
beginning of the scheduling horizon, then constraint (6.94) is modified as follows:


c
ST (s, n) = ST (s, n − 1) + ρsi p
b(ip , n),
ip ∈Is
R
∀s ∈ S , n ∈ N, n > 1, (6.95)


c
ST (s, n) = ST0 (s) + ρsi p
b(ip , n), ∀s ∈ S R , n = 1. (6.96)
ip ∈Is

The total initial amount required from external resources, ST 0 (s), calculated in
Eq. (6.96), is either partly consumed at the first event n = 1 or remains in the
storage, which gets consumed during the subsequent events as described in (6.95)
above. (B) Intermediates:

p


c
ST (s, n) = ST (s, n − 1) + ρsip b(ip , n) + ρsi p
b(ip , n),
ip ∈Is ip ∈Is
IN
∀s ∈ S , n ∈ N, n > 1, (6.97)

p


c
ST (s, n) = ST0 (s) + ρsip b(ip , n) + ρsi p
b(ip , n),
ip ∈Is ip ∈Is
IN
∀s ∈ S , n = 1. (6.98)
Similarly, for the intermediate state s, the material balance is written in constraints
(6.97) and (6.98). ST (s, n) defines the excess amount of state s at event n. In con-
straint (6.97), the first term on right-hand side signifies the amount of state s stored
at the previous event (n − 1) in the storage; the second term represents the amount
of state s produced by the upstream processing task at event n. This total amount is
either consumed in the downstream processing task indicated by the third term or
remains in the storage at event n as shown on the left-hand side. At the first event,
the initial amount of available intermediates is taken into account in Eq. (6.98).
These constraints are based on the condition that an intermediate material is al-
lowed to go directly to the production task bypassing the storage, because only the
excess amount, ST (s, n), is stored. The other case where storage bypassing is not
allowed, irrespective of whether the amount produced by the upstream process-
ing unit is in excess of the amount consumed by the downstream processing, is
discussed later in Section 6.3.5.

6.3.1.4 Demand Constraints

p
Dsmin ≤ ρsip b(ip , n)
Dsmax , ∀s ∈ S P . (6.99)
n∈N ip ∈Is

The material balance for the final product state s is given in constraints (6.99),
where the total production of state s is limited between the specified lower and
upper bounds on the demands of final product.
 6.3 Short-Term Scheduling of Continuous Processes 199

6.3.1.5 Duration Constraints for Processing Tasks

T f (ip , n) − T s (ip , n)
Hw(ip , n), ∀ip ∈ Ip , n ∈ N. (6.100)

The duration constraint for processing tasks is given in (6.100), which states that
the duration is zero if the corresponding processing task is not active, otherwise
the constraint is relaxed.

6.3.1.6 Sequencing Constraints

(A) Same task in the same unit:
T s (i, n + 1)  T f (i, n), ∀i ∈ I, n ∈ N, n = N. (6.101)
Constraint (6.101) states that the start time of a task at the next event should be
greater than the finish time of the same task at the current event. (B) Different tasks
in the same unit: For the case of no changeover times, the two constraints for the
same task in the same unit and different tasks in the same unit can be combined
into one equation as shown in (6.102).
T s (i, n + 1)  T f (i  , n), ∀j ∈ J, i ∈ Ij , i  ∈ Ij , n ∈ N, n = N. (6.102)
For the case of sequence-independent changeovers, the constraint for different
tasks in the same unit requires modification as shown in (6.103), where τj is the
changeover time in unit j .

T s (i, n + 1)  T f (i  , n) + τj w(i  , n),

∀j ∈ J, i, i  ∈ Ij , i = i  , n ∈ N, n = N. (6.103)
For the case of sequence-dependent changeovers, the constraint for different tasks
in the same unit is generalized as follows:

T s (i, n)  T f (i  , n ) +ticl i w(i, n) − H (1 − w(i  , n ))

−H w(i  , n ),
i  n <n <n
∀j ∈ J, i, i  ∈ Ij , i = i  , n, n ∈ N, n > n . (6.104)
Note that in the above constraints, (6.102–6.104), only the last constraint, (6.104),
has big-M terms, unlike in the previous formulations of Ierapetritou and Floudas
[29, 30]. This feature is observed to result in better LP relaxed solutions which
often improve the computational performance. (C) Different tasks in different units
for processing tasks: The start and finish times of different processing tasks ip and
ip , which produce or consume the same state s, need to be aligned, and we have
the following two alternate ways of writing these constraints, (6.105) and (6.106)
and (6.105A) and (6.106A):

T s (ip , n)  T s (ip , n) − H (1 − w(ip , n)),

∀s ∈ S I N , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c
p
< 0, n ∈ N, (6.105)
p
200 6 Short-Term Scheduling of Batch and Continuous Processes

T f (ip , n)  T f (ip , n) − H (1 − w(ip , n)),

∀s ∈ S I N , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c
p
< 0, n ∈ N, (6.106)
p

T s (ip , n)  T s (ip , n) − H (2 − w(ip , n) − w(ip , n)),

∀s ∈ S I N , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c
p
< 0, n ∈ N, (6.105A)
p

T f (ip , n)  T f (ip , n) − H (2 − w(ip , n) − w(ip , n)),

∀s ∈ S I N , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c
p
< 0, n ∈ N. (6.106A)
p

In these constraints the start and end times of the downstream (consuming)
processing tasks are, respectively, enforced to be later than the corresponding up-
stream (producing) processing tasks that processes the same state s. Constraints
(6.105) and (6.106) enforce the condition that the start and end times of the con-
suming tasks need to be always aligned to the corresponding producing tasks
whenever the producing task is active; while (6.105A) and (6.106A) state that the
start and end times of the corresponding processing tasks need to be aligned con-
ditionally, when both the tasks are active.
Constraints (6.105) and (6.106) are required when we use the variable ST (s, n)
in the material balance constraint (6.97), for instance, when we do not consider
storage as a separate task for the unlimited-intermediate-storage case (or for the
dedicated finite-intermediate-storage case discussed later). Otherwise, if we do not
have ST (s, n) (for instance, for the no-intermediate-storage case discussed later)
or if we consider storage as a separate task and use the variable B(ist , n) instead
of ST (s, n) in the material balance (for instance, for the dedicated- and flexible-
finite-intermediate-storage cases discussed later), we need constraints (6.105A) and
(6.106A). The reason for this is that, when we do not consider storage as a separate
task and use the variable ST (s, n) in the material balance constraint (6.97), there is
an implicit assumption that when the downstream consuming task starts at event
n, the amount stored in ST (s, n − 1) would be available for consumption, which
may not be always valid (because of the heterogeneous locations of events used),
unless the consuming tasks are always aligned to the producing tasks whenever the
producing tasks are active, as described in the sequencing constraints (6.105) and
(6.106). It should be noted that the fact that the model of Ierapetritou and Floudas
[30] for continuous processes results in the reported suboptimal solutions [22, 50,
51] is due to lack of the second constraint, (6.106) or (6.106A), relating the end
times of the producing and consuming tasks.
 6.3 Short-Term Scheduling of Continuous Processes 201

6.3.1.7 Extra Tightening Constraint

The following tightening constraint gives a better LP relaxation.



(T f (i, n) − T s (i, n))
H − τjmin , ∀j ∈ J. (6.107)
n∈N i∈Ij

It states that the sum of the durations of all tasks suitable in unit j is limited by the
available time in the horizon (H − τjmin ), where τjmin is a lower bound on the total
clean up time required in unit j .

6.3.2
Dedicated Finite Intermediate Storage with Storage Bypassing Allowed

Next, consider the case where a dedicated finite intermediate storage is available
for each intermediate state s. For this case as well, there is no need to model the
storage tasks explicitly, because the storage tanks are anyway dedicated in nature.
We are only interested in constraining the finite nature of the intermediate storage.
Hence, the model is described here only using the continuous processing tasks ip ,
and the other case when storage is considered as a separate task is discussed later
in Section 6.3.4. All the above constraints for the case of unlimited storage remain
the same except the constraints for different tasks in different units for process-
ing tasks. Here, since we do not consider storage as a separate task, we use con-
straints (6.105) and (6.106). Additional constraints would be required depending on
whether storage bypassing is allowed or not. Initially, we consider the case when
storage bypassing is allowed and then discuss the other case later in Section 6.3.5.

6.3.2.1 Storage Bypassing Allowed

When an intermediate material is allowed to go directly to the production task by-
passing a finite intermediate storage or if there is no intermediate storage available,
then to ensure the feasibility of the inventory capacity balance, the start and finish
times of processing tasks ip and ip , which produce and consume the same state s,
respectively, need to be the same if both tasks are active at the same event point as
shown in the following constraints:
T s (ip , n)
T s (ip , n) + H (2 − w(ip , n) − w(ip , n)),
(6.108)
∀s ∈ S FIS , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c < 0, n ∈ N,
p p

T f (ip , n)
T f (ip , n) + H (2 − w(ip , n) − w(ip , n)),

(6.109)
∀s ∈ S FIS , ip ∈ Is , ip ∈ Is , ip = ip , ρsi  > 0, ρsi
p c < 0, n ∈ N.
p p

So, from constraints (6.105), (6.106), (6.108), and (6.109), both the start and end
times of the producing and consuming tasks of the same state s would be respec-
tively same, if both tasks are active at event n. If either of the tasks is not active, then
constraints (6.108) and (6.109) are relaxed and are trivially satisfied. This zero-wait
condition is required to ensure the feasibility of the inventory capacity balance as
illustrated in [53], because of the unit-specific nature of the continuous-time rep-
resentation used in the formulation. The formulation of Ierapetritou and Floudas
202 6 Short-Term Scheduling of Batch and Continuous Processes

[30] did not take this into account. In the formulation of Giannelos and Georgiadis
[37], they enforced this condition even for batch plants as well, which is unrealistic,
and hence, their formulation led to suboptimal solutions [10, 38].
Now, to constrain the finite nature of the intermediate storage available, the fol-
lowing bounds are added for the states that have the finite storage requirements:
ST (s, n)
STsmax , ∀s ∈ S FIS , n ∈ N. (6.110)

6.3.3
No Intermediate Storage

For the case when no intermediate storage is available, the excess amount of state
s, ST (s, n), is driven to zero at each event n or simply this variable is eliminated.
Then, the material balance constraints (6.97) and (6.98) reduce to the following,
meaning that the amount of state s produced at an event has to be consumed at the
same event.

p

c
ρsip b(ip , n) + ρsi p
b(ip , n) = 0, ∀s ∈ S I N , n ∈ N. (6.111)
ip ∈Is ip ∈Is

The condition of enforcing the same start and end times of the producing and
consuming tasks of the same state s, described in constraints (6.105A), (6.106A),
(6.108), and (6.109) is again applicable because no intermediate storage is available.
Here, (6.105A) and (6.106A) are used because in the material balance constraint,
(6.111), there is no assumption that the consuming task will receive material from
the storage at the previous event, and hence, there is no need to enforce the align-
ment unconditionally.

6.3.4
Flexible Finite Intermediate Storage with Storage Bypassing Allowed

This is a general case where finite intermediate storage is available and for each
material state several suitable storage options exist. A material state can be stored in
all or a limited number of storage units and vice versa. To handle this general case,
separate tasks are introduced for the storage activity, because the storage cannot
have more than one state at any time and we need to accommodate additional
constraints for relating the timing of storage tasks (ist ) to that of the processing
tasks (ip ). The nature of these constraints would be different depending on whether
storage bypassing is allowed or not. Initially, we consider the case when storage
bypassing is allowed and then discuss the other case later in Section 6.3.5. For the
dedicated-finite-intermediate-storage case, in contrast to the model discussed in
Section 6.3.2, if we alternately chose to consider storage as a separate task, then the
following same model would be applicable.
The additional constraints in the mathematical model are described below for
the case of flexible finite intermediate storage when storage bypassing is allowed.
The allocation constraint in (6.91) remains the same except that now it is written
over all units (both processing and storage).
 6.3 Short-Term Scheduling of Continuous Processes 203

6.3.4.1 Allocation Constraints for Storage Tasks

w(ist , n + 1)  w(ist , n) + z(ist , n) − 1, ∀ist ∈ Ist , n ∈ N. (6.112)

Constraint (6.112) states that if a storage task is active and it stores a non-zero
amount at event n, then the same storage task should be active at the next event
n + 1 as well. Additionally, this constraint also avoids unnecessary tank-to-tank
transfer of material because the same storage task would be active at the next event
as well.

6.3.4.2 Capacity Constraints for Storage Tasks

b(ist , n)
Vimax
st
w(ist , n), ∀ist ∈ Ist , n ∈ N, (6.113)
b(ist , n)
Vimax
st
z(ist , n), ∀ist ∈ Ist , n ∈ N. (6.114)
The capacity constraints for processing tasks remain the same as in (6.92) or (6.93),
while for storage tasks the amount of material that can be stored is limited by
the available capacity of the corresponding storage unit as shown in constraint
(6.113). Constraint (6.114) is the same as (6.113), but using a different binary vari-
able z(ist , n) to confine only those instances when b(ist , n) = 0, which is realized
through the penalty term on the number of binary variables in the objective func-
tion described later in (6.135).

6.3.4.3 Material Balances

Constraint (6.94), for calculating the amount of raw material as and when required
from the external resources, remains the same. For the other case when the entire
raw material requirement is supplied at the beginning of the scheduling horizon,
constraints (6.95) and (6.96) are modified as follows:

p

c
ρsist b(ist , n − 1) + ρsi b(i, n) = 0,
ist ∈Ist i∈Is
R
∀s ∈ S , n ∈ N, n > 1, (6.115)



c R
ST0 (s) + ρsi b(i, n) = 0, ∀s ∈ S , (6.116)
n=1 i∈Is

where the set Is consists of both processing and storage tasks. The variable ST (s, n)
is eliminated here because separate storage tasks are defined explicitly. Similarly,
the material balances in (6.97) and (6.98) for the intermediates are modified as
given below.

p
p

c
ρsip b(ip , n) + ρsist b(ist , n − 1) + ρsi b(i, n) = 0,
ip ∈Is ist ∈Is i∈Is

∀s ∈ S F I S , n ∈ N, n > 1, (6.117)

p


c
ρsip b(ip , n) + ST0 (s) + ρsi b(i, n) = 0,
ip ∈Is i∈Is

∀s ∈ S F I S , n = 1, (6.118)
204 6 Short-Term Scheduling of Batch and Continuous Processes

where the set Is consists of both processing and storage tasks. These constraints
are based on the same condition that an intermediate material is allowed to go
directly to the production task bypassing the storage. The demand constraint in
(6.99) remains the same.

6.3.4.4 Duration Constraints for Storage Tasks

T f (ist , n)  T s (ist , n), ∀ist ∈ Ist , n ∈ N. (6.119)

The duration constraint given in (6.100) remains the same for processing tasks,
while for the storage tasks the finish times have to be later than the start times as
shown in (6.119).
The sequencing constraints, (6.101)–(6.104), (6.105A), (6.106A), (6.108) and
(6.109) for the same task in the same unit, different tasks in the same unit, and
different tasks in different units for the processing tasks remain the same. Here,
because storage is considered as a separate task we use constraints (6.105A) and
(6.106A). The sequencing constraints for storage tasks are defined below.

6.3.4.5 Sequencing Constraints. Different Tasks in Different Units for Storage Tasks
The start time of a storage task that stores the intermediate state s should be same
as the start time of the processing task that either produces or consumes the state
s as follows:

T s (ist , n)  T s (ip , n) − H (2 − w(ist , n) − w(ip , n)),

∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N, (6.120)

T s (ist , n)
T s (ip , n) + H (2 − w(ist , n) − w(ip , n)),

∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N. (6.121)
If the corresponding processing task that either produces or consumes the state
s is not active (w(ip , n) = 0), and there is a nonzero amount in the storage
(z(ist , n) = 1), then the start time of the storage task should be equal to the fin-
ish time of the same task at the previous event and is realized through constraints
(6.122) and (6.101).

T s (ist , n)
T f (ist , n − 1) + H (1 − z(ist , n) + w(ip , n)),

∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N, n > 1. (6.122)
The finish time of the storage task for storing state s should be greater than the
finish times of both producing and consuming tasks of the state s as follows:

T f (ist , n)  T f (ip , n) − H (2 − w(ist , n) − w(ip , n)),

∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N. (6.123)
Additionally, if the amount stored in the storage is zero (z(ist , n) = 0), then the fin-
ish time of the storage task should be same as the finish time of the corresponding
consuming processing task and is realized from constraints (6.123) and (6.124).
 6.3 Short-Term Scheduling of Continuous Processes 205

T f (ist , n)
T f (ip , n) + H (2 − w(ist , n) − w(ip , n)) + H z(ist , n),

∀s ∈ S I N , ist ∈ Is , ip ∈ Is , ρsi
c
p
< 0, n ∈ N. (6.124)
However, if the amount stored in the storage is nonzero (z(ist , n) = 1), then it
should remain in the storage until the start time of the processing task at the next
event. So, the finish time of the storage task should be same as the start time of the
next processing task and is realized from constraints (6.125) and (6.126).

T f (ist , n)  T s (ip , n + 1) − H (2 − w(ist , n) − w(ip , n + 1))

−H (1 − z(ist , n)),
∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N, n = N, (6.125)

T f (ist , n)
T s (ip , n + 1) + H (2 − w(ist , n) − w(ip , n + 1))

+H (1 − z(ist , n)),
∀s ∈ S I N , ist ∈ Is , ip ∈ Is , n ∈ N, n = N. (6.126)
The tightening constraint given in (6.107) is again applicable.

6.3.5
Dedicated and Flexible Finite Intermediate Storage without Storage Bypassing

The nature of constraints is different when storage bypassing is not allowed for
the cases where production must go through finite storage (dedicated or flexible)
before consumption in the downstream units. This is a general case and we do not
need to enforce the same start and end times for producing or consuming tasks of
the same state, because the material always goes through storage.

6.3.5.1 Dedicated-Finite-Intermediate-Storage Case without Bypassing of Storage

Here again, we have the option of considering storage as a separate task or not. In
this section, we describe the model without considering storage as a separate task,
and the other case when storage is considered as a separate task is discussed in the
next section along with flexible-finite-intermediate-storage case.
When we do not consider storage as a separate task, the model comprises of
all the constraints, (6.91)–(6.104), (6.105), (6.106), and (6.107), discussed earlier in
Section 6.3.1. Additionally, we need constraint (6.110) and the following material
balance constraints for the intermediate states.

Material Balance for Intermediates

p
ST (s, n − 1) + ρsip b(ip , n) ≤ STsmax ,
ip ∈Is
FIS
∀s ∈ S , n ∈ N, n > 1, (6.127)

p
ST0 (s) + ρsip b(ip , n) ≤ STsmax , ∀s ∈ S FIS , n ∈ N, n = 1. (6.128)
ip ∈Is
206 6 Short-Term Scheduling of Batch and Continuous Processes

In constraints (6.127) and (6.128), the total amount received into the dedicated stor-
age at each event is constrained to be within the maximum capacity limits.

6.3.5.2 Flexible-Finite-Intermediate-Storage Case without Bypassing of Storage

For the dedicated-finite-intermediate-storage case as well, in contrast to the model
discussed in the previous section, if we alternately chose to consider storage as
a separate task, then the following same model would be applicable for both
dedicated- and flexible-finite-storage cases. The constraints, (6.91), (6.92), (6.94),
(6.99)–(6.1041) (6.104), (6.106A), (6.107), (6.112)–(6.114), (6.117)–6.119, (6.122)–
(6.126, discussed in the previous sections are required. Constraint (6.105A) is not
required as it is implicitly enforced, because consuming tasks are aligned to stor-
age tasks which in turn are aligned to the producing tasks. Additionally, we need
the following allocation, material balance, and sequencing constraints.

Allocation Constraints

p
w(ist , n)  w(ip , n), ∀s ∈ S FIS , ip ∈ ρsip , n ∈ N. (6.129)
ist ∈ρsi
c
st

Constraint (6.129) states that whenever a producing task of state s is active then one
or more of suitable storage tasks also need to be active for all intermediate states
that have the finite storage requirement.

Material Balance for Intermediates Here because the storage tasks are modeled ex-
plicitly, constraints (6.127) and (6.128) are modified as follows:

p
p

ρsist b(ist , n − 1) + ρsip b(ip , n) ≤ Vimax
si
w(ist , n),
ist ∈Is ip ∈Is ist ∈Is
FIS
∀s ∈ S , n ∈ N, n > 1, (6.130)

p


ST0 (s) + ρsip b(ip , n) ≤ Vimax
si
w(ist , n),
ip ∈Is ist ∈Is
FIS
∀s ∈ S , n = 1. (6.131)

Sequencing Constraints: Different Tasks in Different Units

T s (ist , n)  T s (ip , n) − H (2 − w(ist , n) − w(ip , n)),

p
∀s ∈ S I N , ist ∈ Is , ip ∈ ρsip , n ∈ N, (6.132)

T s (ist , n)
T s (ip , n) + H (2 − w(ist , n) − w(ip , n)),

p
∀s ∈ S I N , ist ∈ Is , ip ∈ ρsip , n ∈ N, (6.133)

T s (ip , n)  T s (ist , n) − H (2 − w(ist , n) − w(ip , n)),

∀s ∈ S I N , ist ∈ Is , ip ∈ ρsi
c
p
, n ∈ N. (6.134)
 6.3 Short-Term Scheduling of Continuous Processes 207

Constraints (6.132) and (6.133) impose that the start time of the storage task is the
same as the start time of the corresponding processing task that produces the inter-
mediate state, if both the storage task and the producing task are active. Constraint
(6.134) states that the start time of the processing task that consumes the interme-
diate state should be later than the start time of the corresponding storage task if
both tasks are active.

Objective: Maximization of Profit

p


C1 prices ρsi b(i, n) − C2 w(i, n)
s∈S P n∈N i∈Is i∈I n∈N



−C3 z(ist , n). (6.135)
ist ∈Ist n∈N

The objective is maximization of profit due to sales from the production of final
products, and additionally, there are penalties for the total number of binary vari-
ables as shown in (6.135), where Ci is the corresponding cost coefficient.

Objective: Minimization of Makespan For the objective of minimization of makespan

(MS) the following constraints need to be added:

Min MS
T f (i, n)
MS, ∀i ∈ I, n ∈ N. (6.136)
The tightening constraint in (6.107) is modified as follows:



(T f (i, n) − T s (i, n))
MS − τjmin , ∀j ∈ J. (6.137)
n∈N i∈Ij

6.3.6
Computational Study

6.3.6.1 Example 4
An industrial case study of fast moving consumer goods manufacturing plant that
is extensively studied by several authors in the literature [22, 30, 50–53] is con-
sidered. The plant follows a common production sequence: mixing, storage, and
packing. The mixing stage has three parallel mixers (mixers A, B, and C) operat-
ing in a continuous mode, that produce seven intermediates (int1-int7) using three
different base stocks (bases A, B, and C) available as required. These intermediates
may be stored in three storage tanks (tanks A, B, and C) or directly packed in five
continuous packing lines (lines 1–5), thus producing 15 final products (S1–S15). A
time horizon of 120 h is considered. The STN representation of the plant is shown
in Fig. 6.8 along with the unit suitability for each task.
The base stocks are available in unlimited initial amounts. For each task suitable
on multiple units, a separate task is considered. For instance, two making tasks
(m31 and m32) are considered for producing ‘int3’ using mixers B and C, respec-
tively. The 15 final products are produced using 15 packing tasks (p1–p15). The
208 6 Short-Term Scheduling of Batch and Continuous Processes

Fig. 6.8 STN for Example 4.

problem data such as production rates, cleanup times, minimum and maximum
demand specifications, and prices used in the literature [50] are given in Tables 6.9
and 6.10. The minimum production requirements for the final products are shown
Table 6.11.
The following different cases of this process are considered:
• unlimited intermediate storage (Case A),
 6.3 Short-Term Scheduling of Continuous Processes 209

Table 6.9 Production rates for Example 4.

Task (i p ) Unit (J p ) R max

i (ton/h)

m1 Mixer A 17
m2 Mixer A 17
m31 Mixer B 17
m32 Mixer C 17
m41 Mixer B 17
m42 Mixer C 17
m51 Mixer B 12.24
m52 Mixer C 12.24
m61 Mixer B 12.24
m62 Mixer C 12.24
m71 Mixer B 12.24
m72 Mixer C 12.24
p1 Line 3 5.5714
p2 Line 1 5.8333
p3 Line 2 2.7083
p4 Line 1 5.8333
p5 Line 2 2.7083
p6 Line 3 5.5714
p7 Line 4 2.2410
p8 Line 1 5.8333
p9 Line 2 2.7083
p10 Line 5 5.3571
p11 Line 5 5.3571
p12 Line 4 3.3333
p13 Line 4 2.2410
p14 Line 2 2.7083
p15 Line 4 3.3333

Table 6.10 Cleanup times for Example 4.

Changeover (i
p ↔ i p ) Unit (J p ) τ cl
(h) τ min

j (h)
i i

(p2, p4) ↔ p8 Line 1 1 1

(p3, p5) ↔ (p9, p14) Line 2 4 4
p1 ↔ p6 Line 3 1 1
(p12, p15) ↔ (p13, p7) Line 4 2 2

• no intermediate storage (Case B), and

• flexible finite intermediate storage (Case C).

For each of the above cases, we compare the results from the literature [22, 30,
51, 52], except for the model of Giannelos and Georgiadis [50] for which the com-
parison is based on our implementation of their model. Ierapetritou and Floudas
[30] had solved case A; Giannelos and Georgiadis [50] and Mendez and Cerda [51]
210 6 Short-Term Scheduling of Batch and Continuous Processes
Table 6.11 State specific data for Example 4.

State (s) ST 0 (s) ST max

s D min
s P rices

Base A, B, C ∞
Int1-Int7 60
S1 220 1
S2 251 1
S3 116 1
S4 15 1
S5 7 1
S6 47 1
S7 8.5 1
S8 144 1
S9 42.5 1
S10 114.5 1
S11 53 1
S12 2.5 1
S13 16.5 1
S14 13.5 1
S15 17.5 1

solved cases A and C; Castro et al. [22, 52] solved cases A, B, and C in their papers.
All the computations in this work are performed on 3.2 GHz, Pentium 4 machine
with 1 GB RAM using GAMS [41] (distribution 21.7) and CPLEX 9.0.2, and the case
studies are solved with an integrality gap less than 0.3%.

6.3.6.2 Unlimited Intermediate Storage (Case A)

Consider the simple case of unlimited intermediate storage available for all the
intermediates. The objective function is maximization of profit, the first term in
(6.135) with C1 = 1. The model statistics are reported in Table 6.12. For this simple
case, all the models [50, 51, 22, 53] are able to find the global optimal solution of
2695.32. The S&F formulation gave an improvement in both the relaxed (RMIP)
and optimal solution found, compared to the approximate model of Ierapetritou
and Floudas [30], which reported an objective of 2689.42 using seven events. The
Gantt chart for S&F model is depicted in Fig. 6.9.
The unit-specific event-based models of S&F and Giannelos and Georgiadis [50]
require only four events compared to the global event-based model of Castro et al.
[22], which reported a requirement of nine events. The campaign-based algorithmic
model of Mendez and Cerda [51] does not require consideration of event points and
reported the least model statistics.

6.3.6.3 No Intermediate Storage (Case B)

Next, consider the case of no intermediate storage for all the seven intermediates.
The objective function is a maximization of profit, the first two terms in (6.135)
with C1 = 10 and C2 = 1. The S&F model requires consideration of only four
events and takes 157.86 CPU s to find the objective function of 26869.467, which
 6.3 Short-Term Scheduling of Continuous Processes 211

Table 6.12 Results for Case A (UIS) of Example 4.

Shaik & Floudas Castro et al. Giannelos & Mendez &

[53] [22] Georgiadis [50] Cerda [51]

Events 4 9 4
Binary variables 108 236 108 38
Continuous variables 356 762 628 44
Constraints 1040 894 1637 140
Nonzeros 4867 5354
RMILP 2695.32 2695.32 2695.32
MILP 2695.32 2695.32 2695.32 2695.32
Profit 2695.32 2695.32 2695.32 2695.32
Integrality gap (%) 0 0 0
CPU time (s) 1.03 58.5a 186.44 14a
Nodes 69 1041 15310

a CPU times for other models reported for completeness only.

Fig. 6.9 Gantt chart for case A (UIS) of Example 4.

corresponds to a profit of 2689.747. The model statistics are reported in Table 6.13,
and the Gantt chart for S&F model is depicted in Fig. 6.10.
Only Castro et al. [22, 52] reported a solution for this case among the other works.
They classified this problem as intractable and could not obtain the global optimal
solution corresponding to the profit of 2689.747, despite increasing the total num-
ber of event points to 9, for which case they reported an objective of 2672.50 after
an excessive CPU time. They used decomposition strategies [52] to reduce the com-
putational efforts.

6.3.6.4 Flexible Finite Intermediate Storage (Case C)

Now, consider the case of flexible finite intermediate storage for all the intermedi-
ates. Three storage tanks (tanks A, B, and C) are available each with a maximum
212 6 Short-Term Scheduling of Batch and Continuous Processes
Table 6.13 Results for Case B (NIS) of Example 4.

Shaik & Floudas [53] Castro et al. [52]

Events 4 9
Binary variables 108 236
Continuous variables 328 762
Constraints 1240 894
Nonzeros 5926
RMILP 26946.722 2695.32
MILP 26869.467 2672.50
Profit 2689.747 2672.50
Integrality gap (%) 0.26 0.85
CPU time (s) 157.86 200000a
Nodes 4272 8007335

a CPU time reported for completeness only.

Fig. 6.10 Gantt chart for case B (NIS) of Example 4.

capacity of 60 ton, and any intermediate can be stored in any of the three tanks.
Since there are seven intermediates, 21 additional storage tasks need to be consid-
ered. Unlike in the S&F model, the start and end times of storage tasks are not
precisely calculated in the model of Giannelos and Georgiadis [50], rather it seems
that they determine and/or adjust these timings during postprocessing. In our im-
plementation of their model, we used the following equation to calculate the start
times of all active storage tasks during postprocessing.

Tits = τit − θst , ∀s ∈ S f is , i ∈ Isst , t ∈ T . (6.138)

Then, in order to avoid the unnecessary activation of storage tasks in our imple-
mentation of their model, we used the following objective function, (6.139), which
is similar to (6.135) of the S&F model.
 6.3 Short-Term Scheduling of Continuous Processes 213

Table 6.14 Results for Case C (FIS) of Example 4.

Shaik & Floudas Castro et al. Giannelos & Mendez &

[53] [22] Georgiadis [50] Cerda [51]

Events 4 10 4
Binary variables 276 330 220 60
Continuous variables 580 927 712 87
Constraints 4267 1127 2113 361
Nonzeros 21130 6884
RMILP 26946.72 2695.32 2695.32
MILP 26910.18 2695.32 2692.06 2670.28
Profit 2695.32 2695.32 2692.06 2670.28
Integrality gap (%) 0.120 0 0.12
CPU time (s) 465.61 162a 60000 399a
Nodes 7144 3934 3307486

a CPU times for other models reported for completeness only.

C1 prices ST(s, tn ) − C2 x(i, t) − C3 y(s, t). (6.139)
s∈S P i∈I t∈T s∈S f is t∈T

The model statistics are reported in Table 6.14. For the objective function of max-
imization of profit in (6.135) with C1 = 10 and C2 = C3 = 1, the S&F model
requires four events, and the objective function is found to be 26910.181 in 465.61
CPU s with an integrality gap of 0.12%, which corresponds to a profit of 2695.32.
The Gantt chart for S&F model is depicted in Fig. 6.11.

Fig. 6.11 Gantt chart for case C (FIS) of Example 4.

214 6 Short-Term Scheduling of Batch and Continuous Processes

This case is regarded as one of the hardest instances to solve in the literature.
The S&F model finds the global optimum solution and requires consideration of
only four events compared to the global event-based model of Castro et al. [22],
which reported ten events for this case as shown in Table 6.14. The campaign-based
algorithmic model of Mendez and Cerda [51], although it is very compact in terms
of least problem statistics, could not find the global optimal solution for this case
(the best solution reported corresponds to a profit of 2670.28). The formulation of
Giannelos and Georgiadis [50] also reported a suboptimal solution corresponding
to a profit of 2689.48 using four events. On the basis of our implementation of their
model, an improved objective value of 2692.06 was found within a maximum CPU
time of 60,000 s and an integrality gap of 0.12%. However, in the Gantt chart it was
observed that the end times of some storage tasks are not precisely calculated as
discussed in [53].

6.4
Conclusions

An overview of the recent developments in short-term scheduling of batch and con-

tinuous plants is presented. Different continuous-time based scheduling models
from the literature are discussed, followed by a comparative study of these methods
on some benchmark examples. It is concluded that the unit-specific event-based
models require less number of events, have least problem statistics, and have faster
computational performance compared to both slot-based and global event-based
models.

Acknowledgments

The authors gratefully acknowledge support from the National Science Foundation
and BASF Aktiengesellschaft, Ludwigshafen, Germany.

References

1 Floudas, C. A., Lin, X., Continuous- 3 Pinto, J. M., Grossmann, I. E., Op-
time versus discrete-time approaches timal cyclic scheduling of multistage
for scheduling of chemical processes: continuous multiproduct plants. Com-
A review. Comput. Chem. Eng. 28 put. Chem. Eng. 18 (1994), p. 797.
(2004), p. 2109. 4 Pinto, J. M., Grossmann, I. E., A
2 Floudas, C. A., Lin, X., Mixed inte- continuous time mixed integer lin-
ger linear programming in process ear programming model for short
scheduling: Modeling, algorithms, term scheduling of multistage batch
and applications. Ann. Oper. Res. 139 plants. Ind. Eng. Chem. Res. 34 (1995),
(2005), p. 131. p. 3037.
 References 215

5 Pinto, J. M., Grossmann, I. E., An 15 Mockus, L., Reklaitis, G. V., Contin-

alternate MILP model for short-term uous time representation approach
scheduling of batch plants with pre- to batch and continuous process
ordering constraints. Ind. Eng. Chem. scheduling. 2. Computational issues.
Res. 35 (1996), p. 338. Ind. Eng. Chem. Res. 38 (1999), p. 204.
6 Pinto, J. M., Grossmann, I. E., A 16 Schilling, G., Pantelides, C. C.,
logic-based approach to scheduling A simple continuous-time process
problems with resource constraints. scheduling formulation and a novel
Comput. Chem. Eng. 21 (1997), p. 801. solution algorithm. Comput. Chem.
7 Karimi, I. A., McDonald, C. M., Plan- Eng. 20 (1996), p. S1221.
ning and scheduling of parallel semi- 17 Schilling, G., Pantelides, C. C.,
continuous processes. 2. Short-term Optimal periodic scheduling of mul-
scheduling. Ind. Eng. Chem. Res. 36 tipurpose plants. Comput. Chem. Eng.
(1997), p. 2701. 23 (1999), p. 635.
8 Lamba, N., Karimi, I. A., Scheduling 18 Mendez, C. A., Henning, G. P.,
parallel production lines with resource Cerda, J., Optimal scheduling of
constraints. 1. Model formulation. batch plants satisfying multiple prod-
Ind. Eng. Chem. Res. 41, (2002), p. 779. uct orders with different due-dates.
9 Lamba, N., Karimi, I. A., Scheduling Comput. Chem. Eng. 24 (2000), p.
parallel production lines with resource 2223.
constraints. 2. Decomposition algo- 19 Mendez, C. A., Cerda, J., An MILP
rithm. Ind. Eng. Chem. Res. 41 (2002), continuous-time framework for short-
p. 790. term scheduling of multipurpose
10 Sundaramoorthy, A., Karimi, batch processes under different oper-
I. A., A simpler better slot-based ation strategies. Optim. Eng. 4 (2003),
continuous-time formulation for p. 7. 0
short-term scheduling in multipur- 20 Mendez, C. A., Henning, G. P.,
pose batch plants. Chem. Eng. Sci. 60 Cerda, J., An MILP continuous-time
(2005), p. 2679. approach to short-term scheduling of
11 Zhang, X., Sargent, R. W. H., The resource-constrained multistage flow-
optimal operation of mixed produc- shop batch facilities. Comput. Chem.
tion facilities – A general formulation Eng. 25 (2001), p. 701.
and some approaches for the solution. 21 Castro, P., Barbosa-Povoa, A. F.
Comput. Chem. Eng. 20 (1996), p. 897. D., Matos, H., An improved RTN
12 Zhang, X., Sargent, R. W. H., The continuous-time formulation for the
optimal operation of mixed produc- short-term scheduling of multipur-
tion facilities – Extensions and im- pose batch plants. Ind. Eng. Chem.
provements. Comput. Chem. Eng. 22 Res. 40 (2001), p. 2059.
(1998), p. 1287. 22 Castro, M., Barbosa-Povoa, A. P.,
13 Mockus, L., Reklaitis, G. V., Math- Matos, H. A., Novais, A. Q., Sim-
ematical programming formulation ple continuous-time formulation for
for scheduling of batch operations short-term scheduling of batch and
based on nonuniform time discretiza- continuous processes. Ind. Eng. Chem.
tion. Comput. Chem. Eng. 21 (1997), p. Res., 43 (2004), p. 105.
1147. 23 Mendez, C. A., Cerda, J., Optimal
14 Mockus, L., Reklaitis, G. V., Contin- scheduling of a resource-constrained
uous time representation approach multiproduct batch plant supplying
to batch and continuous process intermediates to nearby end-product
scheduling. 1. MINLP formulation. facilities. Comput. Chem. Eng. 24
Ind. Eng. Chem. Res. 38 (1999), p. 197. (2000), p. 369.
216 6 Short-Term Scheduling of Batch and Continuous Processes

24 Majozi, T., Zhu, X. X., Novel continuous-time formulation. Com-

continuous-time MILP formula- put. Chem. Eng. 25 (2001), p. 665.
tion for multipurpose batch plants. 34 Lin, X., Floudas, C. A., Modi, S.,
1. Short-term scheduling. Ind. Eng. Juhasz, N. M., Continuous-time opti-
Chem. Res. 40 (2001), p. 5935. mization approach for medium-range
25 Lee, K., Park, H. I., Lee, I., A novel production scheduling of a multiprod-
nonuniform discrete time formulation uct batch plant. Ind. Eng. Chem. Res.
for short-term scheduling of batch 41 (2002), p. 3884.
and continuous processes. Ind. Eng. 35 Janak, S. L., Lin, X., Floudas, C. A.,
Chem. Res. 40 (2001), p. 4902. Enhanced continuous-time unit-
26 Burkard, R. E., Fortuna, T., specific event-based formulation
Hurkens, C. A. J., Makespan mini- for short-term scheduling of multi-
mization for chemical batch processes purpose batch processes: Resource
using nonuniform time grids. Com- constraints and mixed storage poli-
put. Chem. Eng. 26 (2002), p. 1321. cies. Ind. Eng. Chem. Res. 42 (2004), p.
2516.
27 Wang, S., Guignard, M., Redefining
event variables for efficient modeling 36 Janak, S. L., Lin, X., Floudas, C. A.,
of continuous-time batch processing. Comments on “Enhanced continuous-
Ann. Oper. Res. 116 (2002), p. 113. time unit-specific event-based for-
mulation for short-term scheduling
28 Maravelias, C. T., Grossmann, I. E.,
of multipurpose batch processes: re-
New general continuous-time state-
source constraints and mixed stor-
task network formulation for short-
age policies.” Ind. Eng. Chem. Res. 44
term scheduling of multipurpose
(2005), p. 426.
batch plants. Ind. Eng. Chem. Res. 42
(2003), p. 3056. 37 Giannelos, N. F., Georgiadis, M. C.,
A simple new continuous-time for-
29 Ierapetritou, M. G., Floudas, C. A., mulation for short-term scheduling
Effective continuous-time formulation of multipurpose batch processes. Ind.
for short-term scheduling: 1. Multi- Eng. Chem. Res. 41 (2002), p. 2178.
purpose batch processes. Ind. Eng.
Chem. Res. 37 (1998), p. 4341. 38 Shaik, M. A., Janak, S. L., Floudas,
C. A., Continuous-time models for
30 Ierapetritou, M. G., Floudas, C. short-term scheduling of multipur-
A., Effective continuous-time formu- pose batch plants: A comparative
lation for short-term scheduling: 2. study. Ind. Eng. Chem. Res. 45 (2006),
Continuous and semi-continuous p. 6190.
processes. Ind. Eng. Chem. Res. 37
39 Kondili, E., Pantelides, C. C., Sar-
(1998), p. 4360.
gent, R. W. H., A general algorithm
31 Ierapetritou, M. G., Floudas, C. A., for short-term scheduling of batch op-
Comments on “An Improved RTN erations. 1. Milp formulation. Comput.
continuous-time formulation for the Chem. Eng. 17 (1993), p. 211.
short-term scheduling of multipur-
40 Pantelides, C. C., Unified frame-
pose batch plants.” Ind. Eng. Chem.
works for optimal process planning
Res. 40 (2001), p. 5040.
and scheduling. In Rip-pin, D. W. T.,
32 Ierapetritou, M. G., Hene, T. S., Hale, J. C., Davis, J. (eds.), Proceed-
Floudas, C. A., Effective continuous- ings of Second International Confer-
time formulation for short-term ence on Foundations of Computer-
scheduling: 3. Multiple intermedi- Aided Process Operations, Crested
ate due dates. Ind. Eng. Chem. Res. 38 Butte, Colorado, 1993, pp. 253–274.
(1999), p. 3446. 41 Brooke, A., Kendrick, D., Meeraus,
33 Lin, X., Floudas, C. A., Design, syn- A., Raman, R., GAMS: A User’s Guide,
thesis and scheduling of multipur- GAMS Development Corporation:
pose batch plants via an effective South San Francisco, CA, 2003.
 References 217

42 Sahinidis, N. V., Grossmann, I. E., for scheduling under uncertainty: I.

MINLP model for cyclic multiprod- Bounded uncertainty. Comput. Chem.
uct scheduling on continuous parallel Eng. 28 (2004), p. 1069.
lines. Comput. Chem. Eng. 15 (1991), 49 Janak, S. L., Lin, X., Floudas, C. A.,
p. 85. A new robust optimization approach
43 Pinto, J. M., Grossmann, I. E., Op- for scheduling under uncertainty: II.
timal cyclic scheduling of multistage Uncertainty with known probability
continuous multiproduct plants. Com- distribution. Comput. Chem. Eng. 31
put. Chem. Eng. 18 (1994), p. 797. (2007), p. 171.
44 Munawar, S. A., Bhushan, M., 50 Giannelos, N. F., Georgiadis, M. C.,
Gudi, R. D., Belliappa, A. M., Cyclic A novel event-driven formulation for
scheduling of continuous multi- short-term scheduling of multipur-
product plants in a hybrid flowshop pose continuous processes. Ind. Eng.
facility. Ind. Eng. Chem. Res. 42 (2003), Chem. Res., 41 (2002), p. 2431.
p. 5861.
51 Mendez, C. A., Cerda, J., An efficient
45 McDonald, C. M., Karimi, I. A., Plan-
MILP continuous-time formulation
ning and scheduling of parallel semi-
for short-term scheduling of multi-
continuous processes: 1. Production
product continuous facilities. Comput.
planning. Ind. Eng. Chem. Res. 36
Chem. Eng. 26 (2002), p. 687.
(1997), p. 2691.
46 Janak, S. L., Floudas, C. A., Kall- 52 Castro, M., Barbosa-Povoa, A. P.,
rath, J., Vormbrock, N., Production Novais, A. Q., A divide-and-conquer
scheduling of a large-scale indus- strategy for the scheduling of process
trial batch plant. I. Short-term and plants subject to changeovers using
medium-term scheduling. Ind. Eng. continuous-time formulations. Ind.
Chem. Res. 45 (2006), p. 8234. Eng. Chem. Res. 43 (2004), p. 7939.
47 Janak, S. L., Floudas, C. A., Kall- 53 Shaik, M. A., Floudas, C. A., Im-
rath, J., Vormbrock, N., Production proved unit-specific event-based
scheduling of a large-scale industrial model continuous-time model for
batch plant. II. Reactive scheduling. short-term scheduling of continuous
Ind. Eng. Chem. Res. 45 (2006), p. processes: Rigorous treatment of stor-
8253. age requirements. Ind. Eng. Chem.
48 Lin, X., Janak, S. L., Floudas, C. A., Res. 46 (2007), p. 1764.
A new robust optimization approach
 219

7
Modeling and Optimization of Refinery Operations Considering
Uncertainty
Zhenya Jia, Marianthi G. Ierapetritou

This paper addresses the problem of refinery scheduling under uncertainty

through two approaches: multiobjective robust optimization and parametric mixed
integer linear programming (pMILP). The robust optimization model considers a
number of uncertainty scenarios, and normal boundary intersection (NBI) method
is utilized to determine the Pareto optimal surface in the objective space, on which
each point represents a trade-off between the original objective (minimal cost or
inventory) and the solution robustness. The pMILP approach investigates how the
value of cost or inventory and optimal schedule change with respect to the un-
certain parameter. Three case studies considering the optimal operations of crude
oil unloading and mixing, and gasoline blending and distribution are presented to
illustrate the importance of considering uncertainty in demand in refinery schedul-
ing operations and the results of the two proposed approaches.

7.1
Introduction

A substantial amount of work has been done in the area of refinery planning and
scheduling. A detailed literature review of which can be found in Pinto et al. [22].
The availability of linear programming (LP)-based commercial software for refinery
production planning, such as refinery and petrochemical modeling system (RPMS)
[4] and process industry modeling system (PIMS) [2], has allowed the development
of general production plans of the whole refinery. On the other hand, very few
optimization-based formulations are presented for the refinery scheduling due to
the lack of rigorous models to handle specific plant characteristics.
The problem of crude oil unloading with inventory control is addressed by Lee
et al. [16] based on time discretization that results in mixed integer linear pro-
gramming (MILP) model. The same problem is formulated as an MINLP model
by Li et al. [17], where the nonlinearity arises from calculating the crude oil con-
centration in the storage and charging tanks. Gasoline blending is a crucial step in
refinery operation as gasoline can yield 60–70% of a refinery’s profit. The process

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
220 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

involves mixing various stocks, which are the intermediate products from the re-
finery, along with some additives, such as antioxidants and corrosion inhibitors, to
produce blends with certain properties [6]. A variety of support systems have been
developed to address planning and scheduling of blending operations. StarBlend
[23] for example, which is developed by Texaco, uses a multiperiod blending model
written in GAMS that facilitates the incorporation of future requirements into cur-
rent blending decisions. Singh et al. [25] addressed the problem of blending opti-
mization for in-line blending for the case of stochastic disturbances in feedstock
qualities. They presented a real-time optimization method that can provide signif-
icantly improved profitability. Deterministic global optimization methods, such as
branch-and-bound, cutting plane algorithms were also studied by Horst and Tuy
[9]. Ryoo and Sahinidis [24] proposed a branch-and-bound-based method for dis-
crete/continuous global optimization. Using the idea of range reduction tests of
variables, they [7, 8] developed MILP, which is based on a resource-task network
representation, to solve the task of short-term scheduling of blending processes.
The recipe optimization problem is formulated as a nonlinear program and the
results are returned to the scheduling problem, so that an overall optimization can
be achieved.
As an attempt toward integrating all the stages in refinery operations, Jia and Ier-
apetritou [10–12] developed a comprehensive mathematical programming model
for the efficient scheduling of oil refinery operations. The overall problem is spa-
tially decomposed into three domains: the crude oil unloading and blending,
the production-unit operations, and the product blending and delivery, as shown
schematically in Fig. 7.1. Each of those sub-problems is modeled and solved in an
efficient way using continuous time representation that reduces the overall num-
ber of variables and constraints.
Substantial benefits can be achieved through the use of optimization techniques
in plant operations by improving the resource utilization at different levels of
decision-making process as illustrated for refinery operations by the work reviewed
above. However, uncertainty exists in realistic manufacturing environment due to
lack of accurate process models and variability of process and environment data.
The presence of uncertainty can substantially reduce or eliminate the advantages of
optimization approaches. Therefore, it is of great importance to develop systematic
methods to address the problem of scheduling under uncertainty. Existing work
mainly includes stochastic programming approaches involving chance constraints
and two-stage programming [3, 13], as well as robust optimization methods [1, 19,
26]. A brief overview of these approaches can be found in Jia and Ierapetritou [14].
The variations of certain parameters in refinery, such as feedstock qualities and
yield level are of great importance in operation decisions. In the oil industry, the
prices and demands of crude oil, gasoline, and diesel also fluctuate greatly. In order
to provide a methodology that accurately models operating conditions and yields a
robust solution, it is necessary to evaluate and to accommodate the realities of un-
certainty. However, the issue of uncertainty in refinery operations has not been suf-
ficiently studied for scheduling and planning problems due to the high complexity
of the deterministic case. Li et al. [18] address the problem of refinery planning
 7.2 Refinery Scheduling 221

under uncertainty. A general formulation for revenue and cost calculations is pro-
posed by considering uncertainty in raw material availability and product demand.
Different standard loss function approximation methods are compared and inte-
grated into the planning model. Neiro and Pinto [21] developed a model based on a
nonlinear programming formulation. The model first incorporates multiple plan-
ning periods and the selection of different crude oil types. Uncertainty related to
crude oil and product prices as well as demand is then included as a set of discrete
events with specific probabilities.
In this work, we propose to address the problem of uncertainty in refinery
scheduling through multiobjective robust optimization and a new pMILP ap-
proach. The idea is that when there is sufficient information about uncertain pa-
rameters, robust optimization can be utilized, however, when uncertainty is not
well characterized parameters MILP can be used to determine alternative optimal
schedules to cover the uncertainty range. This paper is organized as follows: the
mathematical models of crude oil unloading and mixing, gasoline blending, and
distribution problems are described in Section 7.2. Section 7.3 briefly explains the
basic ideas of the multiobjective robust optimization and pMILP approaches. Three
case studies are provided in Section 7.4 to illustrate the steps and applicability of
the proposed approaches. Section 7.5 summarizes the work and presents some of
the ideas for future developments.

7.2
Refinery Scheduling

The modeling of the overall refinery operations from the crude oil arrival to the
distribution of oil products gives rise to intractable mathematical models. Thus, de-
composition methodologies have long been recognized as the attractive direction
to overcome this computational burden. The overall refinery system as presented
in Fig. 7.1 is spatially decomposed into three domains: the crude oil unloading and
blending, the production-unit operations and the product blending and delivery.
In particular, the first problem involves the crude oil unloading from vessels, its
transfer to storage tanks, and the charging schedule for each crude oil mixture to
the distillation units. The second problem consists of the production units schedul-
ing which includes both fractionation and reaction processes, and the third prob-
lem describes the finished product blending and shipping end of the refinery. The
mathematical models presented in our earlier work are adopted and the details can
be found in earlier publication [10, 11].
The models involve mainly material balance constraints, allocation constraints,
sequence constraints, and demand constraints. Material balance constraints con-
nect the amounts of material in one unit at one event point to that at the next event
point. Allocation constraints set the delivery assignments between two consecutive
stages, and the beginning and finishing times of each operation are determined
by the sequence constraints. Demand constraints ensure that all the demands will
222 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

Problem 1 Problem 2 Problem 3

Crude-Oil Storage Charging Crude Other Comp. Blend

Marine Tanks Tanks Dist. Prod. Stock Header
Vessels Units Units Tanks

Finished Prod. Lifting/Shipping

Tanks Points

Fig. 7.1 Graphic overview of refinery system.

be satisfied during the time horizon. In addition, constraints related to operating

rules are applied to meet specific requirements.

7.3
Uncertainty Analysis

In most of the existing refinery scheduling studies, the problem data are assumed
to be deterministic. However, uncertainty arises in realistic refinery parameters
such as production recipes, processing times, and order requirements due to lack
of accurate process models and variability of process and environment data. There-
fore, the aim of this chapter is to (a) help the decision maker to select the optimal
overall solution; and (b) to explore the effects of uncertainty on the objective func-
tion which consists of operation cost, inventory or overall production, as well as the
best way to react to uncertain parameter variability. Two different approaches are
used toward this target. For the case where there is sufficient information regard-
ing uncertainty, stochastic robust optimization is used, whereas for the case where
uncertainty is not well characterized parametric programming is used to determine
alternative schedules to cover the uncertainty range.

7.3.1
Multiobjective Robust Optimization

The concept of robust optimization (RO) was first developed by Mulvey et al. [20] to
handle the trade-off associated with solution and model robustness. Assuming that
 7.3 Uncertainty Analysis 223

the problem data are described by a set of scenarios, a solution to an optimization

is considered to be solution robust if it remains close to the optimal for all scenarios.
In addition to the original objective, the solution robustness is also included in the
objective function.
Multiobjective optimization targets the identification of the set of Pareto opti-
mal solutions, where the improvement in one objective can only be improved at
the expense of another objective. NBI is a solution methodology developed by Das
and Dennis [5] for generating Pareto surface in nonlinear multiobjective optimiza-
tion problems. It is proved that this method is independent of the relative scales of
the objective functions and is successful in producing an evenly distributed set of
points in the Pareto surface given an evenly distributed set of parameters, which is
an advantage compared to the most common multiobjective approaches – weight-
ing method and the
-constraint method [27].
For a general multiobjective optimization problem:

 
f1 (x)
 f2 (x) 
 
(MOP) min F (x) =  .  n
2,
x∈C  .. 
fn (x)

where

C = {x : h(x) = 0, g(x)  0, a  x  b}.

A point x ∗ ∈ C is Pareto optimal (or nondominated) for multiobjective optimiza-

tion problem (MOP) if and only if there is no x ∈ C such that fi (x)  fi (x ∗ ) for
all i ∈ 1, 2, . . . , n, with at least one strict inequality.
The basic procedure of NBI method is first to obtain anchor point Fi∗ for the
(MOP), by minimizing the ith objective independently. The shadow minimum
(utopia point) F ∗ , is defined as the vector containing the individual global minima
of the objectives. Then, it searches for the intersection points between the bound-
ary of the objective space and the normal pointing toward the origin emanating
from any point in the convex hull of individual minima (CHIM), as illustrated in
Fig. 7.2. Let  be the n × n matrix whose ith column is Fi∗ − F ∗ , then given a
convex weighting ω, ω represents a point in the CHIM. Let n̂ denotes the unit
normal to the CHIM simplex toward the origin, then ω + t n̂ represents the set of
points on that normal.
As proposed in our earlier work [14] NBI can be successfully used for the solution
of robust scheduling optimization. A multiobjective robust optimization model
was first developed in order to incorporate model robustness and solution robust-
ness in the short-term scheduling formulation, and normal boundary intersection
(NBI) technique is then utilized in order to trace the Pareto optimal surface in the
objective space, on which each point represents a trade-off between the objectives
of economic optimality and solution and model robustness.
224 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

f1(x)
F1*

CHIM

f1* . F* F2*

f2* f2(x)
Fig. 7.2 Pareto optimal surface for a two-objective problem.

7.3.2
Parametric Mixed Integer Linear Programming (pMILP)

An alternative way to deal with uncertainty in scheduling problems is to solve the

problem parametrically with respect to uncertain parameters that gives rise to alter-
native schedules with different characteristics in the face of uncertainty. The solu-
tion of pMILP can determine how the objective value and integer solution change
with respect to the uncertain parameters and provide the functional form of the ob-
jective function. Jia and Ierapetritou [15] proposed an approach to solve the pMILP
problems where uncertain parameters are present on the right-hand side (RHS)
of the constraints. To summarize, the basic steps of the presented approach are
explained here for the case of single uncertain parameter for simplicity in the pre-
sentation.
As illustrated in Fig. 7.3, first the original problem is solved at a starting point
following a Branch and Bound solution procedure, and the dual information is
collected at each leaf node. Then, LP sensitivity analysis is performed for the relaxed
LP problems at the leaf nodes. Then, the next step is to determine θ max beyond
which the structure of the branch and bound may not remain the same. θ max can
be found through the following equation:
  p 
z − z0
θ max = min θ basis , min 0 ,
λ − λp
where z0 and λ0 are the objective value and dual multiplier at the optimal node 0,
respectively, and θ basis is the basis range of node 0. The next step is to update the
tree when θ = θ max +
is slightly larger than θ max . The procedure continues
by determining the new θ max , and these steps are executed repeatedly until no
feasible solution exists beyond the current range.
 7.4 Case Studies 225

Solve the original problem

using a branch and bound tree

max
Find ∆ θ that leaves the structure Collect z,p λp at
of the tree unchanged each leaf node p

max
For ∆ θ = ∆θ + ε
update the branch and bound tree

Fig. 7.3 Flow chart of proposed approach for single pMILP.

7.4
Case Studies

In this section, three refinery scheduling problems are presented to illustrate the
effectiveness of the proposed approach. The first two examples use the NBI method
to obtain Pareto optimal solutions, whereas for the third example, the MILP uncer-
tainty analysis approach is applied to study the effects of uncertain parameter on
the optimal solution behavior.

Example 1. The first example addresses the operation schedule of gasoline blend-
ing and distribution. It involves 5 products being produced and stored in 11 storage
tanks, and then lifted to satisfy 10 orders in a time horizon of 8 days. The demands
of five of the orders are considered to be uncertain parameters with a variability
of ±50%, as illustrated in Table 7.1. Two demand scenarios of each order are con-
sidered, leading to a total of 32 scenarios. The problem is modeled using the for-
mulation presented in Section 7.2 incorporating the 32 scenarios and the following

Table 7.1 Amount of orders for Example 1.

Order Amount

o1 8+
− 50%
o2 3
o3 3
o4 8+
− 50%
o5 3
o6 8+
− 50%
o7 3
o8 88 +
− 50%
o9 3
o10 100 +
− 50%
226 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

35

30

B
25
Expected Positive Deviation

20

15

10 A

0
1100 1120 1140 1160 1180 1200 1220 1240
Expected Inventory and Production

Fig. 7.4 Pareto optimal surface for Example 1.

constraint is included in order to measure the expected positive deviation from the
average inventory and production (solution robustness).

obj 2 = pk δ k , δ k = max 0, obj 1k − pk obj 1k ),

k k

where pk represents the probability of scenario k and δ k corresponds to the positive

deviation of inventory and production (obj 1k ) under scenario (k) from the expected
value. Thus, a dual objective formulation is developed to minimize the expected to-
tal inventory and production (objective 1), and the solution robustness (objective 2).
The basic steps of NBI method as described in Section 7.3.1 are then applied to
the two-objective robust optimization model. The two anchor points obtained by
solving two single objective function problems individually are
1112.76∗ 1234.36
f1 (x ∗ ) = and f2 (x ∗ ) = .
30.27 0∗
1112.76
Consequently, the utopia point is: F ∗ = and the matrix  =
0
0 121.6
.
30.27 0
Choosing a step size of δ1 = 0.05, ω1 and ω2 take the following values:

ω1 = [0, 0.05, 0.1, 0.15, 0.2, . . . , 1],

ω2 = [1, 0.95, 0.9, 0.85, . . . , 1 − ω1 ].
 7.4 Case Studies 227

14.012
pt_11
8
8.000 8.000 8.000 60.195
pt_10
1 4 6 8

pt_9

pt_8

13.792 12.491
pt_7
8 10

pt_6

3.000 3.000
pt_5
3 9
87.509
pt_4
10
3.000 3.000 3.000
pt_3
2 5 7

pt_2

pt_1

0 24 48 72 96 120 144 168 192

Fig. 7.5 Gantt chart of operation schedule A at nominal demand.

The Pareto optimal surface is illustrated in Fig. 7.4, where a number of evenly
distributed Pareto optimal points are obtained through NBI technique. Thus, in
the face of order uncertainty, a number of solutions can be achieved that corre-
spond to different values of average total inventory and production and positive
deviation. For example, at point A ((ω1 , ω2 ) = (0.25, 0.75)) in Fig. 7.4, the average
total inventory and production in the face of uncertainty is 1192.4 and the average
positive deviation from the mean is 4.68. At point B ((ω1 , ω2 ) = (0.8, 0.2)), the two
objective values are found to be 1129.9 and 22.41. The distribution schedules at
these two points show that large orders, such as order 8, is processed three times
and order 10 is processed twice, while the rest are processed once only. However,
the tank utilization of schedule A is splitted among tanks 3, 4, 5, 7, 10, 11, while
schedule B shows that the orders mostly take place in tanks 6 and 10. These two
operation schedules at the nominal demands are presented in Figs. 7.5 and 7.6,
respectively. To satisfy the same amount of demand, the total inventory and pro-
duction is 1161.3 using schedule A, while schedule B only needs 1053.8. However,
schedule A is more flexible to accommodate order fluctuations without large devi-
ations at the inventory and production. Consequently, a risk averse decision maker
would consider schedule A as a better choice.
228 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

12.355 12.491
pt_11
8 10
8.000 8.000 8.000 60.195
pt_10
1 4 6 8

pt_9

pt_8

15.449
pt_7
8
3.000 3.000 3.000 3.000
pt_6
3 5 7 9
3.000
pt_5
2
87.509
pt_4
10

pt_3

pt_2

pt_1

0 24 48 72 96 120 144 168 192

Fig. 7.6 Gantt chart of operation schedule B at nominal demand.

Table 7.2 System information for Example 2.

Scheduling Horizon(# of unit time) 10

Number of Vessel Arrivals 2
Arrival Time Amount of Crude Concentration of key components
Vessel 1 1 60 0.01, 0.04
Vessel 2 5 60 0.03, 0.02
Number of Storage Tanks 2
Storage Tanks Capacity Initial Oil Amount Concentration of key components
Tank 1 100 20 0.01, 0.04
Tank 2 100 10 0.03, 0.02
Number of Charging Tanks 2
Charging Tanks Capacity Initial Oil Amount Initial comp. concentration (min, max)
Tank 1 100 20 0.0167 (0.01, 0.02) 0.0333 (0.03, 0.038)
Tank 2 100 20 0.03 (0.025, 0.035) 0.023 (0.018, 0.027)
Number of CDU 1
Unit costs involved in vessel operation Unloading cost: 8, Sea waiting cost: 5
Tank inventory unit costs Storage tank: 0.05, Charging tank: 0.08
Demand of mixed oils by CDUs Oil mix 1: 60, Oil mix 2: 60

Example 2. The multiobjective robust optimization approach is also applied to a

crude oil unloading and mixing problem as discussed in Section 7.2. The data and
crude oil flow network are as illustrated in Table 7.2 and Fig. 7.7, respectively. As-
suming ±50% variability of demand for crude oil mix in both charging tanks, five
demand scenarios (30, 45, 60, 75, 90) are used to describe the uncertainty of each
tank, leading to a total of 25 scenarios. Similarly, a two-objective formulation is de-
 7.4 Case Studies 229

Vessels Storage Charging CDU

Tanks Tanks

Fig. 7.7 Oil flow network for Example 2.

3.5

B
2.5
Expected Positive Deviation

1.5

A
0.5

0
60 65 70 75 80 85 90
Expected Cost

Fig. 7.8 Pareto optimal surface for Example 2.

veloped to minimize the expected total operation cost (objective 1), which includes
sea waiting cost, unloading cost, inventory cost of storage tanks and charging tanks,
and the expected positive deviation (objective 2) from the average cost.
The resulted Pareto optimal surface is shown in Fig. 7.8. As we expected, a num-
ber of evenly distributed Pareto optimal points are obtained. Moving from any point
to another, the average cost cannot be reduced without decreasing the expected pos-
itive deviation, and vice versa. For example, at point A ((ω1 , ω2 ) = (0.05, 0.95)), the
unloading schedule obtained at the nominal values of uncertain demands shows
that vessel 1 should start unloading from day 1 and finish on day 3, while vessel 2
should unload from day 5 to day 7. Storage tank 1 is scheduled to transfer crude
oil mix to charging tank 1 on day 7 and day 9, while storage tank 2 delivers oil mix
230 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

to charging tank 1 on day 7, and to charging tank 2 on day 6 and 9. The CDU is
charged by charging tank 1 on day 1, 8, 10 and by charging tank 2 on day 1, 7, 10.
At this point, the average cost in the face of uncertainty is 87.232 and the average
metric for robustness is 0.055. At point B ((ω1 , ω2 ) = (0.8, 0.2)), these two objec-
tive values are found to be 66.19 and 2.414. The nominal unloading schedules of
the two vessels are the same as point A but the schedules of the other transfer are
different. In particular, storage tank 1 transfers crude oil mix to charging tank 1 on
day 2 and day 4, and storage tank 2 transfers oil mix to charging tank 1 on day 2,
and to charging tank 2 on day 5 and 7. The CDU is charged by charging tank 1 on
day 1, 3, 5 and by charging tank 2 on day 1, 6, 8. It should be noticed that although
schedule B has a smaller expected cost which means that can fulfill the same de-
mand scenarios with lower cost comparing to schedule A, it exhibits larger variance
indicating higher variability of schedules with respect to different scenarios.

Example 3. The third example is a crude oil unloading and mixing problem, which
considers one crude oil vessel, one storage tank, one charging tanks, and one CDU.
The data and crude oil flow network are as illustrated in Table 7.3 and Fig. 7.9,
respectively. The demand from charging tank 2 is considered to be the uncertain
parameter, which is expected to vary within (60, 100).

Table 7.3 System information for Example 3.

Scheduling Horizon(# of unit time)

Number of Vessel Arrivals 1
Arrival Time Amount of Crude Concentration of key components
Vessel 1 1 80 0.02
Number of Storage Tanks 1
Storage Tanks Capacity Initial Oil Amount Concentration of key components
Tank 1 100 20 0.02
Number of Charging Tanks 2
Charging Tanks Capacity Initial Oil Amount Initial comp. concentration (min, max)
Tank 1 100 20 0.02 (0.015, 0.025)
Tank 2 100 40 0.04 (0.015, 0.045)
Number of CDU 1
Unit costs involved in vessel operation Unloading cost: 8, Sea waiting cost: 5
Tank inventory unit costs Storage tank: 0.05, Charging tank: 0.08
Demand of mixed oils by CDUs Oil mix 1: 60, Oil mix 2: 60

Storage Charging
Vessel CDU
Tank Tanks

Fig. 7.9 Oil flow network for Example 3.

 7.4 Case Studies 231

Table 7.4 Values of binary variables of optimal schedules for Example 3.

(Task, Unit) n0 n1 n2

(Vessel, Storage Tank) 0 1 0

(Storage Tank, Charging Tank 1) 0 1 0

(Storage Tank, Charging Tank 2) 1 0 0

(Charging Tank 1, CDU) 1 0 1

(Charging Tank 2, CDU) 0 1 0

(Schedule A)

(Task, Unit) n0 n1 n2

(Vessel, Storage Tank) 1 0 0

(Storage Tank, Charging Tank 1) 1 0 0

(Storage Tank, Charging Tank 2) 0 1 0

(Charging Tank 1, CDU) 0 1 0

(Charging Tank 2, CDU) 1 0 1

(Schedule B)

The basic methodology presented in Section 7.3.2 is first applied in this example
as follows.
Step 1: The problem is solved at an initial demand value (60) with branch and
bound solution method. The objective value and the dual multiplier at the leaf
nodes of the B&B tree are collected. The optimal schedule is found to be schedule
A with cost 35.87 as shown in Table 7.4.
Step 2: Performing linear sensitivity analysis on node A, we get θ max = 0.
Step 3: For θ = θ max +
= 1 (
= 1), the tree is updated and it is found that
schedule B provides the optimal schedule with objective value of 36.74.
Step 2: Sensitivity analysis on node B gives the current optimal basis θ basis =
zp − z0
39. The value of 0 is calculated at other 16 leaf nodes and the minimum
λ − λp
value of those nodes is 16.5, which is less than 39. That means node B could be
intersected by the node that provides the minima. Hence, θ max is equal to 16.5.
Step 3: For θ = θ max +
= 17 (
= 0.5), node B still provides the optimal
schedule and the new optimal cost is 36.91 when the demand of charging tank 2
becomes 78.
Schedule B continues to be the optimal schedule in the next iteration, in which
θ max = 22 (the demand is equal to 100). After that, the problem becomes in-
232 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

ST 80.0
1

0 1 2 3 4 5 6 7 8

CT2 20.0
1

CT1 80.0
1

0 1 2 3 4 5 6 7 8

CDU 20.0 40.0

1 1
60.0
2

0 1 2 3 4 5 6 7 8

Fig. 7.10 Gantt chart of operation schedule A at nominal demand.

feasible. Therefore, schedule A provides the optimal solution when the demand
from charging tank 2 is equal to 60. Schedule B becomes the optimal for the range
(60, 100), and the linear function of how objective (inventory and production) value
changes with respect to the demand is: z = 36.73 + 0.01θ. Schedules A and B at
the nominal demand (60, 60) are shown in Figs. 7.10 and 7.11. It should be pointed
out that in schedule A, the oil mix in storage tank is transferred to charging tank
2 first and then to charging tank 1, and the charging sequence to CDU is tank 1,
tank 2, and finally tank 1. However, these sequences are exactly the opposite in
schedule B.
The robust optimization approach is also applied to this example, by considering
five scenarios (60, 70, 80, 90, 100) in the uncertain demand space, optimizing the
two objectives of expected operation cost and expected positive deviation from the
mean that represents solution robustness. All the Pareto optimal solutions shown
in Fig. 7.12 are found to correspond to schedule B. Note that the infeasibility is
not considered in the robust optimization model, which means the resulted sched-
ules must be feasible for any realization of the scenarios. Therefore, as expected,
schedule A is not included in the solution set since it becomes infeasible when the
demand is greater than 60.
Comparing the information obtained by the two approaches it can be observed
that parametric studies are independent of the information about uncertain para-
meter variability thus providing a complete picture of optimal solutions.
 7.4 Case Studies 233

ST 80.0
1

0 1 2 3 4 5 6 7 8

CT2 20.0
1

CT1 80.0
1

0 1 2 3 4 5 6 7 8

CDU 40.0 20.0

2 2
60.0
1

0 1 2 3 4 5 6 7 8

Fig. 7.11 Gantt chart of operation schedule B at nominal demand.

0.06

0.05
Expected Positive Deviation

0.04

0.03

0.02

0.01

0
36.9 36.95 37 37.05 37.1 37.15
Expected Cost

Fig. 7.12 Pareto optimal surface for Example 3.

234 7 Modeling and Optimization of Refinery Operations Considering Uncertainty

7.5
Summary and Future Work

This paper deals with refinery scheduling problems with uncertainty considera-
tions. In particular, the crude oil unloading and mixing, and gasoline blending
and distribution problems in the refinery system are studied. Two approaches are
proposed to address the problem of uncertainty. The first approach is multiobjec-
tive robust optimization, where NBI technique is utilized to solve the multiobjec-
tive model and successfully form Pareto optimal surface that captures the trade-
off among different objectives. The second approach is to solve the corresponding
pMILP model of the refinery scheduling problems, so as to determine the opti-
mal operation schedule and objective value with respect to uncertain parameters.
Three cases studies are presented to illustrate the applicability of the proposed ap-
proaches.
Work is currently under development to improve the efficiency of the multiob-
jective framework by advancing the scenario selection process, so that the compu-
tational complexity is not increased substantially compared with the determinis-
tic problem. In addition, work is under progress to further develop the proposed
pMILP approach by improving the retrieval and storage of the required dual in-
formation of the leaf nodes in the branch and bound tree. This will enable the
utilization of the approach in large scale realistic problems.

Acknowledgment

The authors gratefully acknowledge financial support from the National Science
Foundation under the NSF CAREER program CTS-9983406.

References

1 Basset, M. H., Perkny, J. F., Management Science, Houston, TX,

Reklaitis, G. V., Obtaining realistic 1979.
production plans for a batch produc- 5 Das, I., Dennis, J. E., Normal bound-
tion facility. Comput. Chem. Eng. 21 ary intersection: a new method for
(1997), p. S1203. generating Pareto optimal points in
2 Bechtel, PIMS (Process Industry Mod- multicriteria optimization problems.
eling System) User’s manual. Houston, SIAM J. Opt. 8 (1998), p. 631.
TX, Bechtel, 1993. 6 DeWitt, C. W., Lasdon, L. S., Waren,
3 Bonfill, A., Espuna, A., Puigjaner, A. D., Brenner, D. A., Melhem, S.
L., Addressing robustness in schedul- A., Omega: an improved gasoline
ing batch processes with uncertain blending system for Texaco. Interfaces
operations times. Ind. Eng. Chem. Res. 19 (1989), p. 85.
44 (2005), p. 1524.
7 Glismann, K., Gruhn, G., Short-
4 Bonner, Moore, RPMS (Refinery and term planning of blending processes:
Petrochemical Modeling System): A Sys- scheduling and nonlinear optimiza-
tem Description, Bonner and Moore
 References 235

tion of recipes. Chem. Eng. Tech. 24 18 Li, W. K., Hui, C. W., Li, P., Li, A. X.,
(2001), p. 246. Refinery planning under uncertainty.
8 Glismann, K., Gruhn, G., Short- Ind. Eng. Chem. Res. 43 (2002), p.
term scheduling and recipe optimiza- 6742.
tion of blending processes. Comput. 19 Lin, X., Janak, S. L., Floudas, C. A.,
Chem. Eng. 25 (2001), p. 627. A new robust optimizaton approach
9 Horst, R., Tuy, H., Global Optimiza- for scheduling under uncertainty: I.
tion: Deterministic Approaches. 2nd Bounded Uncertainty. Comput. Chem.
edn. Springer, Berlin, 1993. Eng. 28 (2004), p. 1069.
10 Jia, Z., Ierapetritou, M. G., Kelly, 20 Mulvey, J. M., Vanderbei, R. J.,
J. D., Refinery short-term scheduling Zenios, S. A., Robust optimization
using continuous time formulation: of large-scale systems. Oper. Res. 43
crude oil operations. Ind. Eng. Chem. (1995), p. 264.
Res. 42 (2003), p. 3085.
21 Neiro, S. M. S., Pinto, J. M., Multi-
11 Jia, Z., Ierapetritou, M. G., Mixed-
period optimization for production
integer programming model for
planning of petroleum refineries.
gasoline blending and distribution
Chem. Eng. Commun. 192 (2005), p.
scheduling. Ind. Eng. Chem. Res. 42
62.
(2003), p. 825.
12 Jia, Z., Ierapetritou, M. G., Efficient 22 Pinto, J. M., Joly, M., Moro, L. F. L.,
short-term scheduling of refinery op- Planning and scheduling models for
erations based on a continuous time refinery operations. Comput. Chem.
formulation. Comput. Chem. Eng. 28 Eng. 24 (2000), p. 2259.
(2004), p. 1001. 23 Rigby, B., Lasdon, L. S., Waren,
13 Jia, Z., Ierapetritou, M. G., Short- A. D., The evolution of Texaco blend-
tem scheduling under uncertainty ing systems – from omega to star-
using MILP sensitivity analysis. Ind. blend. Interfaces 25 (1995), p. 64.
Eng. Chem. Res. 43 (2004), p. 3782. 24 Ryoo, H. S., Sahinidis, N. V., Global
14 Jia, Z., Ierapetritou, M. G., Generate optimization of nonconvex NLPs and
Pareto optimal solutions of schedul- MINLPs with applications in process
ing problems using normal bound- design. Comput. Chem. Eng. 19 (1995),
ary intersection technique. Comput. p. 551.
Chem. Eng. 31 (2007), p. 268.
25 Singh, A., Forbes, J. F., Vermeer,
15 Jia, Z., Ierapetritou, M. G., Un- P. J., Woo, S. S., Model-based real-
certainty analysis on the right-hand- time optimization of automotive gaso-
side for MILP problems. AICHE J. 52 line blending operations. J. Process
(2006), p. 2486. Control 10 (2000), p. 43.
16 Lee, H., Pinto, J. M., Grossmann,
26 Vin, J. P., Ierapetritou, M. G., Ro-
I. E., Park, S., Mixed-integer linear
bust short-term scheduling of multi-
programming model for refinery
product batch plants under demand
short-term scheduling of crude oil
uncertainty. Ind. Eng. Chem. Res. 40
unloading with inventory manage-
(2001), p. 4543.
ment. Ind. Eng. Chem. Res. 35 (1996),
p. 1630. 27 Zadeh, L., Optimality and nonscalar-
17 Li, W. K., Hui, C. W., Hua, B., Tong, valued performance criteria. IEEE
Z. X., Scheduling crude oil unload- Trans. Autom. Control 8 (1963), p. 59.
ing, storage and processing. Ind. Eng.
Chem. Res. 41 (2002), p. 6723.
 237

8
Production and Inventory Planning for Stock Preparation in the
Tissue Paper Industry
Joakim Westerlund, Mattias Hästbacka, Jarkko Kaplin, Tapio Westerlund

Rising competition in European tissue markets from retailers’ private labels, in-
creased global competition, changes in the energy market, and the concern for
the effects of climate change are forcing tissue paper manufacturers to focus on
key business areas and top performing brands. Furthermore, the enhancement of
supply chain efficiency as well as the optimization of production performance at
existing production facilities is becoming more and more important in this tough
business environment. This chapter presents an industrial case from the Finnish
tissue manufacturing industry, where a mixed integer linear programming (MILP)
based production and intermediate storage planning tool has been developed. The
objective of the tool is to form a knowledge-based decision-making platform, as-
sisting the production organization at Metsä Tissue Mänttä tissue paper mill to im-
prove their operational efficiency. Production-related decisions previously based on
educated guesses and experiences are minimized using the proposed tool, and the
production planning procedure is made more efficient. The mixed integer strate-
gic planning tool (MISPT) may, additionally, be utilized as a strategic support in
investment or production strategy decision-making related to the stock preparation
section of the mill. The MISPT tool is based on a MILP mixed-time scheduling for-
mulation, implemented in a tailored software acting as an access point to various
information systems at the mill. The MISPT tool is, furthermore, integrated into
a user-friendly interface and is concerned with the production and raw material
cost minimization, as well as intermediate storage utilization optimization at the
stock-preparation section of Mänttä mill.

8.1
Introduction

With an annual production value in the EU-25 of some 550–600 BEUR, forest-
based industries account for about 8% of the total value added in the manufactur-
ing industry in the EU [1]. Tissue paper products are one of the biggest nonfood

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
238 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

groups in the daily consumer goods sector, and the European tissue market, repre-
senting around 23% of the global market, is approximately turning over 10 BEUR
annually and is growing at a rate of around 3% [2]. Furthermore, the total Euro-
pean installed tissue production capacity is approximately 7 145 000 ton. 75% of
the value is sold via retail trade and the rest is sold to the AFH (Away From Home)
or “business-to-business” market. The AFH market segment is composed of care
places, hotels and restaurants, industry and more.
The Nordic countries consume rather high levels of tissue paper, with an an-
nual tissue consumption of 15 kg/cap. The European mean tissue use is, however,
relatively low compared to North America where the consumption per capita is ap-
proximately three times as much. Nevertheless, European tissue paper consump-
tion is steadily increasing clearly due to Central Eastern and Eastern European
countries [2].
A wide variety of tissue paper products are used at home, such as bathroom tis-
sue, household towels, hand towels, and tissue handkerchiefs, and as AFH prod-
ucts. Tissue paper is made of one or many plies of paper, depending on the end use.
The tissue furnish consists of bleached hardwood and softwood kraft; additionally,
an increasing amount of recycled paper is also used in tissue making [3].
The operating margins of European tissue manufacturers have been reduced and
large volume low margin, but high quality, products have rapidly been introduced
in the market. At the same time private label products are gaining more and more
market share [4]. Furthermore, global competition and rising energy prices are un-
likely to get any easier over the next 25 years, forcing the tissue paper industry to
continuously improve and optimize their way of doing business. The Integrated
Pollution Prevention and Control (IPPC) directive [5], which, further, will come
into effect in 2007, obliges paper mills to implement the best available techniques
to reduce their environmental impacts. This exigent environment makes a prof-
itable operation very hard for the European tissue paper manufacturing industry,
if it does not maintain and enhance its competitiveness. The locally acting tissue
manufacturers are, however, less influenced than is the paper industry in general,
by the fact that European and North-American forest-based companies are under
strong competition from parts of the world where forests grow faster, production
costs are lower and markets are expanding quicker.
Being one of the biggest suppliers of tissue paper products to households and
consumers in Europe, Metsä Tissue is continuously subject to tough competition
and a turbulent market situation, forcing the company to continuously improve op-
erational performance and supply chain efficiency. To maximize their operational
efficiency and profitability, the continuous development of production solutions
and the optimization of material flows are important factors. Rising demand on
production efficiency, just-in-time delivery, decreased work-in-progress inventory,
etc., is making manual production planning and scheduling a nearly impractica-
ble task. Furthermore, as the paper (or tissue) machines, used to produce the end-
products in the paper industry, are bought from the same manufacturers supplying
everyone in the industry, the competition is fierce and equipment utilization can
make the difference between profit and loss [6].
 8.2 Mathematical Programming as a Tool for Production Optimization 239

In order to manage the rising supply chain complexity, already at mill level, ad-
vanced production planning and scheduling tools play a key role. This chapter
presents an industrial example where advanced MILP-based production planning
and scheduling methodology has been applied to a challenging industrial case,
resulting in a user-friendly interactive software supporting the production organi-
zation at mill level.

8.2
Mathematical Programming as a Tool for Production Optimization

Papermaking is a vast multidisciplinary technology that has expanded tremen-

dously in recent years. Significant advances have been made in all areas of paper-
making, including raw materials, production technology, process control, and end
products. The complexity of the processes, the scale of operation, and production
speed leaves little room for error or malfunction [7]. To minimize losses due to hu-
man error or limitations in the size of problems solvable by a single human being,
mathematical programming-based production planning and scheduling may be of
great help. Production planning and scheduling has also been a very hot topic in
other disciplines of the manufacturing industry over the recent decades.
Scheduling deals with the allocation of scarce resources to tasks, over time,
where the resources may be machines in a workshop, runways at an airport or
crews at a construction site. Scheduling is a decision-making process with the goal
of optimizing one or more objectives [8]. In the recent decades, a significant re-
search effort has been made to develop efficient production planning and schedul-
ing tools and algorithms. Recent examples can be found in [9–11]. Other interesting
formulations can be found in [12–19]. For an in-depth analysis of the most impor-
tant optimization-based techniques and solution methods, applied to production
scheduling, the reader is directed to the review papers [20] and [21].
Despite a comparatively broad availability of models and algorithms, a large
gap still exists between published mathematical programming-based solution al-
gorithms, and industry’s need for efficient logic-based production planning and
scheduling tools. Transforming real-world decision problems into a form that may
be interpreted and solved by a computer is neither self-evident nor easy. Transform-
ing mathematical models, typically consisting of thousands of variables, into a form
that, furthermore, is interpretable by the actual end user is another challenge. It is,
thus, clear that the implementation of existing models and algorithms into an ap-
propriate framework, utilizable by the production planner is of great importance.
Moreover, user-friendliness and practicability tend to be more important than acad-
emic excellence when implementing production planning and scheduling tools in
practice. Furthermore, industrial production facilities are largely unique, making
it extremely hard to provide any easily implementable, general purpose software
for production planning and scheduling. Many actual industrial applications are,
hence, tailored software, based on special purpose models developed for a single
industrial process.
240 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Commercial software packages for production planning and scheduling are usu-
ally available as modules in enterprise resource planning (ERP) systems, such as
SAP, Oracle, and SSA global. These modules usually require a lot of tuning in or-
der to accurately be applicable to a specific production facility. Furthermore, the
overall implementation costs of such systems are very high. The financial benefit
of ERP system implementations is, consequently, a topical and complex issue. An
extensive study of the impact of enterprise systems (ES) on corporate performance
may be found in [22]. Although the results reported in [22] were not uniformly pos-
itive across different ES systems, they are encouraging in the sense that despite the
high implementation costs, they did not find persistent evidence of negative perfor-
mance associated with ES adoption. More general information about ERP systems
and the implementation of such tools may be found in [23], and a survey on recent
research literature on ERP systems in [24].
The industrial case presented in this chapter demonstrates an implementation
of a tailored special-purpose software, developed to support stock preparation pro-
duction planning and intermediate storage handling, as a part of the tissue paper
supply chain of Metsä Tissue Corporation in Finland. The implemented software
is concerned with the production optimization of a fairly large tissue paper mill,
involving particularly complex intermediate storage policies. The software is im-
plemented in the daily production planning at Metsä Tissue Mänttä mill, where a
large number of storage towers are available at the stock preparation section. Some
of the storage towers are available for more than one specific raw material, subject
to cleaning at changes, while other towers are raw material exclusive. The fact that
intermediate products are downgraded in storage over time makes the production
planning procedure utterly complex, and calls for special purpose models to han-
dle the timing aspects. The actual production planning and scheduling software,
called mixed integer strategic planning tool, MISPT, is based on a novel mixed
time mixed integer linear programming (MILP) formulation which is described
more thoroughly in Section 8.5. The actual formulations are based on [25].

8.3
Metsä Tissue Mänttä Mill

Metsä Tissue Mänttä mill was established in 1868 and today the mill employs
over 500 persons and produces some 100,000 ton of tissue paper and 30,000 ton
of greaseproof paper annually. The Mänttä mill consists of two stock preparation
lines, three tissue paper machines (TM), two greaseproof paper machines and sev-
eral intermediate storage towers and converting lines. The number of converted
end-products is roughly 100 and the raw material feed is based on three main
types of raw material: virgin fiber, mixed office waste, and old newspaper. The
end-products are segmented into bathroom tissue, household towels, hand tow-
els, industrial wipes, tissue handkerchiefs, baking, and cooking papers. The main
markets are Finland and Continental Europe.
 8.3 Metsä Tissue Mänttä Mill 241

At the mill, recycled fiber is processed through a stock preparation process where
ink is separated from the fibers before the pulp is processed through the tissue pa-
per machines and converting lines in order to produce high quality end-products.
The stock preparation lines are used to produce de-inked pulp from recycled paper,
where the primary tasks are the removal of contaminants and the elimination of
their effects to meet quality requirements. Separation processes vary according to
the type and quantities of contaminants of removal. The three tissue paper ma-
chines, TM1, TM9, and TM10 produce base paper to the converting lines.

8.3.1
Process Description

Tissue paper is, regardless of its apparent simple nature, a very complex product,
and the production procedure is, accordingly, rather complex. The tissue manufac-
turing process at Metsä Tissue Mänttä mill consists of a multiproduct facility of the
flow-shop type, with all products requiring more or less the same basic production
structure. More specifically, the process consists of three main production stages.
Each stage features a number of continuous lines in parallel, two stock prepara-
tion lines and the pulp center at the first stage, three tissue paper machines (TM)
at the second stage, and several converting lines at the third stage. All production
lines are limited to one quality at a time. The MISPT tool is concerned only with
the two first production stages. The first stage being subject to actual production
planning and scheduling, and the second step providing in-data for the problem.
In practice, this means that the MISPT tool provides the production planner with
optimal production plans and storage utilization profiles for the stock-preparation
section, using the existing production plans for the TM as in-data.
The basic raw material of tissue production is wood fiber, whether used directly
or recycled from waste paper. Recycled fiber plays a very important role in the pa-
per industry today as a substitute for virgin pulps [26]. By the efficient utilization
of recycled fiber as a substitution for virgin fiber, the tissue process is both more
economical and more environmentally friendly. Today the use of recycled fiber is
so important that the paper industry could not exist in many countries without
recycled fibers. In Central, Southern, or Western Europe, this is mainly due to eco-
nomic reasons. In places such as Japan, Korea, Taiwan, or Mexico, the additional
reason is the scarcity of native pulpwood resources [26]. Generally, fiber can be re-
cycled several times before it loses its paper-making properties [2]. Metsä Tissue
Mänttä mill has become the first tissue mill in the world to be granted the right
to use the PEFC (Program for the Endorsement of Forest Certification Schemes)
label. This label tells customers and consumers that the mill’s products are manu-
factured from wood obtained from forests that are managed in a sustainable way.
The right to use the PEFC label concerns all tissue and bleached cooking papers
manufactured at Mänttä. The right to use the label was granted by the Finnish
Forest Certification Council in December 2004.
Three different types of raw materials are processed at Mänttä mill, as mentioned
earlier. Virgin fiber (VF) mixed office waste (MOW) and old newspaper (ONP).
242 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.1 Conceptual flow sheet of the Mänttä mill.

MOW is recovered paper consisting of mainly wood-free fibers. ONP is recovered

newsprint and magazines consisting of mainly wood-containing fibers. In total, the
considered production facility uses six different raw material qualities, two types of
VF, one type of MOW, two types of ONP and, additionally, purchased de-inked
market pulp.
At the first stage of the process, different raw materials, together with some recy-
cled fiber (broke) are prepared for the tissue machines. Fibers are separated from
contaminants in the stock preparation lines but, inevitably, some fibers are lost as
sludge and rejects. The ratio between the inlet flow of material into an equipment
unit and its outlet flow is called the fiber yield. Fiber losses are lower when using
virgin than recycled fiber due to its more homogenous nature, cleanliness and fiber
properties. The use of recycled fiber increases losses by a factor that depends on the
fiber properties and amount of contaminants in the raw material. The overall pro-
duction flow is presented in Fig. 8.1.
After stock preparation, different intermediates are stored in intermediate stor-
age towers. Intermediate storage capacity is available for all raw materials. A more
detailed process flow-sheet of the first two stages of the production facility is pre-
sented in Fig. 8.2. Due to microbiological activities, the quality of the intermediates
suffers if they are kept in storage for too long. The high quality-standards of Metsä
Tissue are, hence, limiting the duration of time that intermediates may be kept
in storage. The time limit alters with the changing surrounding temperature and,
thus, changes according to the current season. To limit any loss of quality, preser-
vatives may be added to the intermediate storages, prolonging the storage times,
but at an additional costs.
In the second stage, production intermediates are processed in the tissue paper
machines in order to obtain different product qualities. It is worth noticing that a
modern tissue machine is not a single concept or design, as each product category
needs an optimum concept [3]. Different product qualities are, hence, produced on
different tissue machines. Some product categories may, however, be produced at
more than one machine while other products are machine exclusive. The fact that
no general modularized production concept, with identical production steps, exists
in the tissue industry, makes production planning and scheduling a very demand-
ing task. In order to obtain efficient production plans, the production planner, at
 8.3 Metsä Tissue Mänttä Mill 243

Fig. 8.2 Flow sheet of stock preparation section and tissue machines at the Mänttä mill.

each production facility, requires an in-depth knowledge of the specific production

process, the produced end-products, and utilized raw material.
Broke is generated when processing fibers on the TM and at the converting lines
following the TM. During a normal operation, the major part of the broke is gen-
erated in the converting process. When generated, broke is circulated back into the
process and some of the processed fibers are lost. Furthermore, broke is generated
on the TM whenever the process changes from one product category to another.
When tissue fibers are dried in the tissue machine, forming the actual tissue base
paper, each tissue reel that is discharged from the TM generates an amount of
broke. Broke is generally recirculated back into the process. Broke generated at
the TM is often circulated back to the wet end of the TM where it is inserted into
the process. When large rolls of tissue paper broke are formed, it is, however, not
possible to convert the material or insert it to the wet end of the TM.
To insert large rolls back into the process requires equipment to cut the rolls,
to separate the paper from the core, and equipment to dissolve the paper back
into fibers. If wet-strength agent has been applied during processing in the TM,
a conventional pulper is not enough to dissolve paper into fiber in a reasonable
time. Broke rolls from TM and converting are therefore transported to the stock
preparation and then recirculated back into the process. Taking into account that
roughly 10% of the produced capacity ends up as broke, it is obvious that the ef-
ficient utilization of the broke is a matter of great relevance [27]. The MISPT tool
is also concerned with the optimal utilization of the generated broke as a valuable
raw material. The production plans proposed by the MISPT tool are, consequently,
maximizing the broke utilization at the stock preparation lines, subject to prede-
fined capacity and quality limitations.
244 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

The intermediate stock products include qualities DI 60, DI 70, and DI 80 where
60, 70, and 80 respectively stand for the brightness in %ISO, while DI stands for
De-Inked. The amount of long and short virgin fiber made of birch and pine is
higher in the production of high-end products, whilst the recycled raw materials
are used to a greater extent in products with a lower brightness. MOW is used
to produce brighter recycled qualities and ONP for darker recycled qualities. Most
end-products are a mixture of 2–5 raw materials, making the production planning
procedure bewilderingly complex. The optimal production strategy for all products
should take into account the chemistry in the water loops of the stock preparation
and minimize fiber losses.
As in other industries such as food, fine chemicals, and pharmaceutical,
changeovers are an important issue. In this case, changeover costs are associ-
ated with the stock preparation lines and are sequence dependent. Changeovers
from dark to bright qualities are more devastating to the production efficiency and
overall profitability than bright to dark. This is because darker raw materials may,
generally, be substituted by brighter qualities whereas lighter qualities may not be
substituted by darker ones. Substituting a dark raw material with a lighter one,
however, adds costs to the production. Hence, changeovers from dark products
to lighter products, on the stock preparation, are penalized more harshly in the
objective function of the optimization algorithm than the reverse way.

8.3.2
Corporate Supply Planning at Metsä Tissue Corporation

The Metsä Tissue Corporation has production at two mills in Sweden, three mills
in Germany, one mill in Finland, two mills in Poland, and one mill in Slovakia.
Strategic product and capacity allocation and planning, as well as the coordination
of shipping and sales to markets, and inventory-cost minimization are performed
at the corporate level. The planning and scheduling of production campaigns, as
well as production control is, on the other hand, performed at the mill level. Some
of the products may be allocated to more than one mill while other products are
mill-exclusive. Likewise, some products are sold only at a specific market while
others are more global. The MISPT tool presented in this chapter is only concerned
with the operative mill level planning and scheduling for a single mill. A schematic
picture of the corporate supply planning activities at the Metsä Tissue Corporation
is presented in Fig. 8.3.
Metsä Tissue started the implementation of a common, tactical, demand, and
supply planning process in 2003. One of the critical issues in this process was to
change the focus from operative planning on mill level to country/region level. To-
day capacity planning is based on a multistep process generating a feasible plan,
including the evaluation of capacity at mill level, with subsequent adjustments
of production location, capacity, and production portfolio [28]. The computational
complexity is, however, often too large to integrate strategic planning, operative
production planning and scheduling, as well as production control activities for
all mills into a centralized activity in the enterprise resource planning (ERP) sys-
 8.4 The MISPT Tool 245

Fig. 8.3 Corporate supply planning at the Metsä Tissue Corporation.

tem. The implementation of a centralized tactical-level planning tool is, however,

an important step toward an actual integration of supply chain management.

8.4
The MISPT Tool

The MISPT tool is used to produce order-driven production plans forming optimal
short-term production schedules for the stock preparation at the Mänttä mill. A de-
tailed production plan for the TM is generated by the mill production planner and
contains due and release dates for orders, tonnages, and maintenance stoppages.
The MISPT tool is concerned with the optimal running of the stock preparation
section, including intermediates storage utilization to ensure that the correct raw
materials are available at the right time for the TM, requiring minimal costs. It is
worth mentioning that the production plans for the TM are generated manually
by the production planner, resulting in suboptimal production strategies due to
ungraspable combinatorial complexity. The incorporation of the TM in the MISPT
would, consequently, be a natural next step in the supply chain optimization strat-
egy of Metsä Tissue.
Advanced process control has yielded great benefits in the pulp and paper indus-
try during the recent decades as continuously controlling complicated processes
optimally is very difficult for an operator. The operator may manage it for a time but
usually lapses in attention in the long run [29]. This is very much the same when
it comes to planning and scheduling. Difficult operations and design and planning
problems that repeat after long periods may require computerized support. Oth-
erwise the human operator may not recall the best solutions available [29]. It is,
furthermore, not unusual that most managers at a production facility only know a
small part of their overall system. Computer-based software tools give them a way
to see the big picture [30].
246 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.4 MISPT tool architecture.

Situations where the production plan for the TM is subject to last-minute

changes due to raw material shortages are commonplace at the Mänttä mill. Mul-
tifaceted product recipes, high order stock, limited deinking capacity, and high-
capacity rigid production processes are factors that significantly add complexity to
the production planning process. Last minute changes in the production demand
tend to result in unnecessary changeovers, substitution of a needed raw material
by a significantly more expensive one, or even production stoppage due to the lack
of raw material. It is patently clear that these last minute changes may be devastat-
ing for production profitability. A logic-based tool was, therefore, created to support
daily production planning. The objective of the tool is, in short, to ensure that the
correct raw material for the TM is available, at the right time; simultaneously min-
imizing production and inventory costs related to the stock preparation section.
The MISPT tool is implemented in a user-friendly graphical interface, providing
the production planner with a controlled environment and a better overview of the
overall production process.

8.4.1
The MISPT Architecture

In Fig. 8.4, the architecture of the MISPT tool is presented. Generally, the MISPT
tool acts as an access point to different subparts, and combines them into a user-
friendly, intuitive application. The actual application consists of several intercon-
nected parts. The enterprise resource planning (ERP) system and the mill control
system, item 5 in Fig. 8.4, are, however, not fully integrated into MISPT.
Currently, the production planner manually imports the production program and
demand, for the TM, from an external resource, and thereafter loads the data into
MISPT. The raw material demand at the TM is translated into a set as hard con-
straints into the model. Storage levels and the current production state are also im-
ported manually. The module for the internal data representation, item 2 in Fig. 8.4,
transforms the data to input for the MILP model generator module, ModGen, item
3 in Fig. 8.4, which creates the MILP-based problem. A typical problem consists of
tens of thousands of variables, of which some thousand are binaries. The graphical
 8.4 The MISPT Tool 247

Fig. 8.5 MISPT tool at a glance.

user interface (GUI), item 1 in Fig. 8.4, is used for input and visualization, and
the commercial MILP solver ILOG CPLEX item 4 in Fig. 8.4, is interactively used
by the user via the GUI. Figure 8.5 presents a simplified version of the MISPT
architecture.
In order to achieve production plans that are practicable, the scheduling func-
tion has to interact with other decision-making procedures used at the mill. The
material requirement planning (MRP) system is one key component that need be
interacting with the scheduling function. After a schedule has been set up, all the
raw materials and resources must be available at specified times [31]. The connec-
tion between the production planning and the MRP should be working at all stages
in the production (stock preparation, tissue machines, and converting). The ready
dates of all orders have to be determined jointly by the production planning and
scheduling system and the MRP system. This is due to the fact that MRP systems
alone ignores capacity constraints, assumes that lead times are fixed, and does not
consider the operation sequences of items [32]. A traditional top-down hierarchical
approach of a general MRP and scheduling system, presented in [32], is demon-
strated in Fig. 8.6. In the multilevel product structure, BOM stands for Bill Of
Materials, describing a product in terms of its assemblies.
Recent developments in the supply of enterprise resource planning (ERP) sys-
tems and related software have also introduced a myriad of different IT-tools, mod-
ules, and systems, developed to rationalize and integrate computational networks
and supportive tools in the manufacturing industry. According to [33], a planning
248 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.6 Traditional MRP and scheduling system structure.

and scheduling system is located between the ERP and manufacturing execution
systems (MES), corresponding to the MRP in [31]. In this particular case, the MES
can be considered as a module of the ERP system. The MES acquires and reports
production-related data from the production, while the MISPT tool acquires the
production-related data through the ERP interface. The contact between the MISPT
tool and the ERP system is, nevertheless, not fully integrated in this tool. To make
the actual planning and scheduling more efficient, and simultaneously maximize
end user practicability, the MISPT tool should be fully integrated into the ERP sys-
tem as a separate interactive module.

8.4.2
The MISPT Graphical-User Interface

Much effort has been put into the usability of this tool. The end users actively took
part in the design and development of the GUI during the development and inte-
gration process. The goal was not to substitute the role of the production planner,
but to give her/him a tool which would produce precise facts to base decisions on.
In short, to minimize decisions based on educated guesswork. With software capa-
bilities increasing rapidly, information technology is no longer the barrier as much
as human and organizational barriers [33].
In Fig. 8.7, the GUI of the MISPT tool is presented. The result from the opti-
mization is transformed mainly into charts and tables, giving valuable information
about the solution. The user can easily compare different charts to get a better
holistic picture of the complex production planning problem. The GUI window is
split into different parts to instantly give a good general overview of the current
production plan for the whole production process. The core of the MISPT tool is,
obviously, the mathematical programming model and the solutions it provides, but
without a user-friendly interface the result would be impossible to utilize in prac-
tice.
 8.4 The MISPT Tool 249

Fig. 8.7 MISPT graphical-user interface.

8.4.3
The MISPT Tool for Daily Production Planning

If no significant changes are made to the production plan for the TM, the produc-
tion planner should run the MISPT tool no more frequently than once a week. In
practice, the production plan for the TM is, however, subject to frequent change,
making the MISPT tool a daily supportive tool. In order to maintain good quality
end-products and minimize the amount of fluctuation in production capacity and
quality at the stock preparation, the production planner should run the optimiza-
tion approximately every morning with a horizon that is longer than the actual fixed
production plan for the TM. This procedure will eliminate possible rapid changes
between each optimization, as the production plans from the MISPT will form
a rolling, yet overlapping, time horizon. The model will, through this procedure,
take possible future production into account at an early stage, providing a kind
of prediction of coming production plans, making production smoother. Utilizing
the “rolling overlapping time horizon” subsequent iterations will start from known
initial conditions.
Rapid changes in the production plan may, as mentioned before, have a great
impact on the overall production efficiency and profitability. The MISPT tool has
been proven very useful, for example, at start-ups after longer production stops
when storage levels are low and the demand at the tissue machines is still opti-
mized for normal production with normal storages levels. The user may, in such
cases, preset the desired storage levels at the end of the optimization horizon as
input in the actual optimization. The desired storage levels, at the end of the plan-
ning horizon, are not hard constraints and the optimization will, therefore, only try
to fulfill them. If the desired storage levels are not reached, the objective function
will slightly be penalized.
250 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

8.4.3.1 Planning of Production Stops

Maintenance stops are of great importance in order to maintain good product qual-
ity, prevent equipment failure, and unscheduled downtime. Machine failure and
other critical issues that may occur if regular service and maintenance are not per-
formed, may affect not only a single isolated part which is directly affected by the
failure, but it may influence the whole production process.
The efficient planning of service and maintenance stops is not an easy task if the
effect of the shutdowns, on the total production, should be minimized. The MISPT
tool offers an easy way to plan service and maintenance stops. The production plan-
ner is able to, initially, make a first optimization run and then evaluate where a
stop would have the smallest impact on production. The production planner then
decides where the stops are held and reoptimizes the problem. The MISPT tool
subsequently recalculates the optimal production plan for the considered produc-
tion horizon, taking into account the service or maintenance stops. The length of
the stops should always be at least the length of the discretization grid (usually
one production shift, 8 h). This fact is, obviously, limiting the applicability of the
MISPT, but is a formulation-related constraint. It is, however, worth noticing that
service and maintenance stops of less than one shift (8 h) are, in practice, rare.

8.4.3.2 Aging Profiles

One of the main features of the optimization model used is the ability to handle
the aging of intermediate products in storage, and sketching up aging profiles for
those storages. Aging profiles are, in the MISPT tool, presented through charts,
automatically generated for each planning horizon. This is very useful since raw
material should always be produced just-on-time. The optimization will, thus, al-
ways try to minimize the amount of intermediate products that have been stored
in the intermediate storage towers for more than a predefined time. At the begin-
ning of each planning scenario, the user may estimate how large a percentage of
the substance in storage should be regarded as having been in storage for a longer
period of time than the proposed time limit. This predefined time limit is specified
by the end user.

8.4.3.3 Time Synchronization

The user can choose the length of the discretization grid and the length of the
planning horizon for each specific case. To improve user-friendliness and ease,
the implementation of production plans, the time-line will be synchronized to real
work shifts at the mill and the model, additionally, adjusts the length of the first grid
to fit the real world. Production shifts at Mänttä mill are 8 h long and start at 06:00,
14:00, and 22:00. If an optimization run is started at, for instance, 11:10, the MISPT
tool synchronizes it to the actual production shift schedule by adjusting the length
of the first shift to start at 11:10 and end at 14:00. After 14:00, the production plan
will again follow the 8-h shift schedule. The optimization can start at any given time
and the discretization grid can be 2, 4, 8, or 24 h, while the planning horizon can
be 1 day to 2 weeks. Current demand at the TM will automatically be transformed
into discrete bounds in the model. The length (and production capacity) of the
 8.4 The MISPT Tool 251

first grid will be adjusted to fit any combination of the defined grid length and the
optimization horizon.

8.4.4
The MISPT Tool for Strategic Planning

The user can change global parameters, such as production unit specific produc-
tion capacities, intermediate storage capacity, and changeover costs at stock prepa-
ration to adjust the optimization. The default purpose of the MISPT tool is to pro-
vide a general platform for daily production planning and scheduling. The MISPT
tool may, however, also be utilized as a strategic tool for investment decision sup-
port. The user can employ the tool in different ways, for example, to evaluate and
simulate new structural investments and new production strategies. In addition to
the ability to change storage capacities and different production-related capacities
through a dialog window, the user can also add a new stock preparation line or
adjust raw material prices to evaluate new production strategies.
A number of strict logical constraints for some parts of the process may also be
replaced by more flexible ones to enable the evaluation of new strategic investments
to improve the process. Some bottlenecks may not be visible in daily production,
but through simulations of new possible improvements, without undergoing ex-
pensive costs, the overall production may be improved in the long run due to more
flexible production strategies. Figures 8.8(a) and (b) and 8.9(a) and (b) present the
ability to use a more flexible storage utilization as a part of strategic planning in
the MISPT tool.
In Figs. 8.8(a) and 8.9(b), the figure on the right is more flexible, since the filling
and emptying of the storage is done using different pipes, and the storages are,
hence, independent from each other.

Fig. 8.8 Flexible storage setup on the right and standard setup on the left.
252 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.9 Flexible storage setup on the right and standard setup on the left.

8.5
The Mixed-Time MILP Model

The presentation of events over time is usually the first aspect needing be taken
into consideration in the problem formulation procedure of scheduling problems.
Roughly, there are two frequently used ways to represent time in scheduling prob-
lems: discrete or continuous. Generally, a discrete time representation divides the
time horizon into a number of uniform time sequences and only allows events to
take place at discrete time points (at the beginning of the time intervals). A contin-
uous time model allows events to take place continuously at any time during the
time horizon.
Recent developments in the scheduling community are to a large extent con-
cerned with continuous-time representations, involving, for example, single [33]
and [18], or multiple time grids [35] and [10]. Continuous-time models are, gener-
ally, outperforming their discrete counterparts regarding computational efficiency.
There are, nevertheless, also some very useful features linked to discrete time for-
mulations. A discrete time formulation is, for example, very practical when mod-
eling nonlinear production profiles in multistage scheduling problems with com-
plex intermediate storages, as is the case at the Metsä Tissue Mänttä mill. Using a
discrete-time model one is able to generate piecewise linearizations of highly non-
linear storage profiles. Further information concerning discrete versus continuous-
time formulations can be found in [21], [20], and [8].
The mixed-time MILP formulation used in the MISPT software is a hybrid-time
formulation using a global uniform time grid but at the same time enabling events,
such as changeovers, to take place at any given time within any of the time grids
during the scheduling horizon. In practice, production plans generated with the
mixed-time model are continuous-time production plans, with some limitations.
This procedure makes the proposed model significantly more flexible than a tradi-
tional discrete-time model, Fig. 8.10, where changeovers may only take place at the
beginning/end of a time grid. The difference between a discrete and a continuous-
time model is presented in Figs. 8.10 and 8.11.
 8.5 The Mixed-Time MILP Model 253

Fig. 8.10 Discrete-time production plan.

Fig. 8.11 Continuous-time production plan.

Fig. 8.12 Mixed-time production plan.

The main novelty in the mixed-time formulation is the ability to make product
changeovers at any time within the scheduling horizon, not only at discrete time
points. This procedure is a requirement to properly address the industrial problem
presented in this chapter. A mixed-time production plan is presented in Fig. 8.12.

8.5.1
Mathematical Formulation

This section gives a short overview of the mathematical MILP formulation used in
the MISPT tool. For a more detailed description, the reader is referred to [25]. The
mixed-time formulation may be described as an extension of a traditional discrete-
time formulation, the main novelty being the ability to make changeovers at any
time within the scheduling horizon, not only at discrete time points. The formula-
tion is grid based and the production of a specific product, i, at machine k during
time grid t is described in Eq. (8.1), using two continuous variables, wi,k,t and pi,k,t ,
where a traditional discrete-time model uses one discrete and one continuous vari-
able.
pi,k,t
wi,k,t · pi,k,max ∀ i ∈ I, k ∈ K, t ∈ T , (8.1)
where pi,k,t represents the momentary production during time grid t and wi,k,t
provides a fractional upper limit for the momentary production and is thereby
bounded to take values between 0 and 1. pi,k,max is the global maximum produc-
tion rate for task i at unit k during one grid. It is worth noticing that the wi,k,t
variable may be 1 although the actual momentary production is not at 100%. wi,k,t
only gives a limit to the maximum allowed momentary production as seen in Eq.
(8.1). The wi,k,t variable is determined by the binaries yi,k,t and yi,k,t+1 as shown
in Eq. (8.2). A changeover may, thus, take place anywhere (any time) within a grid
and not only at the beginning or end of a grid. The conditions for wi,k,t are given
below.
wi,k,t
yi,k,t + yi,k,t+1 ∀ i ∈ I, k ∈ K, t ∈ {1, 2, 3, . . . N − 1}. (8.2)
254 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

At the end of the time horizon, during time grid N, the wi,k,t variable is, obviously,
determined solely by the final yi,k,t variable since no yi,k,t+1 exists.

wi,k,N
yi,k,N ∀ i ∈ I, k ∈ K. (8.3)

To make sure that not more than one task is performed at any unit at any time, ex-
cept for the case when a changeover is taking place, Eq. (8.4) is used. Furthermore,
to make sure that no unit is producing anything at a level exceeding the maximum
throughput, Eq. (8.5) is used.

I
yi,k,t
1 ∀ k ∈ K, t ∈ T , (8.4)
i=1

I
wi,k,t
1 ∀ k ∈ K, t ∈ T . (8.5)
i=1

To make sure that task i may be performed at unit k during time-sequence t, if

yi,k,t is equal to 1; Eq. (8.6) is used.

yi,k,t
wi,k,t ∀ k ∈ K, t ∈ T . (8.6)

Assuming constant production rates within the grids, the exact time of a changeover
may, retrospectively to the actual optimization, be calculated using the solution val-
ues of the wi,k,t variables according to Eq. (8.7).
wi,k,t

ti,k,t = I ·
t, (8.7)
i=1 wi,k,t

where
ti,k,t is the changeover time calculated from the beginning of the time grid
and
t is the length of the specified discretization grid. The
ti,k,t variable is used
in the MISPT tool to produce, for example, continuous-time Gantt charts.
In the formulation of the sequence of dependent changeovers, xCO,ii  ,k,t is a
continuous variable equal to 1 if a changeover from task i to task i  takes place in
the beginning of time grid t; 0 otherwise. The changeover variable does not need
be specified as a binary variable since its integer relaxation will always be 0 or 1 by
using Eq. (8.8). The total number of changeovers is penalized, and consequently
minimized, in the objective function.

xCO,ii  ,k,t  yi,k,t−1 + yi  ,k,t − 1 ∀ i, i  ∈ I, i = i  . (8.8)

The formulation of the intermediate storages is of great importance in the MISPT

model and the general storage balance is given in Eq. (8.9) where the storage in-
ventory at time t is equal to the inventory at time t − 1 plus the addition of material
into storage at time t minus the subtraction of material out of storage at time t.

ms,t+1 = ms,t + ms,in,t+1 − ms,out,t+1 , (8.9)

where ms,t+1 is the amount of intermediate product in storage s at time t + 1.

ms,in,t+1 and ms,out,t+1 are the total batch sizes in and out of storage under a given
time interval. According to Eq. (8.9), the storage contents are always determined at
 8.6 Illustrative Case Examples 255

the end of the time interval. The total flow in or out for a given time period
t is
given in Eq. (8.10).
 t+
t
ms,in,t+1 = ṁs,in,t dt. (8.10)
t
This representation gives a good piecewise linearization of the actual highly non-
linear storage profiles at the Mänttä mill. Furthermore, in the particular industrial
case presented in this paper, it is not enough to know how the storage profiles
develop over time, it is also necessary to know how old the substance in storage
is. The MISPT tool handles the aging aspect by penalizing the objective function
using the constraints in Eqs. (8.11) and (8.12). tkeep represents the time, in grids,
after which the intermediate exceeds the time limit in storage. tkeep is a function of
the grid length
t and the actual (given by the end user) prespecified intermediate
specific aging time limit told .


t
mold,s,t  ms,t−told − ms,out,z told ∈ Z, (8.11)
z=t−told +1
tkeep
told = told ,
t ∈ Z. (8.12)

t
The summarized amount of “old substance”, s, during each time grid with the
length
t is penalized with a substance-specific aging penalty cost factor, pold,s , in
the objective function, where Cold,tot is the total aging penalty/cost, presented in
Eq. (8.13).


T
Cold,tot = pold,s · mold,s,t . (8.13)
t=1

8.6
Illustrative Case Examples

The following case examples will illustrate how some of the features of the MISPT
tool work in practice. The illustrative production programs for the tissue machines
are real programs from production at the Metsä Tissue Mänttä mill during the year
2006, used as in-data in optimization runs with the MISPT tool. The grid length
used in all examples is 8 h, corresponding to one production shift, and the plan-
ning horizon is 1 week. In the charts presented in this chapter, the values on the
horizontal axis illustrate grid number, whereas real dates are used in the real im-
plementation used at the Mänttä mill. The first example, being the most extensive
one, shows most of the basic features of the tool whilst the second example shows
how maintenance may be planned at the stock preparation and the intermediate
storages.
Before each new optimization run, the end user imports the current production
program on the TM from the MES through the ERP system. An illustration of
an imported production program is presented in Fig. 8.13. When the production
256 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.13 Production program for tissue machine 1, grid length of 8 h.

planner loads the continuous tissue machine production program into the MISPT
tool, it is automatically converted to a discrete-time form for each TM, according to
the chosen grid length and optimization horizon. The current production situation
at each TM is taken into account and some orders have perhaps recently started
whilst other are about to end. Since the orders are signed for the rewinders, which
are placed after the tissue machines, and the raw material is fed into the head
box at the beginning of the tissue machines, the dynamics in the reporting system
may cause some minor errors in the demand, in the beginning of the optimization
horizon. Furthermore, some reels may, in general, be finished at the TM but still
need be processed at the rewinders. If an order is nearly done but not marked as
ready, there will still be a demand on the tissue machine, despite the fact that the
raw material is not actually needed.
The demand at the TM is given per raw material per tissue machine, and the
MISPT tool splits the demand in the predefined grids according to the start and
end times of the orders (see Fig. 8.13). It is worth noticing once more that the
start and end times are hard constraints and not release- and due-dates offering
rough frames for the production planning. The smaller chunks of raw material-
need, from each order, are summarized to a combined demand for each raw ma-
terial at each grid (see Fig. 8.13). The summarized need of raw material is then
shown in charts to give an easily examinable overview of the production program
to the production planner. Each raw material is represented by its own color. An
example of the raw material needed (production programs) at the TM is presented
in Fig. 8.13.

8.6.1
Case Example 1

This case example should give the reader a general overview of the in-data needed
in each optimization run, using the MISPT tool, and the output of the MISPT. The
 8.6 Illustrative Case Examples 257

Table 8.1 Production program for tissue machine 1, grid length of 8 h.

Grid Day Start End DI 60 DI 70 DI 80 Short Long

1 1 0:00 6:00 0.0 18.5 8.0 0.0 0.0

2 1 6:00 14:00 0.0 4.2 1.8 15.2 15.2
3 1 14:00 22:00 0.0 0.0 0.0 17.9 17.9
4 2 22:00 6:00 0.0 0.0 0.0 17.1 17.1
5 2 6:00 14:00 0.0 0.0 0.0 17.1 17.1
6 2 14:00 22:00 0.0 0.0 0.0 17.2 17.2
7 3 22:00 6:00 31.0 6.6 0.0 0.1 0.1
8 3 6:00 14:00 35.0 3.9 0.0 0.0 0.0
9 3 14:00 22:00 35.0 3.9 0.0 0.0 0.0
10 4 22:00 6:00 35.5 3.9 0.0 0.0 0.0
11 4 6:00 14:00 35.8 4.0 0.0 0.0 0.0
12 4 14:00 22:00 36.2 4.0 0.0 0.0 0.0
13 5 22:00 6:00 27.4 5.4 7.1 2.2 0.0
14 5 6:00 14:00 0.0 6.8 21.4 6.7 0.3
15 5 14:00 22:00 0.0 0.0 16.2 5.1 5.1
16 6 22:00 6:00 0.0 0.0 13.7 5.4 8.9
17 6 6:00 14:00 0.0 0.0 0.0 7.8 30.0
18 6 14:00 22:00 0.0 0.0 0.0 17.9 17.9
19 7 22:00 6:00 0.0 0.0 0.0 16.2 22.0
20 7 6:00 14:00 0.0 0.0 0.0 17.9 17.9
21 7 14:00 22:00 60.0 0.0 0.0 17.9 17.9

most important in-data of this case example is presented in Fig. 8.13 and Table 8.1,
giving production shift-specific demand for each raw material. During grids 20 and
21, for example, TM 1 is demanding 17.9 ton of virgin short fiber (from birch) and
17.9 ton of virgin long fiber (from pine). The objective of the MISPT tool is now to
fulfill these, as well as all other needs over all time grids, at minimum production
and intermediate storage costs.
The corresponding production program to that presented in Fig. 8.13 is pre-
sented in Table 8.1. In Fig. 8.13, the default discretization grid of 8 h, corresponding
to the actual production shifts, is used. The discretization grid can also be set to 2,
4, or 24 h.

8.6.1.1 Sequence-Dependent Changeovers

Changeovers can, as mentioned in Section 8.5, take place at any given time within
the discrete time grids. All changeovers are considered to be sequence dependent,
meaning that it may take more time or cost more to make a changeover from raw
material A to raw material B than the other way around. A changeover from a
lighter quality to a darker quality is, in this particular case, cheaper than the reverse
way. Situations where a “changeover” from production to no-production takes place
are also taken into account. Changeovers from no-production to production and
vice versa are also minimized, making the production process more fluent. At the
marker in the Gantt chart in Fig. 8.15, one can see that a changeover occurs within
258 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.14 Simplified illustration of a changeover from, for example, DI 60 to DI 80.

the grid (production shift) and not at a discrete time point (at the beginning or end
of the shift). The MISPT tool provides the end user with the production shift and
the precise time for all changeovers. As may be seen in Fig. 8.15, the production
shifts are clearly defined in the production plans generated by the MISPT.
The changeover costs may be calculated from historical data, giving statistically
good assumptions of the fiber-loss, in terms of money, for each specific type of
changeover. Since the changeovers are performed continuously, meaning that no
actual shutdown in the production is needed at changeovers, the fiber losses oc-
cur when high-quality (more expensive) fibers are partially used to produce lower
quality intermediate products during the actual changeover. Moreover, during a
changeover, the more expensive raw material will, explained in a simplified way, be
used for the same purpose as the cheaper quality one involved in the changeover
to produce a cheaper end product. Figure 8.14 presents a changeover from DI 60
to DI 80 at the stock preparation. The vertical axis displaying the material flux and
the horizontal axis displaying time. The dotted black line presents the summarized
stable production flux while the dark-gray line presents the flux of DI 60 and the
light-gray line represents the DI 80 flux.
If a changeover from DI 60 to DI 80 takes place, the DI 80 being produced dur-
ing the changeover is “downgraded” and used as DI 60, as may be seen in Fig. 8.14.
At this point, it is worth mentioning that changeovers are rather time-consuming
events at the stock preparation and unnecessary changeovers may, consequently,
add significant costs to the production, thereby deteriorating overall profitability.
Since the more expensive raw materials are used to produce products yielding a
better end-product price, the DI 80 used to produce products that actually do not
require any DI 80 will result in loss of sales profit. The difference in raw material
price between DI 60 and DI 80 multiplied by the amount of DI 80 getting down-
graded in the changeover is, thus, the raw material related cost for the changeover
from DI 60 to DI 80. It is, hence, clear that changeovers between materials that are
significantly different from each other, and with considerably different raw mate-
rial prices, are more expensive than changeovers between raw materials that are
 8.6 Illustrative Case Examples 259

more homogenous. It is, additionally, worth noticing that no actual cleaning of the
process equipment is needed at regular changeovers at stock preparation.

8.6.1.2 MISPT Results for Case Example 1

When all in-data are gathered into the MISPT and the user has tuned the opti-
mization parameters, if needed, the actual optimization takes place. The results
are, subsequently, illustrated through charts, tables, etc. Depending on the plan-
ners’s choices of parameters, the actual MILP optimization problem may increase
or decrease in size. The discretization grid and the planning horizon are, obviously,
two of the most important factors influencing the problem size, and, thus, also the
solution time. In Table 8.2, the problem size for a 1-week planning problem using
different discretization grid is presented.
The default solution approach in the MISPT tool is to optimize the production
programs for the stock preparation line separately from the pulp center. This pro-
cedure significantly decreases the solution time as the problem size is reduced. It is
important to notice, however, that even though the stock preparation lines and the
pulp center are considered as separate production units in the MISPT optimiza-
tion, optimality is not sacrificed. Since the stock preparation lines and the pulp
center are separate units, processing separate raw materials, and using separate
intermediate storage towers, they do not have to be linked in any way in the opti-
mization. Table 8.3 presents the same problem statistics as in Table 8.2, using an
optimization approach where the production plans of the stock preparation lines
and the pulp center are solved in the same MILP problem, to demonstrate how
much the solution time may be cut down by splitting up the problem into two
separate MILP problems.

Table 8.2 Problem statistics using different discretization grid, default settings.

Variables
Days Grid (h) CPU (s) MIP gap Constraints Binary Continuous

7 8 6 Optimum 5642 525 4092

7 4 48 Optimum 11 228 1050 8103
7 2 4683 Optimum 22 387 2100 16 125

Table 8.3 Problem statistics using a simultaneous solution approach.

Variables
Days Grid (h) CPU (s) MIP gap Constraints Binary Continuous

7 8 1 57 Optimum 5978 525 4092

7 4 600a 1.02% 11 900 1050 8103
7 2 600a 2.33% 23 731 2100 16 125

a CPU time limited to 600 s.

260 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.15 Gantt chart for the stock preparation section.

Fig. 8.16 Production program for stock preparation line 1.

The results from the optimization runs are presented in different forms to re-
spond to different needs. A general holistic production program for the whole stock
preparation section is, for example, illustrated as a Gantt chart which is not as de-
tailed as, for example, a storage profile for a particular storage.

Production Programs Production programs are presented in two different ways,

as Gantt charts giving a holistic view of the whole stock preparation section (see
Fig. 8.15) or as unit-specific production programs (see Fig. 8.16).
While the Gantt chart offers the production planner a good overview of the whole
production plan, the unit-specific production programs present the actual produc-
tion throughput at each specific shift, in addition to changeover times and other
crucial information.

Storage Profiles One of the most important objectives of the MISPT tool is to pro-
duce optimal storage utilization profiles for the intermediate storages. Apart from
the actual storage profiles, the storage charts also contain aging data, highlighting
if any pulp is getting old in storage. All storages profiles are piecewise lineariza-
tions of the actual highly nonlinear profiles. A shorter grid length obviously means
more accurate linearization. Overall optimization times, nevertheless, increase very
rapidly with shorter grid length and rising combinatorial complexity. In Figs. 8.17
and 8.18, the same storage profile for one of the intermediate storages is presented
using discretization grids of 8 and 2 h, respectively.
From Figs. 8.17 and 8.18, it is clear that the overall storage utilization strategy
is the same, in this particular case, even though the chosen discretization grid is
different. One may observe that the fine linearization in Fig. 8.18 generates slightly
smoother storage profiles than in Fig. 8.17, where the linearization is rather coarse.
 8.6 Illustrative Case Examples 261

Fig. 8.17 Storage profile using piecewise linearization with 21 grids of length 8 h.

Fig. 8.18 Storage profile, grid length of 2 h.

It is worth noticing, however, that the finer the discretization, the larger the com-
binatorial complexity and the longer the solution time.

Market Pulp Utilization and Shortage of Raw Material To illustrate scenarios where
the raw material need at the TM is impossible to meet, expensive “fake raw mate-
rial” is used in the MISPT tool. The optimization will, using these fake raw mate-
rials, always reach a feasible solution even if the production program on the TM,
in practice, is unfeasible. In such cases, the model will give a solution indicating
where and when there are problems meeting the raw material demand. These so-
lutions give the production planner, as well as, the production management a good
picture of bottlenecks and a foundation for new production programs. Since the
fake raw material is significantly more expensive than real raw material, the opti-
mization will always avoid using fake raw material if it is not absolutely necessary
in order to meet the demand on the TM.
262 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.19 Own deinked pulp substituted by market pulp up to maximum 30%.

Fig. 8.20 Shortage of pulp during grids 2, 3, and 13. No substituting market pulp available.

Market pulp may be used to substitute some deinked pulp qualities. If the ca-
pacity on the stock preparation lines is not sufficient to feed the TM, market pulp
may be used at a higher raw material cost to compensate for the lack of deinked
pulp produced in-house. Using market pulp is, clearly, not the most cost efficient
solution and the model will minimize its utilization. Market pulp can, as default,
substitute up to 30% of the total raw material. This limit is a design parameter in
the model and may be changed manually by the end user. In Fig. 8.19, the need for
market pulp is visualized in a lighter color in the chart displaying the total need for
raw material DI 60. Market pulp is fed into the TM through a separate pulper, and
does not have to be processed on the deinking lines in advance (see Fig. 8.2).
 8.6 Illustrative Case Examples 263

Fig. 8.21 Aging profile. The dark part is deinked pulp which is exceeding the aging limit.

If the capacity at the stock preparation lines is not enough to meet the demand at
the TM, the pulp shortage will be indicated in the production charts in a dark color.
In practice, this means that something has to be done to the production program
on the TM. In Fig. 8.20, a scenario including shortage of DI 70 at the TM during
shifts 2, 3, and 13 is presented. As in most MILP problems, several equal optimal
solutions may exist. For example, equally good solutions may be received where
shortages are found in different grids, as long as the total amount is the same and
other dependent effects, such as aging profiles, are not affected.

Aging Profiles One main feature in the MISPT tool is keeping track of the age of
the raw material in storage. This is a very important feature, since the quality of the
raw material is of great relevance for the end-product quality in the tissue process.
Furthermore, serious money could be lost if the intermediate products, in the worst
possible case, would get too old and, consequently, be unusable as raw material.
The optimization will, hence, minimize the amount of pulp being stored in the
intermediate storages for more than a prespecified time. The end user can directly
see if, when and where there are intermediates that have been stored for a longer
time than the prespecified limit. If the MISPT tool alerts the production planner
of intermediates that are stored for too long, she/he can take actions according to
that. A scenario where an amount of pulp is stored for too long (giving rise to a
penalty cost rise) is presented in Fig. 8.21.

8.6.2
Case Example 2

The underlying MILP model includes a number of design parameters that may
have an impact on the overall optimization performance and also the solution it-
self. A good example of a typical design parameter is the “smoothing” parameter.
The model strives to make the production as smooth as possible, in the sense that
throughput in subsequent grids it should be as similar as possible, if it does not
264 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.22 Smooth production thanks to adjustable design parameters.

Fig. 8.23 Fluctuating production without smoothing.

otherwise influence the overall performance. If, for instance, 60 relative weight
units, RWU, of a certain raw material should be produced during two subsequent
grids, MISPT will propose a production of 30 RWU for each grid rather than, for
example, 10 RWU in the first and 50 RWU in the latter, if it does not influence
the overall profitability. This procedure is very useful if the production capacity is
not too tight, otherwise the model could set pauses in the production and run at
100% in the next, resulting in huge fluctuations in production capacity which is
obviously undesired in practice. Figure 8.22 illustrates the effect of the smoothing
parameter. Figure 8.23 presents the same scenarios as in Fig. 8.22 but without the
use of the smoothing parameter.
Other examples of design parameters are, for example, the penalties for sequence
dependent changeovers and allowed market pulp proportions. These parameters
may be adjusted manually in the MISPT tool before each optimization, to tune the
outcome to meet specific needs. If the changeover costs are penalized very harshly,
the minimization of changeovers may get a higher priority than the minimization
of old pulp in storage, resulting in fewer changeovers at the cost of, for instance,
more pulp getting old in storage. If the changeovers are penalized very lightly, sev-
eral changeovers may take place at stock preparation while the pulp in storage may
be kept very fresh throughout the whole planning horizon.
Utilization of storages may also be tuned to the specific preferences of the end
user. If the production planner, for example, prioritizes the utilization of only one
intermediate storage for a raw material which actually has two dedicated interme-
diate storages, the MISPT tool will minimize the utilization of one of these storages
 8.6 Illustrative Case Examples 265

by penalizing the fluxes into this specific storage. This strategy should be tuned to
be as realistic as possible, since there is more than one storage for almost all raw
materials. It is, however, very important to bear in mind that every effort to steer the
optimization in a certain direction will be undertaken at the expense of the solution
flexibility. The more flexibility the MISPT tool is given, the closer to an actual global
optimal solution it will obtain. It should, of course, be kept in mind that the model
should only be given flexibility in cases where the production process also, in prac-
tice, is flexible. If the model is not describing the actual production process as it is
in practice, the solutions may be of poor quality or even unfeasible in practice.

8.6.3
Case Example 3

Another important feature with the MISPT tool is the ability to efficiently plan
service and maintenance stops in advance. This example shows how maintenance
planning may be done in an efficient manner using the MISPT tool. The example
below shows a case where a production stop consisting of two grids in sequence
are scheduled on SP line 1.
Comparing Fig. 8.24 to the Gantt chart in Fig. 8.15 (which is a Gantt chart of the
same scenario on the TM but without any production stops on the stock prepara-
tion), one can see that the changeover from the raw material represented by the
light gray to the darker material occurs earlier when the stop is introduced, and
that it also requires one additional changeover at the end of the planning horizon.
Figure 8.25 presents the dialog window for the planning of production stops on
stock preparation lines and storages.
The production planner can also plan stops to each storage separately. The stops
at the storages can be any logical combination of no material flow into the stor-
age (In = 0 in Fig. 8.25), no flow out of the storages (Out = 0 in Fig. 8.25) and no
intermediate storage utilization allowed (Sto = 0 in Fig. 8.25). For example, if main-
tenance work is held on some pumps related to a particular storage, the optimiza-
tion will be aware of it in advance and the model is adapted to the new situation,
Fig. 8.24. The stops are listed in a table where the end user can specify each stop
separately. The two first stops in the table in Fig. 8.25 are logically equal. If no in-

Fig. 8.24 Production plan for stock preparation line 1 with a 16 h production stop.
266 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

Fig. 8.25 The logical combinations how service and maintenance stops can be planned.

termediate storage is allowed, no material flow out of the storage will be possible
per definition in the solution from the planning problem.

8.7
Discussion

In this chapter, an industrial application of mathematical programming MILP-

based production planning and scheduling is presented. The industrial applica-
tion under consideration consists of the stock preparation section at Metsä Tissue
Mänttä mill, which is a large tissue producing mill in Finland. The paper presents
an interactive production planning tool, the MISPT, which is implemented in an
interactive user-friendly graphical interface. The MISPT tool is a tailored tool, based
on a novel mixed-time MILP scheduling formulation. The underlying mixed-time
MILP optimization model is briefly presented in the paper, and the practicability of
the MISPT tool is presented mainly through a number of illustrative examples.
The prevalent challenge of the MISPT tool is the computational complexity
which rises dramatically with a finer discretization grid and longer planning hori-
zon. The MISPT tool is, however, able to reach optimal or very good solutions, in
a practical reasonable CPU-time, even with a fairly coarse discretization grid. The
fact that the final fixed production plans at the TM seldom are more than 1 week
long makes the MISPT an efficient tool for the purpose it was created to serve. In
addition to being a profitable daily operative planning tool MISPT has also been
found to be exceptionally useful in planning profitable start-ups and shot-downs.
To fully be able to take advantage of the possibilities of the MISPT tool, it should
be completely integrated into a modern ERP system supplying the MISPT with
up-to-date production data, intermediate storage levels, and so forth. Furthermore,
 References 267

implementing the production planning and scheduling of the tissue machines into
the same tool as the stock preparation would certainly yield great savings and result
in more efficient production. Implementing the TM into the MISPT tool would,
however, add more combinatorial complexity and, hence, prolong solution times.
Overall the MISPT tool handles an utterly complex industrial production plan-
ning task and is, furthermore, able to produce optimal production plans and strate-
gies in a form that is easily understandable and capable of being put into action
by the end user. For the operating personnel, at the actual mill, the MISPT tool
has led to deeper knowledge of the entire stock preparation process, and to better
understanding of critical parameters in the process. The operating personnel at
the mill has, during MISPT-supported operation, also been able to put more focus
on general operational issues rather than on detailed planning of production pro-
grams. A discernible result of the operational use of the MISPT tool has, beyond a
cost efficient operational planning of the deinking lines including the production
line scheduling and economical use of raw materials, further been a substantial
reduction of aging pulps. The quantities of aging pulps have been reduced close
to zero. In addition to a significant cost reduction this has also had a clear quality
improvement impact for the whole stock preparation section. The MISPT tool is a
result of an intimate cooperation between industry and academia.

References

1 Forest-based sector technology platform, In: Papermaking Science and Technol-

A Strategic Research Agenda for Inno- ogy, Gummerus Printing, Jyväskylä,
vation, Competitiveness and Quality of Finland, 1998.
Life. www.forestplatform.org, 2006. 8 Pinedo, M., Scheduling: Theory, Ap-
2 The European Tissue Paper Industry plications and Systems, Prentice-Hall,
Association, www.europeantissue.com, Englewood Cliffs, NJ, USA, 1995.
2006. 9 Ierapetritou, M. G., Floudas, C. A.,
3 Paulapuro, H., Papermaking Part 1, Effective continuous-time formu-
Stock Preparation and Wet End, Gum- lation for short-term scheduling: 2.
merus Printing, Jyväskylä, Finland, Continuous and semi-continuous
2006. processes, Ind. Eng. Chem. Res. 37
(1998), p. 4360.
4 Björk, K.-M., Carlsson, C., The effect
of flexible lead times on a paper pro- 10 Giannelos, N. F., Georgiadis, M. C.,
ducer, Int. J. Prod. Econ. 107 (2007), A novel event driven formulation for
pp. 139–150. short-term scheduling of multipur-
pose continuous processes, Ind. Eng.
5 European Commission, Reference doc-
Chem. Res. 41 (2002), p. 2431.
ument on best available techniques in the
pulp and paper industry, IPPC, 2001. 11 Castro, P. M., Barbosa-Póvoa, A. P.,
Matos, H. A., Novais, A. Q., Sim-
6 Lehtonen, J.-M., Holmström, J., Is
ple continuous-time formulation for
just-in-time applicable in paper indus-
short-term scheduling of batch and
try logistics? Supply Chain Manage.
continuous processes, Ind. Eng. Chem.
3(1) (1998), pp. 21–32.
Res. 43 (2004), p. 105.
7 Gullichsen, J., Paulapuro, H., Eco-
12 Kondili, E., Pantelides, C. C., Sar-
nomics of the Pulp and Paper Industry.
gent, R., A general algorithm for
268 8 Production and Inventory Planning for Stock Preparation in the Tissue Paper Industry

short-term scheduling of batch oper- review, Comput. Chem. Eng. 28 (2004),

ations: I. MILP formulation. Comput. p. 2109.
Chem. Eng. 17 (1993), p. 211. 22 Hendricks, K. B., Singhal, V. R.,
13 Shah, N., Pantelides, C. C., Sar- Stratman, J. K., The impact of en-
gent, R., Optimal periodic scheduling terprise systems on corporate per-
of multipurpose batch plants, Ann. formance: A study of ERP, SCM and
Oper. Res. 42 (1993), p. 193. CRM system implementations, J.
14 Pantelides, C. C., Unified Frame- Oper. Manag. 25 (2004), pp. 65–82.
works for the Optimal Process Plan- 23 Gupta, M., Kohli, A., Enterprise
ning and Scheduling, in: Proceedings resource planning systems and its
of the Second Conference on Foundations implications for operations function,
of Computer Aided Operations, Cache Technovation 22 (2004), pp. 687–696.
Publications, New York, 1994, p. 253.
24 Botta-Genoulaz, V., Millet, P.-A.,
15 Barbosa-Póvoa, A., Macchietto, S., Grabot, B., A survey on the recent
Detailed design of multipurpose batch research literature on ERP systems,
plants, Comput. Chem. Eng. 18 (1994), Comput. Ind. 56 (2004), pp. 510–522.
p. 1013.
25 Westerlund, J., Hästbacka, M., Fors-
16 Schilling, G., Pantelides, C. C., sell, S., Westerlund, T., A mixed-
Optimal periodic scheduling of mul- time MILP scheduling model, Ind.
tipurpose plants, Comput. Chem. Eng. Engng. Chem. Res. 46 (2007), pp. 2781–
23 (1994), p. 635. 2796.
17 Heo, S. K., Lee, K. H., Lee, H. K., Lee, 26 Göttsching, L., Pakarinen, H., Re-
I. B., Park, J. H., A new algorithm for cycled Fiber and De-inking, in: Paper-
cyclic scheduling and design of multi- making Science and Technology, Gum-
purpose batch plants, Ind. Eng. Chem. merus Printing, Jyväskylä, Finland,
Res. 42 (2003), p. 836. 2000.
18 Castro, P. M., Barbosa-Póvoa, A. P.,
27 Castro, P. M., Westerlund, J., Fors-
Novais, A. Q., Simultaneous design
sell, S., Optimal periodic scheduling
and scheduling of multipurpose
of a tissue paper mill with recycling
plants using resource task network
of by-products, Comput. Chem. Eng.
based continuous-time formulations,
(2006) (submitted).
Ind. Eng. Chem. Res. 44 (2005), p. 343.
28 Kortmann, J., Karlsson, P., Better
19 Pinto, T., Barbosa-Póvoa, A. P., No-
Delivery Accuracy and Capacity Uti-
vais, A. Q., Optimal design and retro-
lization using mySAP SCM: Common
fit of batch plants with a periodic
Planning at Corporate Level, SAP Info,
mode of operation, Comput. Chem.
November 28, 2005, www.sap.info.
Eng. 29 (2005), p. 1293.
29 Leiviskä, K., Process Control, Paper-
20 Méndez, C. A., Cerdá, J., Gross-
making Science and Technology, Gum-
mann, I. E., Harjunkoski, I., Fahl,
merus Printing, Jyväskylä, Finland,
M., State-of-the-art review of optimiza-
2005.
tion methods for short-term schedul-
ing of batch processes, Comput. Chem. 30 Kumar, S., Phrommathed, P., Im-
Eng. 30 (2006), pp. 913–946. proving a manufacturing process by
21 Floudas, C., Lin, X., Continuous-time mapping and simulation of critical
versus discrete-time approaches for operations, J. Manuf. Technol. Manage.
scheduling of chemical processes: A 17(1) (2006), pp. 104–132.
 269

9
Production Planning in Process Systems Engineering
Charles Sung, Christos T. Maravelias

The goal of this chapter is to provide a review of the methods that have been pro-
posed for the solution of production planning problems in the chemical industry.
We begin by presenting the classical formulations used in operations research and
the main software tools currently used in practice. Next, we discuss the limitations
of current methods to address production planning problems in the chemical in-
dustry. We then present the modeling approaches and optimization strategies for
the integration of production planning with scheduling. We conclude with a brief
discussion of specific applications and the integration of production planning with
other models.

9.1
Introduction

The management of the supply chain of a chemical enterprise requires answers

to multiple questions at different levels and time scales: from procurement of
raw materials to sales management and from long-term (5–10 years) to short-
term (1–4 weeks) decisions. A supply chain (SC) is a network of linked organi-
zations that work together to provide the end-consumer with products and ser-
vices [1–3]. A company-centric or intraorganizational SC places the organization
in the center, connected upstream to first-tier suppliers and downstream to first-
tier customers [4]. A SC is often represented with a network, where the different
SC entities/nodes are categorized by their function: procurement (suppliers), pro-
duction (manufacturers), distribution (warehouses, distribution centers), or sales
(retailers).
Planning problems can also be categorized as long-term strategic planning,
mid-term tactical planning, or short-term operational planning problems. These
decisions are made hierarchically with different information, frequency, and re-
strictions [5]. Strategic planning determines the structure of the SC network sub-
ject to suppliers and manufacturing facilities being ultimately connected to end-
consumers; it determines the location of production sites, warehouses, the capacity

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
270 9 Production Planning in Process Systems Engineering

Procurement Production Distribution Sales

Long-Term Supply Network Design

Mid-Term Master Planning

Demand
Planning
Production Distribution
Materials Planning Planning
Short-Term Requirement
Planning Transport Demand
Scheduling
Planning Fulfillment

Fig. 9.1 A matrix used by supply chain management software systems (from [6]).

of these facilities, and the transportation means [1]. Tactical planning coordinates
flows between sites assigning production targets and determines inventory levels
at the nodes of the SC, while operational planning is run within each site to deter-
mine the detailed assignment and sequencing of tasks in equipment units.
Clearly, planning at higher SC levels affects decisions made at lower levels (e.g.,
facility location affects future capacity expansions). In addition, business functions
are interconnected (e.g., procurement decisions affect production), thus, leading
to a need for horizontal as well as vertical integration of planning methods. The
structure of the different planning tasks based on the business function and time
scale is shown in the supply chain planning matrix in Fig. 9.1.
Here, we are specifically interested in production planning, a subset of opera-
tional planning that is studied in operations research (OR). Production planning
is solved to optimize trade-offs between economic objectives (e.g., minimization
of inventory cost) and customer satisfaction [7]. It is solved to determine produc-
tion and inventory targets over a medium time horizon (e.g., of several months).
As seen in Fig. 9.1, production planning is traditionally separated from scheduling
and demand planning [8]. In the process industry, however, this separation can be
adverse because scheduling decisions have a major impact on what production tar-
gets are attainable [9, 10]. To address this limitation, researchers in process systems
engineering (PSE) have developed methods to more accurately solve the production
planning problem using some form of integration with scheduling. This has led to
a number of papers that now comprise a significant body of literature. It is the
objective of this paper to provide a review of these methods.
This paper is structured as follows. In Section 9.2, we introduce the produc-
tion planning problem, briefly discuss the main OR-based methods to address it,
review the main software tools currently used in practice and present the main
challenges in production planning in the chemical industry. In Section 9.3, we re-
view the different modeling approaches for the integration of production planning
 9.2 Background 271

with scheduling and we discuss the main optimization strategies exploited by re-
searchers to solve integrated models effectively. Integration of production planning
with other models and specific applications are discussed in Section 9.4.

9.2
Background

9.2.1
Problem Statement

The objective in production planning (PP) is to fulfill customer demand at mini-

mum total (i.e., production + inventory) cost. Formally, we are given:
1. A planning horizon divided into a set T of NT time periods
(i.e., time buckets).
2. A set I of NI products (items) with customer demand for
each item due at the end of each time bucket.
3. Inventory and production costs.
4. Resource capacities.

Optimization decisions include:

1. Production level (amount) Xi,t of item i ∈ I in period t ∈ T.
2. Inventory level Si,t of item i at the end of period t.

The goal usually is the optimization of an economic objective.

9.2.2
Classical Mathematical Programming Formulations

The simplest formulation for the production planning problem (PPP) is the single-
item, uncapacitated lot-sizing problem with no setup time or cost [11]:


min (pt Xt + ht St ), (9.1)
t
St−1 + Xt = St + dt ∀t, (9.2)
St , Xt
0 ∀t, (9.3)
where dt is the demand for the single item at the end of period t. In Eq. (9.1), the
objective function is the minimization of production and holding costs. Equation
(9.2) expresses the conservation of flow. A network representation of this standard
PPP is shown in Fig. 9.2, where the inventory Si,t−1 from the previous time period
and the production level Xi,t in the current time period (incoming flows) are equal
to customer demand di,t and the inventory Si,t in the current time period (outgoing
flows).
Partition of the time horizon into time periods is a decision to be made by the
practitioner [12] and may be influenced by when raw materials become available
272 9 Production Planning in Process Systems Engineering

t=1 t=2 t=3 t=4

P r odu ction P la n ning

Production Xit

Initial Inventory Inventory Si,t-1 Sit

Demand dit

Fig. 9.2 Flow conservation in the network representation of the PPP.

and when demand is due. For example, nonuniform time periods may be chosen
such that later time periods are more aggregated (i.e., have longer length).
The single-item, uncapacitated lot-sizing problem in Eqs. (9.1)–(9.3) can be used
in material requirements planning, but cannot be readily used for production plan-
ning because it does not account for multiple products competing for shared lim-
ited resources. A more general formulation, given by Eqs. (9.4)–(9.8), allows for
multiple items (single-level or multilevel), capacity constraints (which make the
model “capacitated”), backlog, and setup time and cost. Binary variable Yi,t is equal
to one if item i is produced during period t and the backlog level of item i at the
end of period t is denoted by Bi,t .


min = (fi Yi,t + pi Xi,t + hi Si,t + ui Bi,t ), (9.4)
i,t

Si,t−1 + Xi,t + Bi,t = Si,t + di,t + Bi,t−1 ∀i, t, (9.5)

αi Xi,t + βi Yi,t  Ct ∀t, (9.6)
i i

Xi,t  Ct Yi,t ∀i, t, (9.7)

Bi,t , Si,t , Xi,t
0 Yi,t ∈ {0, 1} ∀i, t. (9.8)
The objective in Eq. (9.4), includes setup, production, inventory, and backlog costs,
while the flow conservation constraint in Eq. (9.5) accounts for backlog. Equation
(9.6) is a capacity constraint, where Ct can be viewed for example as the available
time in period t, αi is the reverse of production rate for item i and βi is the setup
time for item i. If multiple resources j ∈ J need to be modeled, Eq. (9.6) is replaced
by Eq. (9.6’),

j
j j
αi Xi,t + βi Yi,t  Ct ∀j, t, (9.6’)
i i
j j j
where is the capacity of resource j , and αi and βi are the inverse of production
Ct
rate and setup time, respectively for item i in resource j . Finally, item i can be
produced only if there is a setup. Note that setups may incur a fixed cost (penalty
in objective function), consume time (penalty in capacity constraint), or both [13].
 9.2 Background 273

The above multiitem capacitated lot-sizing formulation can be extended to ac-

count for multiple levels (echelons) of items using the bill of materials (BOM), an
acyclic graph that records for every product i what products i  require i, in what
ratio ri,i  and with what lead-time γi,i  . In this case, the flow balance constraint in
Eq. (9.5) must be replaced by Eq. (9.5’):


Si,t−1 + Xi,t = Si,t + di,t − ri,i  Xi  ,t+γi,i  ∀i, t, (9.5’)
i  ∈BOMi,i 

where production of A in time period t + 1 depletes the end of period t inventory

of product B.
Lot-sizing models can also be modified to account for lost sales instead of back-
logged demand, minimum production levels, safety stocks, discrete lot-sizes, ca-
pacity expansion, overtime, etc. An excellent classification of lot-sizing models is
presented in Pochet and Wolsey [7]. It is important to note that since most of these
mixed-integer programming (MIP) formulations are hard to solve, they are decom-
posed into simpler single-item subproblems and solved using existing algorithms
and reformulation results [7].

9.2.3
Software Systems

Decision support tools (i.e., software) can be divided into transactional IT systems
and analytical IT systems. Transactional IT systems such as enterprise resource
planning (ERP) can maintain and react to data. ERP systems integrate a range
of databases such as financial, controlling, manufacturing, and human resources.
Analytical IT systems such as material requirements planning (MRP) and ad-
vanced production and scheduling (APS) can plan ahead using data from ERP
systems [14].
An MRP system is shown in Fig. 9.3. First, aggregate planning (or master plan-
ning) sets capacity utilization and aggregate inventory levels to meet forecast de-
mand over a medium-term horizon [16–18]. Next, master production scheduling
(MPS) computes production and inventory for end-products. MPS is based on a
multi-item, single-level capacitated lot-sizing problem. Multiple levels (echelons)
of single-level capacitated lot-sizing problems may be combined to form the mater-
ial requirements planning (MRP-I or MPS/MRP) model. Different levels are linked
by the BOM, and precedence is enforced using lead-time. As seen in Fig. 9.1, de-
mand planning is separate from both master planning and production planning,
so MRP is sometimes solved in two phases: the push phase, based on demand
forecasts, plans production from raw materials to specified levels of intermediate
products. The pull phase, based on confirmed orders and due dates, schedules pro-
duction from intermediates products to end-products. These two phases are similar
to Make-to-Stock Planning and Final Assembly Scheduling, respectively [7].
Beyond production planning, APS systems have been introduced to integrate the
entire supply chain. APS systems follow the general structure shown in Fig. 9.1,
viewing individual modules as linked but separate [1, 18]. The optimization meth-
274 9 Production Planning in Process Systems Engineering

Aggregate Planning Forecasting

Master Production
BOM Scheduling
Inventory
Product
Status
Structure Material Requirements
Planning

Scheduling Vendor Control

Fig. 9.3 An MRP-II system [7].

ods used by APS systems may be rigorous or nonrigorous and are not always de-
scribed in detail by their vendor.

9.2.4
Challenges in Production Planning in the Chemical Industry

Despite the multiple extensions of lot-sizing formulations, the methods described

above are limited in addressing production planning problems in the chemical
industry due to the existence of large sequence-dependent changeover times and
several subtle features of chemical process networks [19, 20].
To accurately determine the reduction in equipment capacity (i.e., loss of produc-
tion) due to sequence dependent changeovers, sequencing should be taken into ac-
count. Lot-sizing models, however, do not account for the sequencing of items; at
best, they subtract a fixed setup time if a product is produced within a time period
[see Eqs. (9.6) and (9.6’)]. Furthermore, setups are often modeled as beginning and
ending in the same time period, but in the presence of large changeover times this
restriction leads to suboptimal solutions. To address this limitation, setup cross-
over should be allowed. In addition, most lot-sizing formulations assume that setup
times are shorter than the planning period. This, however, is not necessarily true in
sectors such as pharmaceuticals where changeover times can be as long as several
weeks. The formulations proposed by Drexl, Suerie, and Sung and Maravelias [21–
23] address some of these limitations, but are computationally expensive.
Another subtle feature of process networks is coproduction, which occurs in
multi-stage continuous facilities where a unit is connected to two or more down-
stream units or in processes with significant amounts of byproducts. Coproduction
cannot be modeled via Eq. (9.6) and requires the introduction of detailed process
models. Furthermore, storage capacity constraints can affect the overall production
 9.3 Integration of Production Planning and Scheduling 275

rate. Finally, the concept of lead-time is too coarse to accurately enforce precedence
among items within time periods.

9.3
Integration of Production Planning and Scheduling

To effectively solve PPPs involving complex chemical process networks, re-

searchers have proposed methods that incorporate scheduling information. This
is achieved via the integration of the production planning formulation with a form
of a scheduling subproblem. Integration allows the higher level production plan-
ning model to be aware of the lower level scheduling constraints, and thus obtain
appropriate production targets [9, 10, 24]. The proposed methods can be classified
based on (a) the models employed to account for scheduling constraints, and (b)
the corresponding solution methods.
Before we proceed with their classification, it is important to highlight that so-
lution methods for integrated production planning formulations depend on the
modeling approach. For example, if the PPP is solved using a detailed scheduling
model, the goal of the solution method is the computational enhancement. On the
other hand, if the PPP is solved using an approximate scheduling submodel, then
the goal of the solution method is the development of an effective iterative scheme
that allows us to correct infeasible production targets. Thus, the classifications in
Section 9.3 and Section 9.3.1 should be considered simultaneously.

9.3.1
Scheduling Subproblems

Scheduling constraints can be incorporated using the following three general

classes of models.

9.3.1.1 Integration with Detailed Scheduling Formulations

The advantage of this class of methods is that production targets are exact because
they satisfy a detailed schedule, so no corrective reoptimization is necessary. The
disadvantage is that these approaches tend to be computationally expensive, and
thus difficult to use for the solution of practical problems. For the solution of larger
instances, specialized solution methods that exploit the structure of the integrated
formulation have been proposed. If solved to optimality, these methods provide
production targets that neither under-utilize capacity (suboptimality) nor require
backtracking to correct infeasibilities. A thorough review of scheduling formula-
tions can be found in Méndez et al. [25].

9.3.1.2 Integration with Approximate Scheduling Models

In an effort to develop methods that provide near feasible production targets while
being computationally effective, several researchers have proposed the integration
with approximate scheduling models. These models are either relaxations of large-
276 9 Production Planning in Process Systems Engineering

scale MIP scheduling formulations or aggregate scheduling formulations. Relaxed

problems can be obtained by considering only a subset of scheduling decisions and
dropping the irrelevant constraints, e.g., account for assignment but not sequenc-
ing decisions [26]. Aggregated formulations can be obtained by lumping together
decisions for multiple small-bucket scheduling periods [27]. If the predicted pro-
duction targets are infeasible, then the integrated model can be solved iteratively.
The challenge in this case, is the communication of infeasibility information back
to the PPP.

9.3.1.3 Integration with Surrogate Models Obtained Offline

The main idea behind these methods is the offline development of an accurate
but simple surrogate model that enables us to easily determine what production
targets are feasible for a given facility. Note that although a detailed formulation
can be used for the development of the surrogate model, the surrogate model itself
does not have to have the structure or the variables of this formulation. The ap-
proximations provided by these surrogate models can be of high quality for certain
process networks [20, 28].

9.3.2
Solution Methods

In general, integrated models can be solved holistically (“full-space”), hierarchi-

cally, or iteratively. For hierarchical methods, the higher level model is solved using
an a priori approximation of the lower level model. Then the lower level model is
solved to guarantee feasibility. For iterative methods, the higher level contains an
approximation of the lower level that is updated each time the lower level is solved.
Solution methods can be broadly divided into the three classes shown in Fig. 9.4.

Higher Level Higher Level Higher Level

PP Model PP Model PP Model

Detailed Approximate Approximate

Lower Level Scheduling Scheduling
Scheduling Model Model
Model

Detailed Detailed
Scheduling Scheduling
Model Model

a) Full-space method b) Hierarchical method c) Iterative method

Fig. 9.4 Three classes of methods for solving integrated models.

 9.3 Integration of Production Planning and Scheduling 277

9.3.2.1 Full-Space Methods

The first class of methods solves the integrated full-space model using rigorous
methods (e.g., mathematical programming) [29, 30]. Although one may be able
to yield the overall optimal solution for small instances, this method cannot be
readily applied to real world applications because rigorous methods scale poorly
with model size (e.g., MILP branch-and-bound). In addition, the higher level PP
model involves many periods and resources and the detailed lower level model
is defined for each period and unit, thereby resulting in large integrated models.
The full-space model may also be solved using nonrigorous methods such as ge-
netic algorithms or simulated annealing [31]. Solutions found by these methods
are not provably optimal, but are reportedly good nevertheless. Lot-sizing heuris-
tics are reviewed by Hariga and Alyan [32] and compared by Heemsbergen and
Malstrom [33]. Genetic algorithms in production and operations management are
reviewed by Chaudhry and Luo [34] and used for integrated production planning
and scheduling by Yan and Zhang [35], Berning et al. [36], and Romero et al. [37].
Research efforts for this class of solution methods focuses on the development
of smaller models (e.g., continuous-versus discrete-time and slot-based versus
sequence-based models), tighter formulations, and generation of problem-specific
cuts [38].

9.3.2.2 Hierarchical Methods

The second class of methods solves the higher level model in the presence of a
scheduling approximation. Approximations may be created: (a) by dropping or ag-
gregating constraints, (b) from historical precedent or policy, (c) using partial enu-
meration, or (d) from optimization. First, Wilkinson et al. [27] present an approach
where variables and equations are aggregated by time. Vancza et al. [39] discuss
an aggregation of standard data by time, capacity, and operations. Zhu and Ma-
jozi [40] present a planning model linked by production targets to state sequence
network scheduling models and comment on the block angular structure. A hierar-
chal decomposition is proposed in which the planning problem is solved to assign
production targets, subject to aggregate capacity constraints. Approximate subprob-
lems based on aggregation are model-specific but require no offline computation.
Second, surrogate constraints may be suggested by historical precedent [7] or pol-
icy rules [41]. Third, computation can be used to identify favorable lower level so-
lutions. Birewar and Grossmann [42] recommend finding cycle and slack time for
enumerated single- and multi-product campaigns. Sukoyo et al. [43] do the same
for product mixes. Henriques [44] constrains processing time by limiting produc-
tion rate between rate of minimum cost and rate of maximum production. Fourth,
computation may be used to develop a holistic approximation of the scheduling
model. Wan et al. [28] uses simulation to create a support vector machine. Mul-
tiobjective and parametric methods have also been proposed [45]. Wellons and
Reklaitis [46] identify cyclic campaigns with dominant production rates. Sung and
Maravelias [20] identify the attainable region of production. The disadvantage of
hierarchical methods is the lack of a recourse mechanism that reacts to suboptimal
directives from the PP model. Although rolling horizon algorithms are often used
278 9 Production Planning in Process Systems Engineering

to address this limitation [47], the development of compact yet accurate approxima-
tions for hierarchical approaches remains a major challenge.

9.3.2.3 Iterative Methods

The third class of methods rigorously improves the quality of the approximate
scheduling model through iterative solution of a high-level planning (master) and
a lower level detailed scheduling (slave) model, as seen in Fig. 9.4. These methods
are usually problem-specific because they exploit the structure of the lower level
scheduling model. Generally, the original approximate model is a relaxation of the
detailed lower level model, so the solution of the high-level model provides an op-
timistic solution relative to the integrated model (lower bound if minimizing). By
comparing this solution to the detailed lower level model, a new solution can be
found that is feasible for the integrated model (upper bound if minimizing). The
approximate model can be updated via generation of new columns or addition of
new linear constraints. Most methods in this class employ Lagrangean decompo-
sition and column generation [48] (Kondili et al. (2003) [74]). Chu and Xia [72] pro-
pose a Benders-type decomposition in which a specially formulated subproblem is
solved to determine the constraint to be added. For mixed-integer approximations,
integer cuts are used to exclude previously found solutions. Dogan-Erdirik and
Grossmann [18] propose a decomposition in which the higher level model assigns
products to time periods and the lower level returns integer cuts indicating the to-
tal transition time for that assignment. Maravelias and Grossmann [73], Hooker
and Ottoson [49], and Timpe [50] describe frameworks in which the lower level
model is solved using constraint programming to provide a new cut to the higher
level model. Maravelias [51] presents a hybrid iterative scheme that integrates MIP
methods with problem-specific heuristics. Papageorgiou and Pantelides [52] pro-
pose an iterative decomposition approach for optimal campaign planning. Finally,
Stefansson et al. [53] present a three-level iterative scheme for production plan-
ning in the pharmaceutical industry. Although iterative methods can in principle
yield the optimal solution in a finite number of iterations, their performance de-
pends strongly on the number of iterations needed to find the optimal solution and
prove optimality. The number of iterations can be reduced by (a) tightening the
higher level problem, (b) excluding nonpromising solutions using preprocessing
algorithms, and (c) developing integer cuts during the search that cut-off many in-
feasible or suboptimal solutions [51]. Although significant progress has been made
in this area, the development of strong integer cuts and powerful preprocessing
algorithms remains an open research question for most process networks.

9.4
Applications and Extensions

Specific applications of integrated production planning and scheduling have been

presented for the refining [54], pharmaceutical [53], sheet-metal [55], and fruit pack-
aging [56] industries. Production planning has also been integrated with distribu-
 9.5 Conclusions 279

tion [57–61], and maintenance policy [62, 63] models. Kondili et al. [64] discuss
the integration of production planning with energy. Grau et al. [65] consider si-
multaneously planning and energy use, while Romero et al. [37] consider the inte-
gration of production planning and scheduling with budgeting. Methods for cam-
paign scheduling and planning are developed by Oh [66] and Papageorgiou and
Pantelides [52]. Finally, van den Heever and Grossmann [67] consider production
planning and rescheduling.
In addition, the area of production planning under uncertainty has received con-
siderable attention for the last 10 years (see [68], and [69] for reviews). Specifically,
Neiro and Pinto [70] present a mathematical programming model for production
planning in petroleum refineries under price and demand uncertainty. Ierapetri-
tou et al. [71] discuss operational planning under uncertainty, while Gupta and
Maranas [48] consider demand uncertainty in supply chain planning. Pratikakis
et al. [75] present a dynamic programming method for production planning and
scheduling under uncertainty.

9.5
Conclusions

Production planning has been a very active research area in process systems engi-
neering for the last 18 years. Despite the significant progress that has been made
both in terms of modeling approaches and solution algorithms, the solution of
large practical production planning problems in the chemical industry remains
elusive. Research challenges in this area are very broad. They include (a) the devel-
opment of tighter scheduling formulations and the generation of problem-specific
cutting planes for the solution of large-scale integrated planning-scheduling for-
mulations, (b) the development of accurate surrogate models for hierarchical pro-
duction planning and scheduling, and (c) the enhancement of iterative schemes
through the development of effective preprocessing algorithms and strong integer
cuts. Another promising area, that is currently virtually unexplored, is the develop-
ment of decomposition schemes that exploit known algorithms and reformulation
results for the solution of different variants of the lot-sizing problem.

Acknowledgments

The authors gratefully acknowledge financial support from the Department of

Chemical and Biological Engineering at the University of Wisconsin – Madison
and the National Science Foundation, under Grant no. CTS-0547443.
280 9 Production Planning in Process Systems Engineering

References

1 Stadtler, H., Supply chain manage- 12 Liu, M. L., Sahinidis, N. V., Com-
ment and advanced planning – basics, putational trends and effects of ap-
overview and challenges. Eur. J. Oper. proximations in an MILP model for
Res. 163 (2005), pp. 575–588. process planning. Ind. Eng. Chem. Res.
2 Christopher, M., Logistics and sup- 34 (1995), pp. 1662–1673.
ply chain management, Strategies for 13 Zhu, X., Wilhelm, W. E., Schedul-
Reducing Cost and Improving Service. ing and lot sizing with sequence-
2nd ed., Financial Times, Pitman, dependent setup: A literature review.
London, 1998. IIE Trans. 38 (2006), pp. 987–1007.
3 Beamon, B. M., Supply chain design 14 Knolmayer, G., Mertens, P., Zeier,
and analysis: Models and methods. A., Supply Chain Management Based
Int. J. Prod. Econ. 55 (1998), pp. 281– on SAP Systems, Springer, Berlin, Ger-
294. many, 2002.
4 Shah, N., Process industry supply 15 Buxey, G., Aggregate planning for sea-
chains: advances and challenges. sonal demand reconciling theory with
Comput. Chem. Eng. 29 (2005), pp. practice. Int. J. Oper. Plan. Manu. 25
1225–1235. (2005), pp. 1083–1100.
5 Lasschuit, W., Thijssen, N., Sup- 16 Behnezhad, A. R., Khoshnevis, B.,
porting supply chain and planning Integration of machine requirements
decisions in the oil and chemical in- planning and aggregate production
dustry. Comput. Chem. Eng. 28 (2004), planning. Prod. Plan. Control. 7 (1996),
pp. 863–870. pp. 292–298.
6 Meyr, H., Wagner, M., Rohde, J., 17 Nam, S. J., Logendran, R., Aggregate
Structure of advanced planning sys- production planning – A survey of
tems. in: Supply Chain Management models and methodologies. Euro. J.
and Advanced Planning – Concepts, Oper. Res. 61 (1992), pp. 255–272.
Models, Software, and Case Stud-
ies, Stadtler, H., Kilger, C. (eds.), 18 Neumann, K., Schwindt, C., Traut-
Springer, Berlin, 2002, pp. 99–104. mann, N., Advanced production
scheduling for batch plants in process
7 Pochet, Y., Wolsey, L. A., Production industries. OR Spectr. 24 (2004), pp.
Planning by Mixed Integer Program- 251–279.
ming, Springer, New York, 2006.
19 Dogan-Erdirik, M., Grossmann,
8 Saharidis, G. K., Dallery, Y., Karaes-
I. E., A decomposition method for the
men, F., Centralized versus decentral-
simultaneous planning and schedul-
ized production planning. Rairo-Oper.
ing of single-stage continuous multi-
Res. 40 (2006), pp. 113–128.
product plants. Ind. Eng. Chem. Res.
9 Grossmann, I. E., Enterprise-wide op- 45 (2006), pp. 299–315.
timization: a new frontier in process
20 Sung, C., Maravelias, C. T., An At-
systems engineering. AIChE J. 51
tainable region approach for pro-
(2005), pp. 1846–1857.
duction planning of multi-product
10 Eggersmann, M., von Wedel, L., processes. AIChE J. 53 (2007a), pp.
Marquardt, W., Management and 1298–1315.
reuse of mathematical models in the
21 Drexl, A., Haase, K., Proportional
industrial design process. Chem. Eng.
lotsizing and scheduling. Int. J. Prod.
Technol. 27 (2004), pp. 13–22.
Econ. 40 (1995), pp. 73–87.
11 Brahimi, N., Dauzere-Peres, S., Na-
jid, N. M., et al., Single item lot sizing 22 Suerie, C., Modeling of period over-
problems. Eur. J. Oper. Res. 168 (2006), lapping setup times. Eur. J. Oper. Res.
pp. 1–16. 174 (2006), pp. 874–886.
 References 281

23 Sung, C., Maravelias, C. T., A mixed- MRP lot-sizing heuristics – an analy-

integer programming formulation sis of performance by rule. Prod. Plan.
for the general capacitated lot-sizing Control. 5 (1994), pp. 381–391.
problem with setup cross over. To ap- 34 Chaudhry, S. S., Luo, W., Application
pear in Comput. Chem. Eng. (2007b). of genetic algorithms in production
24 Subrahmanyam, S., Pekny, J. F., and operations management: A re-
Reklaitis, G. V., Decomposition ap- view. Int. J. Prod. Res. 43 (2005), pp.
proaches to batch plant design and 4083–4101.
planning. Ind. Eng. Chem. Res. 35 35 Yan, H. S., Zhang, X. D., A case
(1996), pp. 1866–1876. study on integrated production plan-
25 Méndez, C. A., Cerda, J., Gross- ning and scheduling in a three-stage
mann, I. E., Harjunkoski, I., Fahl, manufacturing system. IEEE Trans.
M., State-of-the-art review of optimiza- Automat. Sci. Eng. 4 (2007), pp. 86–92.
tion methods for short-term schedul- 36 Berning, G., Brandenburg, M.,
ing of batch processes. Comput. Chem. Gursov, K., Kussi, J. S., Mehta, V.,
Eng. 30 (2006), pp. 913–946. Tolle, F. J., Integrating collaborative
26 Jain, V., Grossmann, I. E., Algo- planning and supply chain optimiza-
rithms for hybrid MILP/CP models tion for the chemical process industry
for a class of optimization problems. (I) – methodology. Comput. Chem.
INFORMS J. Comput. 13 (2001), pp. Eng. 28 (2004), pp. 913–927.
258–276. 37 Romero, J., Badell, M., Bagajewicz,
M., Puigjaner, L., Integrating bud-
27 Wilkinson, S. J., Shah, N., Pan-
geting models into scheduling and
telides, C. C., Aggregate modeling
planning models for the chemical
of multipurpose plant operation. Com-
batch industry. Ind. Eng. Chem. Res. 42
put. Chem. Eng. 19 (1995), pp. S583–
(2003), pp. 6125–6134.
S588.
38 Miller, A. J., Nemhauser, G. L.,
28 Wan, X., Pekny, J. F., Reklaitis, G. V., Savelsbergh, M. W. P., A multi-
Simulation-based optimization with item production planning model
surrogate models – application to with setup times: algorithms, refor-
supply chain management. Comput. mulations, and polyhedral charac-
Chem. Eng. (2006), pp. 1317–1328. terizations for a special case. Math.
29 Kallrath, J., Planning and schedul- Program. 95 (2003), pp. 71–90.
ing in the process industry. OR Spectr. 39 Vancza, J., Kis, T., Kovacs, A., Aggre-
24 (2002), pp. 219–250. gation – the key to integrating produc-
30 Bassett, M. H., Pekny, J. F., tion planning and scheduling. CIRP
Reklaitis, G. V., Decomposition Ann.-Manuf. Technol. 53 (2004), pp.
techniques for the solution of large- 377–380.
scale scheduling problems. AIChE. 42 40 Zhu, X. X., Majozi, T., Novel contin-
(1996), pp. 3373–3384. uous time MILP formulation for mul-
31 Reklaitis, G. V., Overview of plan- tipurpose batch plants. 2. Integrated
ning and scheduling technologies. planning and scheduling. Ind. Chem.
Latin Am. Appl. Res. 30 (2000), pp. Eng. Res. 40 (2001), pp. 5621–5634.
285–293. 41 Schmidt, C., Performance guarantee
32 Hariga, M. A., Alyan, A. A., The dy- of two simple priority rules for pro-
namic lot-sizing problem (DLSP) – duction scheduling . Int. J. Prod. Econ.
Review, extension, and comparison. 68 (2000), pp. 151–159.
Arabian J. Sci. Eng. 17 (1992), pp. 387– 42 Birewar, D. B., Grossmann, I. E.,
402. Simultaneous production planning
33 Heemsbergen, B. L., Malstrom, and scheduling in multiproduct batch
E. M., A simulation of single-level plants . Ind. Eng. Chem. Res. 29 (1990),
pp. 570–580.
282 9 Production Planning in Process Systems Engineering

43 Sukoyo, Matsuoka, S., Muraki, M., 53 Stefansson, H., Shah, N., Jensson,
Production planning for multi-site P., Multiscale planning and schedul-
batch plants with the MILP method. J. ing in the secondary pharmaceuti-
Japan Petrol. Inst. 47 (2004), pp. 318– cal industry. AIChE J. 52 (2006), pp.
325. 4133–4149.
44 Henriques, E., Towards the inte- 54 Joly, M., Moro, L. F. L., Pinto, J. M.,
gration of process and production Planning and scheduling for petro-
planning: an optimisation model for leum refineries using mathematical
cutting parameters. Int. J. Adv. Manuf. programming. Brazilian J. Chem. Eng.
Technol. 28 (2006), pp. 117–128. 19 (2002), pp. 207–228.
45 Chen, Y. K., Liao, H. C., An inves- 55 Cattrysse, D., Beullens, P., Collin,
tigation on selection of simplified P., Duflou, J., Van Oudheusden,
aggregate production planning strate- D., Automatic production planning
gies using MADM approaches. Int. J. of press brakes for sheet metal bend-
Prod. Res. 41 (2003), pp. 3359–3374. ing. Int. J. Prod. Res. 44 (2006), pp.
4311–4327.
46 Wellons, M. C., Reklaitis, G. V.,
Scheduling of multipurpose batch 56 Blanco, A. M., Masini, G., Petracci,
chemical plants. 2. Multiple-product N., Bandoni, J. A., Operations man-
campaign formation and production agement of a packaging plant in the
planning. Ind. Chem. Eng. Res. 30 fruit industry. J. Food Eng. 70 (2005),
(1991), pp. 688–705. pp. 299–307.
57 Oh, H. C., Karimi, I. A., Global mul-
47 Dimitriadis, A. D., Shah, N., Pan-
tiproduct production-distribution
telides, C. C., RTN-based rolling
planning with duty drawbacks.
horizon algorithms for medium term
AIChE. 52 (2006), pp. 595–610.
scheduling of multipurpose plants.
Comput. Chem. Eng. 21 (1997), pp. 58 Ryu, J. H., Dua, V., Pistikopoulos,
S1061–S1066. E. N., A bilevel programming frame-
work for enterprise-wide optimization
48 Gupta, A., Maranas, C. D., Manag- networks under uncertainty. Comput.
ing demand uncertainty in supply Chem. Eng. 28 (2004), pp. 1121–1129.
chain planning. Comput. Chem. Eng.
59 Jackson, J. R., Grossmann, I. E.,
27 (2003), pp. 1219–1227.
Temporal decomposition scheme
49 Hooker, J. N., Ottosson, G., Logic- for nonlinear multisite production
based Bender’s decomposition. Math. planning and distribution models.
Prog. 96 (2003), pp. 33–60. Ind. Eng. Chem. Res. 42 (2003), pp.
50 Timpe, C., Solving planning and 3045–3055.
scheduling problems with combined 60 Lee, H. K., Lee, I. B., Reklaitis, G. V.,
integer and constraint programming. Capacity expansion problem of multi-
OR Spectr. 24 (2002), pp. 431–448. site batch plants with production and
51 Maravelias, C. T., A decomposi- distribution. Comput. Chem. Eng. 24
tion framework for the scheduling (2000), pp. 1597–1602.
of single- and multi-stage processes. 61 Timpe, C. H., Kallrath, J., Optimal
Comput. Chem. Eng. 30 (2006), pp. planning in large multi-site produc-
407–420. tion networks. Eur. J. Oper. Res. 126
52 Papageorgiou, L. G., Pantelides, (2000), pp. 422–435.
C. C., Optimal campaign plan- 62 Pistikopoulos, E. N., Vassiliadis,
ning/scheduling of multipurpose G. C., Arvela, J., Papageorgiou, L. G.,
batch/semicontinuous plants. 1. Interactions of maintenance and pro-
Mathematical decomposition. Ind. duction planning for multipurpose
Chem. Eng. Res. 35 (1996), pp. 510– process plants – A system effective-
529. ness approach. Ind. Eng. Chem. Res.
40 (2001), pp. 3195–3207.
 Appendix 283

63 Dedopoulos, I. T., Shah, N., Long- opportunities. Comput. Chem. Eng. 28

term maintenance policy optimization (2004), pp. 971–983.
in multipurpose process plants. Chem. 70 Neiro, S. M. S., Pinto, J. M., Multi-
Eng. Res. Des. 74 (1996), pp. 307–320. period optimization for production
64 Kondili, E., Shah, N., Pantelides, planning of petroleum refineries.
C. C., Production planning for the ra- Chem. Eng. Commun. 192 (2005), pp.
tional use of energy in multiproduct 62–88.
continuous plants. Comput. Chem. 71 Ierapetritou, M. G., Pistikopoulos,
Eng. 17 (1993), pp. S123–S128. E. N., Floudas, C. A., Operational
65 Grau, R., Graells, M., Corominas, planning under uncertainty. Comput.
J., Espuna, A., Puigjaner, L., Global Chem. Eng. 20 (1996), pp. 1499–1516.
strategy for energy and waste analysis 72 Chu, Y. Y., Xia, Q., A hybrid algo-
in scheduling and planning of mul- rithm for a class of resource con-
tiproduct batch chemical processes. strained scheduling problems. Lec-
Comput. Chem. Eng. 20 (1996), pp. ture Notes Comp. Sci. 3524 (2005), pp.
853–868. 110–124.
66 Oh, H. C., Karimi, I. A., Planning 73 Maravelias, C. T., Grossmann, I. E.,
production on a single processor with A hybrid MILP/CP decomposition ap-
sequence-dependent setups part 1: proach for the short-term scheduling
determination of campaigns. Comput. of multipurpose batch plants. Comput.
Chem. Eng. 25 (2001), pp. 1021–1030. Chem. Eng. 28 (2004), pp. 1921–1949.
67 Van den Heever, S. A., Grossmann,
74 Kondili, B., Fatemi Ghomi, S. M. T.,
I. E., A strategy for the integration
Wison, J. B., The incapacitated lot
of production planning and reactive
sizing problem: A review of models
scheduling in the optimization of a
and algorithms. Omega 31 (2003), pp.
hydrogen supply network. Comput.
365–378.
Chem. Eng. 27 (2003), pp. 1813–1839.
75 Pratikakis, N. E., Lee, J. H., Realff,
68 Mula, J., Poler, R., Garcia-Sabater,
M. T., A real time adaptive dynamic
J. P., Lario, F. C., Model for produc-
programming approach for planning
tion planning under uncertainty: A
and scheduling. Proceedings 16th Eu-
review. Int. J. Prod. Econ. 103 (2006),
ropean Symposium on Computer Aided
pp. 271–285.
Process Engineering, Marquardt, W.,
69 Sahinidis, N. V., Optimization un- Pantelides, C. C. (eds.), Elsevier, Am-
der uncertainty: state-of-the-art and sterdam.

Appendix

Notation

Indices/Sets
i item (product)
t time period
j resource

Parameters
bi,t backlog cost for one unit of i from end of t – 1 to end of t
Ct capacity in t
284 9 Production Planning in Process Systems Engineering

j
Ct capacity of unit j during period t
di,t demand for i due at end of t
hi,t holding cost for one unit of i from end of t – 1 to end of t
fi,t setup cost for i in t
ct cost of overtime for
ri,i  amount of i needed for one unit of i 
j
αi variable capacity requirements of item i in unit j
j
βi fixed capacity consumed by setup for item i in unit j
γi,t lead-time that i needs before i 

Planning Variables
Bi,t backlog, demand for i still unfulfilled at end of t – 1
Xi,t production, production amount of i during t
Si,t inventory, inventory of i at end of t
Yi,t setup = 1 if product i is produced in t
 285

10
The Supply Chain as a Dynamical System
Kendell R. Jillson, Eduardo J. Dozal-Mejorada, B. Erik Ydstie

In this paper we approach inventory and flow control in complex supply chains
using methods developed in the area of process control. Manufacturing sites, dis-
tribution centers, retailers and customers are modeled as networks of semiau-
tonomous, interacting activities. Inventories and flows of orders and products are
controlled using distributed decision making. It is well known that feedback inter-
action in distributed, dynamical systems can give instability when the information
flow is not properly organized. In this paper we show that the instabilities can be
stabilized and the supply chain optimized using dynamic feedback control. We in-
troduce new methods for inventory control using frequency response analysis. We
discuss an approach for feedback scheduling of assembly, disassembly and repack-
aging lines. Feedback scheduling is a relatively new technique which has roots in
system dynamics rather than optimization. We finally develop a method for flow
control which optimizes the intrinsic value of a business by managing inventory
and flows so that costs are minimized.

10.1
Introduction

Significant benefits are realized when internal and external business relations are
integrated. Organizations move in this direction by deliberately engaging resources
to improve their supply chain management capabilities. Good management of
business-to-business (B2B) relationships, in addition to business-to-consumer re-
lationships have become necessary to ensure smooth product and service flow. Re-
sources and information must be available and cost must be reduced. The most
important enablers for this business trend are enterprise research planning (ERP)
tools provided by companies like ORACLE, SAP, J.D. Edwards, and others.
It does not suffice to focus on the internal processes alone. Upsets are created
by factors beyond the control of a single company and business environments may
change very rapidly. Advanced planning systems (APS) have been developed to link
the database capabilities of the ERP system to market forecasts and process models.

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
286 10 The Supply Chain as a Dynamical System

Such tools enable a company to evaluate scenarios and respond to changes in the
market place. Robust and near optimal performance can be expected when these
tools are combined with feedback control to give improved agility. Computational
complexity can be avoided and structural stability can be achieved when informa-
tion is available throughout the supply chain so that distributed and structurally
coordinated decision making can take place on a routine basis [21, 40, 41]. By struc-
turally coordinated we understand that the architecture of the supply chain control
system has been designed so that distributed decision makers are enabled to make
quick and optimal decisions without referring to a central coordinator.
To move in the direction of optimal distributed decision making it is necessary
to view the supply chain system as a dynamical system. Therefore, stability, robust-
ness, and optimality must be embedded in the control laws that govern the system
behavior. Three distinct aspects of the supply chain management problem must be
addressed.
1. Inventory control.
2. Assembly, packaging, and disassembly.
3. Flow control and load balancing.

The models we develop are based on a mix of discrete and continuous model-
ing approaches. The developments are based on the idea that there are capacity
constraints in the form of constraints on inventory as well as on processing rate.
Successful supply chain management systems recognize constraints and ensure
that order, service, and product flows operate within their limitations. The follow-
ing constraint is fundamental and cannot be violated:


τi di < 1,
products

where τi (processing time/unit (h)) is the time it takes to process a product (assem-
ble pack and ship), and di is the rate at which orders are received for product i.
Basically the system has to be able to process orders faster than they are received.

The backlog grows as the load L = products τi di increases. Huge delays and
poor customer satisfaction result when L is close to 1. We interpret this constraint
in the context of the fluid model as constraints on shipping and receiving rates. In
other words the capacity of the system is limited by its ability to process and ship
orders.
Another fundamental limitation in retail is that the storage capacity is limited so
that
0
Imin
I
Imax ,
where I is the inventory level. The supply chain management system must be able
to operate within these constraints so that customer expectations are met while
inventory is not too high.
 10.2 Literature Review 287

10.2
Literature Review

A supply chain is an integrated system of suppliers, manufacturers, distributors,

and retailers, which share a common objective of adding value to a product or
service through transformation and transportation [4, 34, 36, 35]. The process of
planning and implementing strategies for controlling the inventories and flow of
products and orders throughout the supply chain system is called supply chain
management (SCM). In this chapter, we briefly review how the field of dynamic
analysis of supply chain systems got started, what problems have been addressed
and what tools are being implemented.

10.2.1
How did the Field Start?

It is widely acknowledged that Forrester provided a key contribution in the dynam-

ical system theory for business decision making [8]. His studies typically involved
a one-product supply chain with no branching. As limited as the case study was, it
still allowed Forrester to observe that a sudden increase in product demand at the
final consumer level can result in amplification of demand throughout the supply
chain system. The amplification increases in severity as the information progresses
up the supply chain to the manufacturer. Forrester’s work is important since he
demonstrated a link between business decision making and dynamical systems
theory, thereby showing that complex dynamics can be understood in terms of
feedback stabilization and equilibrium theory. He laid the foundation for fruitful
developments that continue to have significant impact to this day [13, 35].
Simon [31] provided other key contributions in business dynamics. He used
ideas based on feedback and dynamical systems theory to model industrial dynam-
ics and process decision making. Professor Simon received the 1970 Nobel Prize
in Economics “for his pioneering research into the decision-making process within
economic organizations.” Simon studied enterprise dynamics using Laplace trans-
form methods and developed the separation principle that plays a critical role in
the development of modern process control systems.
Forrester [8] developed the “MIT Beer Game” to illustrate demand amplifica-
tion. Four players make independent decisions regarding their inventories without
consulting other members of the supply chain. The players therefore have ade-
quate local information but poor global information, typically regarding only their
incoming orders as viable information. The game demonstrates that the order vari-
ance increases and amplifies as one moves up the supply chain as predicted by
Forrester. The game shows that the player’s decisions contribute to demand ampli-
fication since they do not account for how their decisions affect their teammates’
decisions. During simulations “emotions run high and many players report feel-
ings of frustration and helplessness” [35].
Alternative monikers for Forrester’s demand amplification problems have been
proposed. Lee et al. [18] argue that it is the bullwhip effect which tends to be mis-
288 10 The Supply Chain as a Dynamical System

understood or ignored by industry. They explain four causes for the bullwhip: de-
mand signal processing, the rationing game, order batching, and price variation.
They propose that the problem can be solved if manufacturers are given access to
demand data across the supply chain so that they can provide better forecast of
future demands. Hayes and Wheelwright [14] describe four stages of manufactur-
ing competitiveness which are useful in dealing with demand amplification. They
point out that one common misunderstanding is that SC managers believe that the
whip-lash effect is not due to instabilities in the feedback loops arising from their
own decisions. They believe that it is an unfortunate and unavoidable fact of life.

10.2.2
What Needs and Areas are Being Addressed?

Supply-chain management is challenging because it is necessary to bridge tradi-

tionally disjoint areas such as product development, manufacturing, and financial
planning. Common problems include (1) demand planning, (2) production plan-
ning, (3) production scheduling, (4) distribution planning, (5) coordination and in-
tegration, (6) demand forecasting, (7) disturbance forecasting, amongst others [27].
These diverse activities must be integrated to improve competitiveness by differen-
tiation, efficient manufacturing, and rapid penetration of diverse markets. In order
to achieve these goals, quantitative analysis, modeling, design, optimization, and
control have been identified as essential tools for supporting the decision-making
process [41].

10.2.2.1 Environmental
Hugo and Pistikopoulos [16] consider coupling the environmental concerns with
economic criteria in formulating a multiobjective optimization problem for sup-
ply chain management. The authors combine concepts from multienterprise SCM
and life cycle assessment (LCA) using mixed integer linear programming (MILP)
models to support strategic planning decisions. The decisions include the choice of
processing technologies, allocating technologies to potential sites, capacity expan-
sion, and implementation of new technologies over a given horizon [16]. The key
ingredient is the incorporation of the environmental performance objective. The
optimization objective function is based on the Eco-Indicator 99 model which is
incorporated into the quantitative LCA part of the model.

10.2.2.2 Pharmaceutical
Papageorgiou et al. [22] select product development and market introduction strate-
gies coupled with capacity and inventory planning using MILP. The approach takes
into account active ingredient manufacturing such as the two-stage production of
the high purity, high value chemicals, and global trading structures. One of the
main goals is to capture production and distribution capabilities accurately with-
out considering detailed scheduling. Maravelias and Grossmann [20] single out the
process of effectively selecting, developing, and manufacturing new products as
one of the greatest challenges facing the pharmaceutical industry. In order to solve
 10.2 Literature Review 289

this problem, the authors propose the use of a multiperiod MILP model that maxi-
mizes the ENPV of multiple projects. The specific model developed by the authors
takes multiple tradeoffs as inputs and determines which products should be tested,
test schedules, design decisions for the process network, and production profiles
for the different scenarios defined by the various testing outcomes [20].

10.2.2.3 Semiconductors
Wang et al. [38] present a model predictive control (MPC) approach for supply chain
management in the semiconductor industry. The MPC maintains inventory levels
at prespecified setpoints while a Kalman filter mitigates the effect of integrating
noise due to variations in customer demand. The selection of the Kalman filter
gain provides the flexibility to improve performance and robustness. The authors
conclude that MPC is a viable, flexible, and powerful tool for making daily deci-
sions in SCM in semiconductor manufacturing. Schwartz and Rivera [29] improve
the consumer demand modeling by implementing control relevant prefilters to
provide demand forecasts to a tactical MPC-based inventory management policy. A
systematic approach for generating these prefilters is presented. The ability to inte-
grate demand modeling and inventory control allows the use of low-order models
which result in efficient computation.

10.2.2.4 Energy
Goel and Grossmann [11] develop a stochastic programming model to plan de-
velopment of gas and oil fields when there are uncertainties in reserve esti-
mates. They consider the decision dependence of the scenario tree and present
a decomposition-based approximation algorithm. The proposed method yields
higher expected net present value (ENPV) solutions than a traditional determin-
istic approach. Goel and Grossman [12] extend the model to decision scenarios
which have uncertain future and include the influence of information discovery.

10.2.3
What Modeling Approaches and Solutions Exist?

Beamon [2] reviews and discusses multistage supply chain models. The models
surveyed in her work are divided into four categories: (1) deterministic analyti-
cal models, (2) stochastic analytical models, (3) economic models, and finally (4)
simulation models. Beamon’s review shows that there is a growing number and
diversity of models developed for supply chain optimization. Shapiro [30] presents
four overlapping approaches which he identifies as strategic challenges of SC plan-
ning and modeling:
• enlarging the scope of SC planning models;
• introducing strategy theories into data-driven optimization
models;
• standardizing scenario planning while incorporating
stochastic programming and modeling;
• expanding business processes to exploit fact-based analysis
of strategic plans.
290 10 The Supply Chain as a Dynamical System

Significant time and resources have been invested in trying to lay the foundations
for stronger and more powerful tools to address the second and third items. The
classic approach tries to forecast future inventory demand using statistical trend-
ing and “best fit” based on historic demand. For example, Disney et al. [6] use an
auto regressive moving average (ARMA) demand model to analyze 16 real-world
demand patterns from the consumer goods industry. They show that these meth-
ods can be used to prevent bullwhip and reduce inventory holdings. Warburton
and Disney [39] develop differential equations for inventory control and develop
an ordering policy to obtain the continuous time domain evolution of the free sys-
tem response. Using the simulated response of the system, the authors calculate
the bullwhip effect and inventory level variation which result from the continuous
time representation of the control policy. The resulting signals are compared to
known results for the discrete manifestation. The comparison shows the discrete
time solution to be managerially equivalent to the continuous case.
Tsiakis et al. [37] propose a mathematical programming formulation to address
the complexity of supply chain network design and optimization. Their approach
considers uncertain product demands, flexible production facilities, flexible trans-
portation, and variable delays. Product demand uncertainty is managed by setting
safety inventory levels adapted to the forecast.
In an interesting paper, Singhvi et al. [32] use pinch analysis to maximize profit
while satisfying demand over a specified time horizon. By providing operating tar-
gets, the method helps to alleviate the combinatorial complexity faced within dis-
crete mathematical programming. The authors state that the pinch analysis pro-
vides insight by balancing “quality” versus “quantity.” Riddalls and Bennett [26]
review optimal control applied to batched production and its effect on the bullwhip
effect. A differential equation model of a production–inventory system is coupled
with optimal control. Optimal responses to demand fluctuations are generated by
the control system. The authors also use the model to mimic the rational behavior
of supply chain managers to quantify bullwhip due to variable batch size. The am-
plification is linear with respect to the remainder of the ratio of demand rate and
batch size. The result is not surprising. When batches produce a large remainder
it is cost effective to produce a new batch rather than a fraction of a batch.
Disney and Grubbstrom [7] analyze inventory and order related cash flows result-
ing from a generalized Order-Up-To (OUT) control policy. They use a combination
of z-transforms, net present value calculations, and probability density functions in
their analysis. They show that the OUT policy can give better dynamic performance
when proportional controllers are included in the feedback loop.

10.2.4
Back to Dynamics

Significant advances in supply chain management have been made by treating SCs
as dynamical systems. Sterman [35] gives an overview of dynamics in business de-
cision making research and Geary et al. [10] present a historical review of the bull-
whip effect from a dynamics perspective. De Souza et al. [5] argue that supply chain
 10.2 Literature Review 291

dynamics increase cost. At the same time, Badell [1] notes that a dynamical econ-
omy continuously increases production alternatives and hence makes the empiric
management of the financial resources intractable. A tradeoff between increasing
production alternatives and decreasing cost is expected. Complicating this tradeoff
is the fact that presently enterprises are depicted as integrated networks of coop-
erating companies rather than isolated hierarchical ones [16]. This depiction leads
directly to information and material delays which are considered very important
contributors to the bullwhip effect in the supply chains.
Perea-Lopez et al. [23] model decentralized supply chains and develop systematic
decision-making processes to improve the modeled supply chains using dynamic
analysis. The main goal is to construct a model which allows managers to design,
evaluate, and compare the performance of various control policies for periodic fluc-
tuations and unexpected arrival of orders. The model captures system dynamics by
considering material and information flow within the chain and considers supply
chains to be decentralized systems with strong global coordination. Echelons in the
chain make decisions at the node level with only local information. The study is un-
dertaken with the goal of analyzing the impact of various heuristic control laws on
the performance of supply chains consisting of multiproduct, multiechelon distri-
bution networks, and manufacturing sites with single-unit, multiproduct, nonded-
icated batch, or continuous processes. The work analyzes the effects of operational
costs, customer satisfaction level. and inventory stability on SC performance.
In trying to build on their heuristic work, Perea-Lopez et al. [24] present an MPC
method to determine the optimal decisions in order to maximize profit. The sup-
ply chains under consideration are multiproduct, multiechelon distribution net-
works with multiproduct batch plants. The work focuses on developing a discrete
time MILP dynamic model that is able to consider all the flows within the chain,
both material and informational in nature. An optimization approach is presented
which considers all the elements of the supply chain as well as their interactions.
The rolling horizon implementation allows decision variables to be updated when-
ever current operating conditions change. Lastly, the authors present a comparison
between centralized and decentralized control showing that the former yields bet-
ter profit results when all information is considered centrally.
Pin et al. [19] model information and material flows in supply chain system and
develop a time series model for system flows. Transfer functions are obtained for
every node in the chain by z-transforms. The supply chain system is modeled by
linking together the transfer functions. The inventory level is managed to satisfy
the demands by ordering products from suppliers using proportional, proportional
integral, and cascade control. The cascade control policy provides inventory control
without causing bullwhip and raises the customer satisfaction by providing more
active tracking of the customer demand.
Yu and Cassandras [42] propose to use a stochastic fluid model (SFM) to simu-
late supply chains. The SFM captures stochastic variations in the supply processes
by specifically treating all fluid rates as random processes. The performance analy-
sis takes into consideration factors such as the structure of the underlying system
in addition to the nature of the performance metrics of interest. They note that
292 10 The Supply Chain as a Dynamical System

an SFM can be used to determine settings for control and optimization purposes,
even when it fails to provide accurate performance estimates. The authors point
out that feedback systems can provide accurate solutions even if the corresponding
performance cannot be estimated with accuracy. Furthermore, by utilizing pertur-
bation analysis along with the SFM, gradient estimators of throughput and buffer
overflow metrics with respect to production control parameters are derived.
Ydstie [41] introduced the idea of adaptive enterprises and distributed decision
making. The adaptive enterprise meshes decentralized and centralized decision
making with dynamical systems theory. Flexible systems result that allow innova-
tion and continuous improvement. The decentralized and distributed systems have
been developed to improve flexibility, optimize inventory, and minimize wastage.
In order to achieve these goals, adaptivity and learning in complex systems and
networks need to be better understood. The adaptive enterprise needs to be flex-
ible and quite agile so that it can respond to challenging needs as the boundary
conditions and even its internal structure changes.
The main issue can be summarized as follows: A given customer places orders
at the preferred retailers; the retailers keep track of what is being sold (inventories)
and places “enough” orders to fulfill the customers’ product demands. This prob-
lem superficially seems to be a simple one; however, it has been shown to be rich
in dynamics and difficult to address because most supply chains involve thousands
of customers and products. Thousands of people and processes are involved in the
manufacture and transportation of these products. Customers and suppliers have
conflicting objectives, the supplier wants a high price for a cheap product whereas
the customer wants a high quality product for a low price.

10.3
Problem Formulation

A supply chain is to be thought of as a “network of organizations involved, through

upstream and downstream linkages, in different processes and products [4].” The
objective of the supply chain is to create value through a sequence of operations
which we refer to as activities. Such activities include transportation, storage, man-
ufacture, routing, and shipping/receiving. In this context, Stadtler and Kilger [34]
define supply chain management (SCM) as “the task of integrating organizational
units along a supply chain and coordinating materials, information, and financial
flows in order to fulfill (ultimate) customer demands with the aim of improving
competitiveness of a supply chain as a whole.”
The SCM perspective therefore includes the idea of two or more legally separate
partners working together toward a common goal within a business sector. A typi-
cal supply chain with raw material suppliers, processing plants, plant warehouses,
distribution centers, retailers, and end customers is depicted in Fig. 10.1.
 10.3 Problem Formulation 293

D i s t r i b u t io n
S up p lie r Plant W a r eh o u se Ret ail er Consumer
Ce n t e r

Fig. 10.1 Three interconnected supply chains moving products

from the raw material suppliers to the (end) consumers.

The problem we address is the following:

How does one manage the flow of orders and products so that customer
satisfaction is high while maintaining low inventory? Furthermore,
disturbances should be filtered out and demands should be appropriately
forecasted so that the plant can adapt its production schedule without running
idle or running up large backlogs.

The logistics are difficult to manage since there are hundreds or even thousands
of possible different products, material and information delays, and uncertainties
that depend on issues that cannot easily be controlled, e.g., the weather, how end-
customers respond to advertising and changes in fashion, macroeconomics, and
difficulties in rapidly changing production capacity.
Consider the storage capacity shown in Fig. 10.2. It may correspond to the storage
at the factory, the distribution center, or the retailer. The inventory is denoted by I ,
the order backlog by O, the receiving rate is denoted by φr , the shipping rate is
denoted by φs , the incoming order rate is denoted by φso . Inventory, receiving rate,
and shipping rate are capacity limited so that

0
Imin
I
Imax ,
0
φr,min
φr
φr,max , (10.1)
0
φs,min
φs
φs,max .

The inventory balance can now be written as

dI
= φr − φs . (10.2)
dt
294 10 The Supply Chain as a Dynamical System

Transportation and Order delays

Orders φos,j+1

Receiving φr,j-1
Shipping φs,j
S Distribution center

Orders φor,j
Storage Capacity j

Fig. 10.2 Storage capacity in node j of a supply chain.

Receiving and shipping rates can further be divided into subcomponents which
represent shipping to different customers and receiving from different suppliers
so that


φr = fr,i ,
receiving


φs = fs,i , (10.3)
shipping

where the summations are over all shipping and receiving lines. The receiving is
controlled by sending orders at the rate


φro = o
fr,i ,
suppliers

where the summation is over all suppliers. In a system with no delays we have
φr = φro .
Equations (10.3) show that the plant, distribution center, and retailers may have a
large number of different suppliers and customers. The load on these supply lines
must be balanced to maximize end customer value. Similar balances are written
for the orders O so that we obtain
dO
= φso − φs , (10.4)
dt
where


φso = fs,i
customers
and the summation is over all customers. It is worth pointing out that in this model
the orders are immediately sent to the supplier. The order book fills as orders are
received and empties as goods are shipped.
 10.3 Problem Formulation 295

We now introduce some definitions motivated by control theory. Incoming or-

ders from customers φso and receipts φs from suppliers are disturbances. Shipping
rates φs and orders sent to suppliers fro are the controls. The state of the dynamical
system is given by the vector of I = (I1 , . . . , Inp ) of the different stock keeping
units (SKU) and the vector O = (O1 , . . . , Onp ) of outstanding orders. These vec-
tors may have several thousand entries. There are two control problems:
• control the inventories I and O. In general these should be
kept as low as possible.
• choose suppliers so that cost is minimized.

One difficulty in supply chain management arises from the fact that product in-
ventory is controlled indirectly by sending orders to suppliers. Shipping may be
delayed due to limited capacity, transportation, and slow response.
To introduce the subject of automated supply chain management, we consider a
system with no delays and sufficient capacity. It is then possible to achieve perfect
control. To see that this is the case we note that the mappings
 
dI ∗
φr − φs + → (I − I ∗ ),
dt
 
dO ∗
φso − φs + → (O − O ∗ )
dt
are passive. A passive operator is an operator which preserves sign so that if the
mapping u → y is passive then there exists a storage function W so that
dW

uT y, W  0.
dt
It is easy to prove that a supply chain system is passive using the input–output
signals defined above and the storage function
1 1
W = (I − I ∗ )T (I − I ∗ ) + (O − O ∗ )T (O − O ∗ )  0.
2 2
By differentiating the storage function we have
   
dW dI ∗ dO ∗
= (I − I ∗ )T φr − φs + + (O − O ∗ )T φs o − φs +
dt dt dt
and the result follows.
The main advantage of passivity is that it is a property of the system which is es-
tablished independent of the control method of choice. Any strictly passive control
can be used to control a passive system using the relations
dI ∗
φro = φs + Cs (I − I ∗ ) + ,
dt
dO ∗
φs = φso − C0 (O − O ∗ ) + ,
dt
where Cs (·) is any strictly passive shipping policy and C0 (·) is any strictly passive
ordering policy. Examples of strictly passive controls include
296 10 The Supply Chain as a Dynamical System

• PID control,
• adaptive control,
• back stepping nonlinear control,
• linear quadratic finite and infinite horizon control,
• nonlinear programming based finite horizon control.

We will use proportional control to illustrate our approach. In this case, we get the
following strictly passive controls:

dI ∗
φro = φs − Kc (I − I ∗ ) + , (10.5)
dt
dO ∗
φs = φso + Kc (O − O ∗ ) + , (10.6)
dt
where I ∗ is the level of safety stock and O ∗ is the setpoint for the backorders.
In the just-in-time (JIT) [36] supply chain, the aim is to reduce capacitance to
improve agility and reduce cost. In the ideal case, we set O ∗ = 0 and I ∗ = 0. Such
a strategy will not work in retail since zero capacity implies that the product has to
be assembled and shipped before delivery. However, the approach can be applied
when the customer can wait for delivery. JIT is ideally suited for e-business. Orders
are placed before delivery is expected and capacitance can be moved back in the
supply chain, sometimes all the way to the production plant. For example, after
the customer has placed an order for a Dell computer, necessary components are
purchased from Intel and other suppliers off trucks at the loading dock and brought
into the assembly plant. Assembly is carried out to customer specifications and
shipped within days.
In retail it is necessary to carry inventory since the customer expects to be able to
carry the purchase out as soon as she has paid for the item. However, giant retailers
like Wal–Mart and COSTCO have been able to reduce capacity and move toward
JIT by eliminating distribution centers and take direct delivery from the factory to
the retail store. COSTCO and Sam’s Club further streamline operation and reduce
supply chain cost by only carrying limited range of products which are tuned to
meet specific seasonal demands. In this way, the retailers are able to reduce risk
and cost by shifting supply chain and inventory management to suppliers such as
Procter & Gamble, Johnson & Johnson, Unilever, etc.
JIT cannot be implemented exactly along a supply chain with shipping, reload-
ing, repackaging, and production. Each of these steps incurs transaction delays so
that

φr (t) = φro (t − td ), (10.7)

where td is the delay which is incurred by finite communication speed in the order-
ing system, finite transportation rate, and limited supplier availability due to poor
management of the supply chain. The delays are state and supplier dependent and
the ordering policy (10.5) will not work, as we will see below since orders are placed
as long as the inventory is low, leading to significant overstocking.
 10.3 Problem Formulation 297

1 .6 5

1 .6 U S M a n u f a ct u r es: rati o o f i n v e n t orie s to sh ip m e n t s

1 .5 5

1 .5

1 .4 5

1 .4

1 .3 5

1 .3
US Census Bureau
1 .2 5

1 .2
1990 1992 1994 1996 1998 2000 2002

Fig. 10.3 United States Census Bureau data showing a steady

decline in inventory levels due to throughput increases of about
3% per year corresponding to growth in GDP.

Overstocking is prevented by including outstanding orders in current inventory.

This approach makes good sense since these items are due. Let M represent out-
standing orders, then
dM
= φo − φs , M(0) = 0. (10.8)
dt
The feedback control policy (10.5) is then implemented using the following modi-
fication:
dI ∗ dφso
φro = φso + Cs (I + M − I ∗ ) + + , (10.9)
dt dt
where we have included feedforward from the shipping order rate and its deriva-
tive. The stability result is now developed using the storage function
1 1
W = (I + M − I ∗ )T (I + M − I ∗ ) + (O − O ∗ )T (O − O ∗ )  0.
2 2
Stability follows as before provided that the inventory setpoint I ∗ is chosen large
enough so that market demands can be met while products are in transit, namely
I ∗ (t)  M(t). (10.10)
We will address this issue in the next section.
Retailers sometimes adapt the inventory so that more stock is kept of products
that are in high demand. This leads to the setpoint modification
I ∗ (t) = Nd φro (t). (10.11)
The strategy allows a storage location to hold Nd time periods (e.g., weeks) of stock
on hand at current ordering rate. Figure 10.3 shows that most US companies do
not use such a policy, however, they try instead to run down inventories even as
demand increases. US manufactures have been able to streamline operations while
increasing throughput over a long period of time by controlling the supply chain.
298 10 The Supply Chain as a Dynamical System

Order flow

Raw material suppliers

End customers
Pr oduc t an d s er vi c e fl ow

Fig. 10.4 A linear supply chain consisting of interlinked capacities.

Fig. 10.5 Flowsheet of the linear supply chain utilized in the “Beer game.”

The model and inventory control systems extend to multiechelon supply chains.
Each unit is given an index j = 1, . . . , nU , where nU is the number of units as
shown in Fig. 10.4. The state of the supply chain system is given the vectors Ij =
(I1 , . . . , Inp )j of stock keeping units (SKU) and the vector Oj = (O1 , . . . , Onp )j of
outstanding orders. The controls do not necessarily become more complex since
they are decentralized.

10.3.1
The Beer Game

The “MIT Beer Game” [35] represents a linear supply chain with retailer, distribu-
tor, wholesaler, and plant warehouse as shown in Fig. 10.5. Orders are represented
by the flows fi0 coming from the left. Product flows are represented by the flows
fi from the right. The system is called the Beer Game since its motivation can be
found in the manufacture, distribution, and sales of beer.
The orders suffer a delay of 1 week and the products suffer a delay of 2 weeks.
The delays are modeled as a “bucket brigade” where the orders and products accu-
mulate for a week before they are sent to the next stage for processing. The factory
is assumed to have “infinite capacity” so that it can meet any demand issued from
the retailer. The complete model is given by
 10.3 Problem Formulation 299

Retailer: The disturbance is given by the customer demand fo (1).

Dynamical model:
dI (1)
= f (2) − f (1),
dt
dO(1)
= fo (1) − f (1).
dt
Constraint:
f (1) = min[I (1) + f (2) , O(1) + fo (1)].
The lower bound ensures that the storage is nonnegative so that goods are not
shipped if the current inventory is empty.
Delays:
f (t, 2) = i(t, 2) = f (t − 1, 2) ,
f (t, 2) = i(t, 2) = f (t − 1, 2).
Distribution: The disturbance:
fo (t, 2) = o(t, 2) = fo (t − 1, 2).
Dynamical model:
dI (2)
= f (3) − f (2),
dt
dO(2)
= fo (2) − f (2).
dt
Constraint:
f (2) = min[I (2) + f (3) , O(2) + fo (2) ].
Delay:
f (t, 3) = i(t, 3) = f (t − 1, 3) ,
f (t, 3) = i(t, 3) = f (t − 1, 3).
Wholesale: The disturbance:
fo (t, 3) = o(t, 3) = fo (t − 1, 3).
Dynamical model:
dI (3)
= f (4) − f (3),
dt
dO(3)
= fo (3) − f (3).
dt
Constraint:
f (3) = min[I (3) + f (4) , O(3) + fo (3) ].
300 10 The Supply Chain as a Dynamical System

Delay:
f (t, 4) = i(t, 4) = f (t − 1, 4) ,
f (t, 4) = i(t, 4) = f (t − 1, 4).
Manufacture: The disturbance:
fo (t, 4) = o(t, 4) = fo (t − 1, 4).
Dynamical model:
dI (4)
= f (5) − f (4),
dt
dO(4)
= fo (4) − f (4).
dt
Constraint:
f (4) = min[I (4) + f (5) , O(4) + fo (4) ].
Delay:
f (t, 5) = i(t, 5) = f (t − 1, 5) ,
f (t, 5) = i(t, 5) = f (t − 1, 5) = fo (5) .

We will first use the ordering and shipping policies presented in (10.5) and (10.6)
with the setpoint modification (10.11). Inventory levels therefore increase during
periods of high demand. These features are implemented so that
fo (j ) = φr (j ) + Kc (I (j ) − I ∗ (j )), j = 1, 2, 3 (10.12)
with controller gains Kc = 0.25 in all echelons and
I ∗ (j ) = fo (j )Nd + O(i), j = 1, 2, 3. (10.13)
The desired stock level is chosen so that Nd = 3, corresponding to 3-week demand
at the current demand rate. The predictive term dI ∗ /dt has not been included in
the control. The feedback control policy (10.12) is implemented with the constraint
fo (j ) = max[0, fo (j ) + K(I ∗ (j ) − I (j ))]. (10.14)
The lower bound has been added to Eq. (10.5) to prevent “negative orders,” which
would represent returning goods.
The simulations represent a 100-week period. The retailer demand (fo (1)) is ini-
tially 100 units/week. After 10 weeks, the demand is stepped up to 200 units/week,
and then back to 100 units/week at week 40.
Case 1. The inventories, orders, and overall service level are shown in Figs. 10.6–
10.8. Large transients result due to increasing demand and stock levels. Upstream
nodes ramp up shipping rates to account for increases in demand as well as in-
creases in stock levels. The amplification effect propagates through the system all
 10.3 Problem Formulation 301

4
2.5
x 10 Inventory Level vs Time
Retailer
Who l e sal er
2 Distributor
Manufacturer
Inventory Level

1.5

0.5

-0.5
0 20 40 60 80 100
Week #

Fig. 10.6 Case 1: Inventory level when backorders are not

included in the policy where negative inventory level indicate
unmet orders.

Ord ers Mad e vs Time

6000 Retailer
W holes aler
Dis tribut or
5000 Manufac turer

4000
Orders Made

3000

2000

1000

0
0 20 40 60 80 100
Week #

Fig. 10.7 Case 1: Orders placed.

the way back to the factory and the supply chain saturates with overstocked goods.
It takes a long time to bring the system back to steady state after the demand
drops off.
Negative inventories in the three upstream nodes refer to backlogs of orders.
Backorders at the retailer node give lost sales since the customer instead of waiting
decides to buy a substitute product. In the US, lost sales due to poor stocking at the
retailer level account for 5–10% of total sales. The major problem with the ordering
policy developed here is that it continues to order as long as the inventory is low,
which leads to significant overstocking.
302 10 The Supply Chain as a Dynamical System

1
Total Service Level vs Time

Total Percentage of Demand Met

0.98

0.96

0.94

0.92

0.9
0 20 40 60 80 100
Week #

Fig. 10.8 Case 1: Service level.

3000
Inventory Level vs Time
Retailer
W holesaler
2000 Distributor
Manuf acturer
Inventory Level

1000

-1000

-2000
0 20 40 60 80 100
Week #

Fig. 10.9 Case 2: Inventory levels when outstanding orders are

included in the policy where negative inventory levels indicate
unmet orders.

Case 2. Overstocking can be reduced by including stock on hand as well as out-

standing orders when we calculate the inventory. Stock on hand is calculated so
that
dM(1)
= fo (2) − f (2) , M(0) = 0.
dt
The feedback control policy (10.14) is now implemented so that
fo (j ) = φr (j ) + Kc (I (j ) + M(j ) − I ∗ (j )), j = 1, 2, 3,
where the setpoint is adapted as in (10.13) with the constraints given by (10.14).
This policy results in the dynamics shown in Figs. 10.9–10.11. The demand ampli-
fication is still present, but less pronounced. There is still significant loss of sales.
 10.4 Inventory Control 303

2000
Orders Made vs Time
Retailer
W holesaler
Distributor
1500 Manuf acturer
Orders Made

1000

500

0
0 20 40 60 80 100
Week #

Fig. 10.10 Case 2: Orders placed.

1
Total Service Level vs Time
Total Percentage of Demand Met

0.98

0.96

0.94

0.92

0.9
0 20 40 60 80 100
Week #

Fig. 10.11 Case 2: Service level.

The main problems with the heuristic policies developed here are that the feed-
back gains have not been optimized and that inventories are extremely high during
high demand periods, contrary to the expectations provided by Fig. 10.3 and the
resultant cost of keeping inventory. Therefore, we believe that filtering is needed to
ensure better optimization.

10.4
Inventory Control

Extra storage allows the supply chain to absorb unexpected changes in the demand
and supply so that the supplier can operate with more flexibility in a fluctuating
market. Consider, for example, the distribution center shown in Fig. 10.12. Varia-
304 10 The Supply Chain as a Dynamical System

Maximum Inventory

Current Inventory level I

Incoming ordering rate

Maximum rate

Minimum Inventory M i n i m u m ra t e

Time Time

Receiving φr Shipping φs
D D is t r ib u t i o n ce n t e r

Fig. 10.12 Distribution center with bounds on receiving and shipping rates.

tions in demand can be met while the disturbances are dampened by using inter-
mediate storage to absorb the demand variations. Demand can be met all the time
while keeping the outgoing order rate to upstream suppliers smooth if the storage
is sufficiently large.
In this section, we show how to eliminate bullwhip by using an idea developed
for surge tank control in chemical plants. The control policy is optimal in the sense
that it gives the best damping possible for demand variations with a flat power
spectrum while satisfying the constraints.
Consider a single distribution center modeled using Eq. (10.2) so that
dI
= φr − φs .
dt
The product and order flows satisfy the constraints (10.1) which we rewrite so that
• Imax is the maximum allowable inventory;
• Imin is the minimum acceptable inventory;
• a is the minimum customer order rate;
• b is the maximum customer order rate.

The objective is to develop a control system which maximizes the filtering effect
of the intermediate storage with respect to order variations. We use an ordering
strategy adapted from policy (10.8)
φro = a − Kc (I − I ∗ ). (10.15)
The main differences are that we use a fixed feedforward a and that we choose
the setpoint I ∗ and the feedback gain Kc so that the damping is maximized while
satisfying the constraints. We assume for now that there are no delays in receiving
and shipping so that
φro = φr , φso = φs .
 10.4 Inventory Control 305

At this moment we do not include the backorders M since these are easy to imple-
ment later on.
By combining the controller (10.15) and inventory balance (10.2) we have the
closed loop
dI
= a − Kc (I − I ∗ ) − φs .
dt
Solving this equation for steady-state inventory gives
a − φs
I = I∗ + .
Kc
The last term is negative for positive gain since φs  a. The extreme demand
conditions give
I = I∗ for φs = a, (high demand)
I =I ∗ + a−b
Kc for φs = b, (low demand).
To satisfy the inventory constraints we therefore choose

I ∗ = Imax
b−a
Kc = .
Imax − Imin
To show that this is indeed the best possible feedback control solution for distur-
bances with a flat power spectrum we use frequency response analysis and model
the customer demand so that
φs = φ̄s + A sin(ωt),
where φ̄s = 12 (b − a) + a is the mean value of the expected shipping rate and
A = 12 (b − a) is the amplitude. The dynamics of the distribution center are now
modeled by the following system of differential and algebraic equations (DAE):
dI
= φr − φs ,
dt
φs (t) = φ̄s + A sin(ωt),
φr (t) = a − Kc (I − I ∗ ).
By solving the forced DAE system we get the frequency response
φr = B sin(ωt + ϕ) + φ̄s .
The amplitude and phase shift are defined so that
A
B=  ,
ω2
Kc2
+1

ϕ = tan−1 (−ω/Kc ),
T
ω= ,
2π
where T is the period of the sinusoid.
306 10 The Supply Chain as a Dynamical System

2,000 ≤ I 50,000 units

0≤ f s ,1 ≤ 1,000 units/day Incoming orders φos
500 ≤ f s, 2 ≤ 3,000 units/day
200 ≤ f s ,3 ≤ 1,500 units/day Shipping fs,1

Receiving φr
D Shipping fs,2

Shipping fs,3
Outgoing orders φor
Fig. 10.13 Storage capacity with one supplier, three customers, and the respective constraints.

The ratio of the output to the input amplitudes, called the amplitude ratio (AR)
provides a measure of the damping provided by the distribution center. We have
output amplitude 1
AR = =  2 .
input amplitude ω
+1
Kc2

The AR is small if Kc small. More generally we have

lim AR = 0, lim AR = 1.
Kc →0 Kc →∞

It follows that the damping is excellent if Kc is small and that there is no damping
if Kc is large. In order words, we should not control the inventory too tightly in a
supply chain system and we should definitely not use positive feedback. Instead,
we need to chose Kc as small as possible, still large enough to satisfy the storage
constraints imposed by the size of the distribution center. All these objectives are
achieved using the control law developed above.
A fundamental result in control theory states that the optimal control for a linear
system under disturbances with a flat power spectrum in fact is a linear controller.
This means that the system presented here is the best possible in the absence of
additional knowledge about the disturbance spectrum.

Example 1. The distribution center shown in Fig. 10.13 supplies three retailers at
varying rates. These satisfy the constraints

2000
I
55,000 inventory,

0
fs,1
1000 units/per day,
500
fs,2
3000 units/per day,
200
fs,3
1500 units/per day,
 10.4 Inventory Control 307

-0.2
10
Amplitude ratio

-0.4
10

-0.6
10

-0.8
10

-3 -2 -1 0 1
10 10 10 10 10
Frequency [rad/day]

Fig. 10.14 Amplitude ratio for a distribution center under feedback ordering policy.

and the total demand is

φs = fs,1 + fs,2 + fs,3 .
The minimum inventory corresponds to about 1/3 of the day’s shipping whereas
the maximum inventory corresponds to 10-days supply under the highest rate of
sales for all three customers.
The problem is adapted to the theory developed above when we write
2000
I
55,000,
700
φs
5500.
The latter equation describes the bounds for the total variation. The corresponding
controller has parameters
Kc = 0.0906, I ∗ = 5500, b = 55,000.
The optimal damping is shown in Fig. 10.14.
Consider, for example, a situation where the total sales are approximated by the
expression
φs = fs,1 + fs,2 + fs,3 = 3100 + 2400 sin(ωt)
with
2π
ω=
= 0.8976 rad/day.
7
This corresponds to a sales cycle with a period of 1 week. The corresponding am-
plitude ratio is
1
AR =  = 0.1004.
(ω/K)2 + 1
Therefore, the amplitude of the ordering rate is only one-tenth of the amplitude of
the sales. The phase lag is
 ω
P = 3100 + 240 sin(ωt + ϕ), ϕ = arctan − = −1.4702 days.
K
308 10 The Supply Chain as a Dynamical System

500
Inventory Level vs Time

400
Inventory Level
300

200

Retaile r
100 W ho lesaler
Distributor
Manuf acturer
0
0 20 40 60 80 100
Week #

Fig. 10.15 Case 3: Inventory level with optimal damping.

It follows that the maximum ordering rate lags the maximum demand by about 1.4
days. The PD controller introduced below provides a phase lead which provides a
mechanism to put in advance orders.

The controller addresses the problem associated with the bullwhip in two impor-
tant ways.
1. It provides a systematic method for tuning the feedback
gains so that these are optimal in the sense of providing
disturbance rejection while satisfying customer demand.
2. It provides a way to choose inventory setpoints.

The ordering policy with phase lead is implemented so that

b−a dφ o
φr0 = a − (I + M − Imax ) + τD s . (10.16)
Imax − Imin dt
A centralized decision maker may be needed to determine Imax , Imin , a, and b and
how these should be distributed throughout the supply chain. In the case of a single
stranded supply chain, these numbers correspond to the customer demand. In a
later section, we address the problem of how to distribute the demand to different
suppliers.
MIT Beer Game Case 3. In order to illustrate the performance of the new ordering
policy with optimal damping we revisit the beer game and run a third case study us-
ing a 100-week simulation of the modified game developed in the previous section.
The inventory capacity used to develop the controller is set so that the minimum
and maximum are set 50% above and below the safety stock level. The safety stock
is set at 3 weeks’ worth of supply material. The maximum and minimum demands
correspond to maximum and minimum step disturbances.
The inventory level, order rates, and service level for the simulation are shown
in Figs. 10.15–10.17, respectively. The controller works quite well as can be seen in
 10.5 Assembly, Disassembly, and Packaging 309

200
Orders Made vs Time
Retailer
Who lesaler
180 Distributo r
M anuf acturer
Orders Made

160

140

120

100

80
0 20 40 60 80 100
Week #

Fig. 10.16 Case 3: Orders placed.

2
Total Service Level vs Time
Total Percentage of Demand Met

1.5

0.5

0
0 20 40 60 80 100
Week #

Fig. 10.17 Case 3: Overall service rate.

the figures. The inventory levels are bounded and service is maintained at 100%.
To get better damping at the factory site it is necessary to install larger inventories.

10.5
Assembly, Disassembly, and Packaging

In process control it is customary to make a distinction between automation and

control. Automation concerns discrete feedback implemented in PLC control logic.
Control concerns the implementation of continuous feedback as exemplified by
the PID controller. Dynamical systems that incorporate elements of discrete and
continuous variables are called hybrid systems.
310 10 The Supply Chain as a Dynamical System

Unit 3 Unit 8

Unit 1 U n it 10

Unit 6 Unit 9

Unit 5 Unit 7

Unit 2 Unit 4

Fig. 10.18 In a repackaging center different products arrive at

Units 1 and 2 where pallets are opened and distributed to
different work stations where products are reassembled and
repackaged. The finished products finally arrive at Units 8 and
10 where they are shipped on for delivery to retailers or
distribution centers.

About 70–90% of all decisions made in process control are discrete. Discrete de-
cisions dominate during start-up and shut down of continuous production lines, in
batch control, safety interlock systems, and grade transition. Automation is needed
in supply chain systems whenever products are assembled, put on pallets for ship-
ping, repackaged for specialized campaigns, and small retailers. In this section we
develop an approach for automation in hybrid systems which is based on feedback
scheduling and on the idea of modeling the system (10.2) as one where we only al-
low work on one product at any given time. The material presented was developed
in [33].
Consider the assembly line shown in Fig. 10.18. Workstations receive items,
assemble, store, repackage, shrink-wrap, etc. The connections show that at least
one product type can flow in the direction indicated. Between each workstation,
there may be storage facilities where items are held until the workstation is avail-
able. The flow control problem then consists of organizing the flow of material
so that
1. Order backlogs are small,
2. Process constraints are satisfied.

A single unit in such a processing center can be seen in Fig. 10.19. The order level
is modeled using the order balance (10.4). It is convenient to rewrite this equation
so that
dxi
= di − σi (t)τi−1 ,
dt
where xi is the current backlog of orders at workstation i, di (units/time) is the
rate of incoming orders, and τi (time/unit) is the time required to process one
unit. We also note that in some applications it is convenient to let x represents
 10.5 Assembly, Disassembly, and Packaging 311

Order rate for products

d1 d2 d3 A,B,C,….

A B C

W o r k p r o ce s s f low
P
{A,B,C,B,B,C,A,…..}
Fig. 10.19 In this schematic work flows from left to right. The
machine can only work on one product at a time. The machine
is now working on product A.

the inventory in (10.2). We therefore use the generic symbol x to represent either
O or I .
Every time we switch from one product to another there is a setup time δi . During
setup the workstation is idle as it is being cleaned and readied for new products.
The control function σi (t) is equal to one when product i is processed and zero
otherwise. In the context of (10.4) we have


p
O(i) = xi
i=1

so that the backlog for the assembly/repackaging plant is equal to the sum of all
the backlogs at all workstations throughout the plant.
We assume that material is always available so that the order backlog is limited
by processing rate rather than availability of material when the scheduling theory
is developed. The policies will be robust with respect to upsets and lack of material
since they are based on feedback. We also note that we must have the following
load constraint satisfied:


p
τi di < 1. (10.17)
i=1

The plant must have capacity to produce faster than orders arrive! One or more
backlogs will inevitably increase without bound if the constraint (10.17) is not sat-
isfied, leading to reduced customer satisfaction, loss-of-sales, and long delay since
the machine is a bottleneck.
The scheduling problem is simplified considerably when we consider a special
class of scheduling policies called clearing policies. Suppose that the plant is pro-
ducing product A. The production schedule is said to be clearing if the production
run continues until the order queue for A is empty. The clearing policy can be mo-
312 10 The Supply Chain as a Dynamical System

tivated on the grounds that if we switch to another product before A is finished

then we lose time due to setup since we must later return to producing A. Let us
suppose that we decide to make product C. The production of C continues until all
backorders for C are processed. Now we can start to produce B or we may switch
back to A again, thus generating a production schedule

S = {A, C, B, A, B, C, B, A, ...}

in real time.

Control objective: Define a feedback policy which keeps the order levels (the back-
log) as small as possible.

The problem consists of finding the feedback policy which minimizes the order
backlog. Inventory control problems of this kind are discrete and they are called
scheduling problems. In this case we are faced with a hybrid problem since there is a
mix of discrete and continuous variables.
Let σ (t) = (σ1 , σ2 , ..., σp ) be the control signal, where σi = 1 if the unit is setting
up to or working on product i and σ0 = 0 otherwise. We will restrict the unit so
that it can only operate on one product at a time, i.e.,

1 for i = k
σi σk =
0 for i = k

By letting xi (t)  0 represent the order backlog and defining x(t) = (x1 , x2 , ..., xp )T
we have the state equation
dx
= F (x(t), σ (t)), x(0) = xo . (10.18)
dt
The vector field F (x(t), σ (t)) is defined so that
 
f1 (x1 , σ1 )
 .. 
F (x, σ ) =  . ,
fp (xp , σp )

where each row represents the right-hand side of the differential equation so that
 
σi
fi (xi , σ1 ) = di − max[t − (tk + δi ), 0] ,
τi
where tk are the times when we switch from one work order to another, i.e.,
 
t a control action takes place
tk = .
tk−1 otherwise

With this notation it is natural to define a sampled version

x(k + l) = (x(k), σ (k), Tk ),

 10.5 Assembly, Disassembly, and Packaging 313

where Tk = tk+1 − tk defines the control dependent sampling interval. The state
transition function satisfies
 
φ1 (x1 (k), σ1 (k), Tk )
 .. 
(x(k), σ (k), Tk ) =  . ,
φ1 (xp (k), σp (k), Tk )
where
σi
φi (xi , σ, Tk ) = xi (k) + Tk di −
max[(Tk t − δi ), 0].
τi
The rest of the section is aimed toward developing feedback control policies for the
hybrid control system defined above.

Definition 1 [Feedback Control Policy]. A control policy for the system described above
is a feedback control policy if
σ (k) = G(x(k)).

From this definition it is clear that a feedback policy depends on the current state,
in contrast to a feedforward policy (a schedule) where the control signals are chosen
ahead of time, i.e.,
σ (k) = G(t, x(0)).
The main advantage of feedback is that it adapts to changes so that it works well
even when there are upsets in the process. Model predictive control with receding
horizon is an example of a feedback control policy. Feedback is sometimes called
reactive scheduling.
We now define the notion of stability and performance before discussing the
relative merits of different types of feedback control policies.

Definition 2 [Stable Scheduling Policies]. A scheduling policy σ (x(k)) is stable if

there exists a constant M < ∞ so that
lim sup x(t)
M.
t

Stable scheduling policies ensure that material levels in storage remain bounded.

Definition 3 [Periodic State]. A scheduling policy σ (x(k)) is periodic with period T if

◦ (x(k), G(x(k), T ) = x(k).

The smallest period T ∗ is called the minimum cycle time.

Definition 4 [Load]. The numbers ρi = τi di are called the load due to product i. The
total load on a work station is defined so that


p
ρ= ρi
i=1
314 10 The Supply Chain as a Dynamical System

The necessary condition for the existence of a stable policy is

p
ρ= τi di < 1. (10.19)
i=1

10.5.1
Bounds on Storage Requirements

In this section we will develop upper and lower bounds on storage requirements
for a semicontinuous processing unit which works on p products. To simplify the
problem we make the following assumptions:
1. The setup times are sequence independent, i.e., δij = δki .
2. The policies we consider are clearing, i.e., once production
has started on a particular product it will continue until the
order buffer is empty.
3. No “idling” is allowed, i.e., the production rate is either 1/τi
or 0.

Let x̌ni represent the maximum backlog of product i at the beginning of production
run n as illustrated in Fig. 10.20. Let

x̂i = max{x̌1i , , ..., x̌li }

j
represent the maximum buffer level for a period T of l production runs. Let µi be
the length of the j th production run of product i.
j
µi = zi x̌ji , (10.20)

Machine works on product i

Machine sets up to work on product i

Product i
xi2
Machine works on another product
xi3
xi1 δi µi2

t1 t2 t3
Fig. 10.20 Trajectory of level of product i. x̌1i = backlog for
workorder i at beginning of first production run. µ2i = length of
production run 2 of workorder i. t1 = time at which production
run 1 starts.
 10.5 Assembly, Disassembly, and Packaging 315

where
τi
zi = .
1 − τ i di

The best policies result when all production runs are clearing; at the end of pro-
duction run of product k we have xk = 0. Consider Fig. 10.21. Suppose no produc-
tion run is clearing as depicted by the continuous line in Fig. 10.21(a). By translat-
ing this graph downward to the dashed line we obtain another valid trajectory with
x̂p (dashed line) < x̂p (continuous line). Therefore, at least one production run must
be clearing. Note also that every production run prior to a clearing run must also
be clearing. To see this, translate the continuous line just above the dashed line in
Fig. 10.21(b) to the dashed line, and we get a lower x̂p . A similar argument shown
in Fig. 10.21(c) shows that every production run after a production run must also
be clearing. These translations continue to satisfy the production requirements.
Hence, we note that every single production run is clearing.
Now we claim that we minimize x̂p if all x̌ip are equal, i.e., the order backlog
of p at the beginning of every production run is the same. If this was not so, as
shown by Fig. 10.21(d), then a simple translation of the trajectory from the contin-
uous line to the dashed line will result in a lower x̂p while still meeting production
j
requirements. This implies that µi = µki for all j, k. Henceforth, we shall drop the
subscript and Eq. (10.21) can be rewritten as

j
µi = zi x̌ i . (10.21)

These arguments were presented persuasively in [25].

(a) (b)

( c) (d)
Fig. 10.21 Arguments for clearing production runs.
316 10 The Supply Chain as a Dynamical System

Product i
x̂i
Fi

δ i µi
t
Fig. 10.22 Trajectory of level of backlog with definition of parameters.

Lemma 1 [A lower bound for the total storage]. For all policies the total weighted
storage S is bounded from below so that


p
S γi x̂LB,i ,
i=1

where γi > 0 are weights and

p  
γj dj (1−ρj 0.5 
i=1 δj δj δi di (1 − ρi )
x̂LB,i = .
1−ρ γi

Proof. From above, we note that a policy that minimizes the total storage will
ensure that every single production run is clearing and that the heights of the buffer
levels of product p at the beginning of every production run are the same, we now
refer to Fig. 10.22. Let li be the number of production runs of product i in the
minimum cycle time T ∗ . Define
T∗
Fi =
li
By material balance

Fi di = x̂i + µi di .

Since the machine is always working on a product or setting up for production we

have


p
T∗ = li (zi x̂i + δi ).
i=1

Now consider the following minimization problem:

p
min γi x̂i
Fi
i=1

subject to
 10.5 Assembly, Disassembly, and Packaging 317

0 = Fi di − x̂i (1 + zi di )

p
1
0= (zi x̂i + δi ) − 1. (10.22)
Fi
i=1

Using the method of Lagrange multipliers we set

 p   p  
∂
Fi di γi
z i di δi
+λ + − 1 = 0.
∂Fi 1 + z i di 1 + z i di Fi
i=1 i=1

Solving for Fi gives


λδi (1 + zi di )
Fi = .
di γi
Using the expression for Fi developed in Eq. (10.22) gives
p  0.5
γi di
√ i=1 δi δi (1+zi di )
λ= p zi d i
.
1 − i=1 1+z i di

The result follows.

We now develop an upper bound for the optimal steady-state feedback policy
using the wraparound or “round-robin” policy.

Definition 5 [Wraparound Policy]. A policy in which every product is produced once

and only once in every cycle and in the same order, is called a Wraparound Policy. The
steady-state storage is given by Swa defined in Theorem 1.

The wraparound policy is implementable and gives a way to calculate an upper

bound for storage requirements for an optimal feedback control policy.

Theorem 1 [The Upper Bound on Storage Using a Wraparound Policy for the
Optimal Policy]. The optimal policy S 0 satisfies the bound
S 0
Swa
where Swa denotes the order backlog using the wraparound policy given by


p
Swa = γi x̂U B,i
i=1

and
 p 
i=1 δi
x̂U B,i = di  p (1 − τi di ).
1 − i=1 τ di

Let T be the time required for one complete cycle, for example A, B, C, D, ....
Using the definitions introduced above we can assert
318 10 The Supply Chain as a Dynamical System

p
T = (µi + δi ),
i=1
xi = di (T − µi ),
µi = xi zi ,
τi
zi = .
1 − τ i di
It follows that
µi = di (T − µi )zi .
Hence
di z i
µi = T = T τ i di .
1 − di z i
Solving for the cycle time gives


p

p
T τ i di + δi = T ,
i=1 i=1

whence
p
δi
T = i=1
p ,
1− i=1 τi di

we therefore get
µi τ i di
xi = =T
zi zi
and the result follows.

Example 2 (The Holland Tunnel effect). In this example we examine the sensitiv-
ity of the lower bound on total storage to the load on the processing unit. Consider
a unit that works on three products with parameters as listed in Table 10.1. The
plot in Fig. 10.23 shows the sensitivity of the lower and upper bounds on total stor-
age with respect to load as τ1 varies from 0.0 to 0.91. We note that when ρ < 0.9
(corresponding to τ1 < 0.75) the minimum storage required is quite small. As
we approach the maximum capacity of the processing unit, the minimum storage
required begins to climb very rapidly. Higher loads imply higher processing unit
utilization but also mean that small changes in load represent large changes in the

Table 10.1 System parameters.

τ d δ γ

Product 1 0 < τ1 < 0.91 0.6 1.0 1

Product 2 0.9 0.3 1.0 1
Product 3 0.9 0.2 1.0 1
 10.5 Assembly, Disassembly, and Packaging 319

500
Storage Caoacity (Units)

400
upper bound
300 lower bound

200

100

0
0.4 0.5 0.6 0.7 0.8 0.9 1
System Load
Fig. 10.23 The Holland Tunnel effect.

minimum storage required. Low loads, on the other hand, imply low unit utiliza-
tion but the storage requirement becomes relatively insensitive to the load. Thus
Fig. 10.23 represents the tradeoff between ease of operation (corresponding to low
loads) and high process utilization. This phenomenon is known as the Holland
Tunnel effect and it leads to bottlenecks.

10.5.2
Feedback Scheduling

We now develop feedback control policies which achieve the production demands
while attempting to minimize the order queue. We will only consider clearing poli-
cies and use these to develop a class of feedback policies called the largest-backlog
policy or LBP. Briefly, the LBP clears a buffer and then starts to work on the product
which has the largest weighted backlog. The weighting factor γi > 0 can be used
to assign priority.

Definition 6 [Clearing Policies]. A clearing policy

1. Continues a production run until the storage tank is empty.
2. Switches to another product at the end of a production run.

It follows that if the processing unit began setting up for product k at t0 , then a
clearing policy terminates the production run k at time tl so that xk (tl ) = 0, where
(xk (t0 ) + δk dk )τk
tl = t0 + + δk .
1 − τ k dk
320 10 The Supply Chain as a Dynamical System

We have already seen that a clearing policy results in the lowering of storage re-
quirements compared to policies whose production runs are not all clearing. The
intuitive idea is that one must not switch between products (setup or cleanup) too
frequently.
All policies we discuss will be of the clear-a-fraction kind which are a special case
of the clearing policy.

Definition 7 [Clear-a-Fraction (CAF) Policies]. Fix 0 < < 1. Let t0 = 0. A clearing

policy is said to be a CAF policy if at time tn it chooses to set up for product type p∗ (tn )
such that


p
xp∗ (tn ) (tn )  xi (tn ).
i=1

Production is commenced at time tn +δp∗ (tn ) , and continued until the buffer level
xp∗ (tn ) (t) hits 0 at time t = tn + l given by
τp∗ (tn )
tn+l − tn = δp∗ (tn ) + xp∗ (tn ) (tn + δp∗ (tn ) .
1 − τP ∗ (tn ) dp∗ (tn )
In other words, a CAF policy has the following characteristics:
1. It will continue the production until no more material is left.
2. At this point the policy will choose another product for the
processing unit to work on. This clearly cannot be the
product just processed, but must be one that has a nonzero
amount of material in storage.

We now develop a CAF policy that attempts to keep the maximum (or peak)
buffer levels close to the lower bounds x̂i∗ . The arguments (equal production
lengths) used in the proof of Theorem 1 when combined with operability require-
ments result in the Wraparound Policy, which requires no less storage than that
predicted by the lower bound developed in Lemma 1. Hence a policy that will do
better than the wraparound policy will not be strictly symmetric. Rather, the peak
buffer levels x̌i (k), k = 1, 2, ..., of product i will vary with time. For a given set
of parameters, let the optimal sequence of peak buffer levels be x̌i∗ (k). Define the
normalized backlog of product i to be the quantity
xi (k) − x̄ ∗ (k)i
Bi (k) = γi ,
x̄ ∗ (k)i
where x̄(k)i  0 is an estimate of the peak level x̌i∗ (k) for the order backlog of
product i. Two factors come into play here
1. If a production run of i starts with a small backlog (Bi < 0),
this will result in a short production run and less product
will be made. This will require a longer run at a later time to
make up for the shortfall, and this may result in larger x̂i .
2. Similarly, if the backlog gets too large, then x̂i is large, the
long production run keeps the processing unit busy for a
 10.5 Assembly, Disassembly, and Packaging 321

long time, and results in inventory buildup in the storage

tanks for other products.

Using these ideas, the following policy which chooses to work on the product that
has the largest backlog becomes obvious. The definition of the backlog allows us to
define a strategy for feedback control called the Largest Backlog Policy.

Definition 8 [The Largest Backlog Policy (LBP)]. When the processing unit has just
cleared product j begin setup to produce product k = j so that
 
xi − x̂i∗
k = arg{max }, i = 1, ..., p
x̂i∗
where x̂i∗ has been previously defined.

Theorem 2 [Stability of the Largest Backlog Policy]. LBP can be implemented with
fixed and finite storage provided


p
τi di < 1.
i=1

Proof. From the definition of the LBP we have

(xi − x̂i∗ )
xp  x̂p∗ + x̂p∗ .
x̂i∗
Choose
x̂p∗
¯
.
x̂i∗
Hence

xp  ¯ (xi − x̂i∗ ) + x̂p∗ ,

¯
 p(xi − x̂i∗ ) + x̂p∗ ,
p
¯

p
 (xi − x̂i∗ ) + x̂p∗ ,
p
i=1

¯
p
¯
∗
p
 xi − x̂i ) + x̂p∗ .
p p
i=1 i=1

Choose
¯
∗
p
x̂i
x̂p∗ .
p
i=1
322 10 The Supply Chain as a Dynamical System

Hence
x̂ ∗
0 < ¯
p p i ∗
.
i=1 x̂i

Thus the LBP is a clear-a-fraction policy. CAF policies can be implemented with
fixed and finite storage [25]. The result follows.

In the LBP, the decision to choose the next product is made at the end of the
production run of the current product. The aim in designing the LBP is to try
and start production runs such that maximum buffer levels are close to the values
dictated by the lower bound. Since the largest amount of material in a buffer occurs
just after a setup, it might be advantageous to consider the effect of setup while
calculating the backlog.
The LBP gives a closed-loop system which can be implemented with fixed and
finite storage provided that the load condition

p
τ i di < 1
i=1

is satisfied. It is difficult to study the asymptotic properties of the feedback sys-

tem explicitly because it is nonlinear. To give some insight we shall discuss a few
simulation examples.

Example 3 (Performance of the Largest Backlog Policy and Wraparound Policies).

Consider a unit manufacturing three products with parameters as shown in Ta-
ble 10.3. The storage requirements for the LBP and the WA are shown in Table 10.4.
When the system is under the control of the LBP the production sequence that
evolves is

1232123212....

The wrap around sequence gives

123123123...

Table 10.2 System parameters.

τ d δ γ

Product 1 0.932 0.381 0.147 1

Product 2 2.8152 0.066 8.166 1
Product 3 0.055 6.150 0.024 1
 10.6 Flow Control and Load Balancing 323

Table 10.3 System parameters.

τ d δ γ

Product 1 0.01 0.6 25 1

Product 2 0.9 0.3 1.0 1
Product 3 0.9 0.2 1.0 1

Table 10.4 System parameters.

Lower bound Upper bound Simulation

LBP 41.2 48.6 47.03

WA 41.2 48.6 48.6

Mark-up

Sale price cs
Unit Value $

Internal Value cp

A c ti vi ty C o s t = V a l u e A d d e d
Added value in manufacture

Purchase price cr

Move Supplies Manufacture Store Move Products Mark-up

Fig. 10.24 In activity-based analysis the value of products is

increased by transportation and manufacture.

10.6
Flow Control and Load Balancing

Activity-based analysis [15] and lean manufacture [28] approach supply chain man-
agement from the point of view that profit P for a reporting period is maximized
when the activity cost is minimized (Fig. 10.24). The profit for an activity is calcu-
lated so that
 t
P (t) = (R − C)ds,
t−tAP

where tAP is the reporting period, R is the rate of income from sales, and C is the
rate of cost. The difference R − C is called rate of accounting earnings. The activity
costs include transportation, storage, manufacture (assembly), and purchase. By
differentiating the expression above, we get the differential balance
dP
= R − C, P (t − tAP ) = 0. (10.23)
dt
324 10 The Supply Chain as a Dynamical System

Manufacture Storage

M S
Out going orders Incoming orders

Routing
Receiving φr Shipping φs

Fig. 10.25 Representation of an activity, containing

manufacturing, storage, routing, shipping, and receiving nodes.

Factories, distribution centers, transportation, and retailers contribute to value cre-

ation as illustrated in Fig. 10.25.
In order to link supply chain operation to value creation we need to introduce the
idea of value more formally. Four different components are introduced as follows:
• transportation and movement of goods, services, and
information,
• manufacture, including assembly, packaging, quality control,
cost of processing of orders, and accounting.
• storage of materials and supplies, and
• receiving, shipping, sales, and purchasing.

We must also define the basic dynamic elements needed to describe the evolution
of the state of the system and how it is constrained by stability considerations.
1. The state: The vector of length 2p of current assets

 
I
Z=
O

comprising inventories and outstanding orders, represents

the state of the supply chain system.
This vector evolves according to the differential Eqs.
(10.2) and (10.4) so that

dZ
= φs − φr + p, (10.24)
dt
where p is the rate of manufacture. We model manufacture
so that

i Mi = 0,
 10.6 Flow Control and Load Balancing 325

where i is the processing rate for activity i and M is a

conservation vector which provides the recipe for
manufacture, assembly, and repackaging. For example, the
vector

M5 = (−2, −2, −5, 0, 4, 1, −3)T

assembles two units of items 1 and 2, 5 units of item 3 and 3

units of item 7 from the list of SKUs to produce 4 units of
item 5 and one unit of item 6.
2. Valuation: The positively homogeneous degree one function
A(Z, t) defines the current (intrinsic) value of the assets,
including purchased and manufactured goods and orders.
The valuation depends on time since market conditions and
product properties change. Homogeneity implies that the
value is proportional to the assets so that

A(λZ, t) = λA(Z, t).

3. Cost: All activity costs are nonnegative. This implies that

circular activity, like moving an item from point A to B and
back to A incurs a cost without creating value. This key
assumption allows us to relate value creation to cost and
show that the performance of the supply chain in fact is
optimized when it is operated as lean as possible. The
nonnegativity of cost also leads to the following important
result: [9].

Lemma 2. The value function A(Z, t) is concave in Z.

We now define the dual

A∗ (c) = max(A(Z) − cT Z). (10.25)

Z

The vector of per unit values c = (c1 , . . . , cN , cp+1 , . . . , c2p )T represents the value
of adding one unit of the corresponding asset, which may include supplies, prod-
ucts, and orders. Homogeneity of A(Z) then implies that


A = ZT c = ZiT ci . (10.26)
assets

By differentiating A(Z) we get

dA dZ dc
= ZT + ZT . (10.27)
dt dt dt
326 10 The Supply Chain as a Dynamical System

The last term is equal to zero since Z T dc = 0, which is referred to as the Gibbs–
Duhem equation in thermodynamics. The inventory balance (10.24) is now com-
bined with Eq. (10.27) to give
dA
= crT r − cpT s + pA . (10.28)
dt
The variable
pA = (c − cr )T r − (c − cp )T s + cT p (10.29)
   
transportation production

represents the activity cost, cr represent cost per unit purchased, cp is the per unit
value before mark-up while cs is the sales price per unit. The cost of orders include
marketing, research, and communication, but not the mark-up.
In a system with assets Zi , i = 1, . . . , p, p + 1, = . . . , 2p, we have the following
interpretation of mark-up:

> 0, sell with a profit
cs,i − cp,i = = 0, sell at cost (no mark-up)

< 0, sell with a loss.
We can view the sales price cs as a control variable. The variables of internal prices,
cp are intrinsic variables which measures how well the supply is run. We want the
internal price of an item to be as low as possible since these represent the cost
whereas sales variables should be as high as possible.
The activity cost in Eq. (10.23) can now be defined so that
C = pA + crT r.
Combining Eqs. (10.23) and (10.28) gives
 
dP dA
=R− + cpT s . (10.30)
dt dt
By using the definition for the income from sales we get
dP
dA
= (cs,i − cp,i )si − . (10.31)
dt dt
sales

In this expression we let cs,i denote the ith component of the vector cs and cp,i
denote the ith component of the vector cp . We note that cs,i denotes the sales price
whereas cp,i denotes the price “at cost” for item i. The cost of outgoing orders is
zero in a COD/FOB purchase system.
In the case cs − cp = 0 there is no mark-up. This is often the case for internal
customers and the cost cp is then referred to as a “fair transfer price.”
Expressions (10.29) and (10.31) highlight the main issues in supply chain man-
agement.1)

1) There can be a considerable phase shift between the movement of goods and the associated finan-
cial transaction. Ignoring this phase shift is referred to as accrual.
 10.6 Flow Control and Load Balancing 327

1. The profit increases at a faster rate when the mark-up

cs − cp > 0 is large. Larger mark-up can be achieved by
raising the per unit price of the item sold. But higher prices
also tend to give reduced sales. The focus has therefore been
placed on reducing the manufacturing cost.
2. The sales volume should be high.
3. The profit can be increased by reducing current inventory
and fixed assets since we have dA/dt < 0.
4. The transportation and production costs as defined in
Eq. (10.29) should be minimized.

Example 4 (Minimization of activity cost by selecting cheapest suppliers). Con-

sider an activity with one product, as shown in Fig. 10.26. There is a single cus-
tomer and three different suppliers. Each supplier has a shipping rate constraint
so that

0
fr1
200,
0
fr2
200,
0
fr3
200,
and a fixed unit price (low, medium, high). The customer demand fluctuates in the
range
300
φso
500.
The inventory balance is
dI
= φr − φs , (10.32)
dt
where


3
φr = fri .
i=1

We assume there are no delays so that the order balance is superfluous. There is
no assembly or repackaging. However, there is a penalty for having an inventory
that is too high leading to higher capital cost. We model the sales price so that
cs = c∗ − a(I − I ∗ ) + C((I − I ∗ )),
     
storage cost mark-up

cheap

me d i u m demand
RECEIVING SHIPPING
e x p e ns i v e

Fig. 10.26 One node with three suppliers and a varying demand.
328 10 The Supply Chain as a Dynamical System

where c∗ is the base cost which is independent of how much is stored, and a  0
is a constant which reflects the change in per unit cost of holding inventory. The
negative sign makes the per unit value in storage decreases as the inventory level
increases (i.e., the storage penalty per unit in storage increases). Finally, C(·) is a
strictly input passive operator, e.g., PID control, which controls the inventory level
by adapting the mark-up to the inventory level. We use a small mark-up if the
inventory is too high, indicating that we do a sale, whereas we use a higher mark-
up if inventory is low, indicating that the item is popular and is almost sold out. An
example of the price rises and falls is shown graphically in Fig. 10.29 for a linear
storage cost a and no control action C. In this case c∗ is equal to 7 for an inventory
setpoint of 10. We use a low-bandwidth PI type control to adjust the sales price.
The controller is defined so that

1 t
C(e) = Kp e + edt,
τI 0
where

e = I − I∗

is the tracking error. The inventory I ∗ can be set using the theory developed in the
previous section.
The flow amongst suppliers and customers is assumed to be driven by price
differential so that

fj,j −1 = Ki (ci − cr ),

where

c1 = 2,
c2 = 5,
c3 = 10.

and K = 100.
The demand varies randomly between 300 and 500 units using Simulink’s uni-
form random number generator, with a change in random demand every 5 time
units as shown in Fig. 10.27(a). At t = 50 the desired inventory changes from 500
to 600. This change is implemented as a setpoint change in the control law and
may be due to a desire of carrying higher safety stock. The controller implements
the change by temporarily adapting the purchase price so that the inventory, shown
in Fig. 10.27(b), tracks the setpoint well. The order rates are shown in Fig. 10.28(a)
and the resulting price is shown in Fig. 10.28(b). The large spike in the price is
solely due to the immediate step change.
The example shows that it is possible to control inventory by adapting the sale
price to the market conditions. In practice such changes are implemented as sale
campaigns and are mainly used to reduce stock due to seasonal variations in de-
mand. The base prices in general follow a smoother trend and correlate more
closely to the consumer price index.
 10.6 Flow Control and Load Balancing 329

a) b)
500
Demand vs Time Inventory Level vs Time
620

600
450
580

Inventory Level
Demand

560
400
540

350 520

500

300 480
0 20 40 60 80 100 0 20 40 60 80 100
Time Time

Fig. 10.27 Demand (left, a) and inventory level (right, b) for one node.

a) b)
200
Orders Made vs Time Controlled Price vs Time
120
Orders Made from Suppliers

Lo w 100
150 Medium
Price at node

High 80

100 60

40
50
20

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Time
Time

Fig. 10.28 Orders to suppliers (left, a) and price at node (right, b).

12
Sales Price vs Inventory
Price curve
Setpo int
10
Sales Price

2
0 5 10 15 20
Inventory

Fig. 10.29 Example curve of how price per unit changes as the
inventory differs from the setpoint.

We now want to model an interconnected network of activities. The added value

of a path consisting of several activities of the type, shown in Fig. 10.24, is given by
the formula


X= Xi , (10.33)
Segments
330 10 The Supply Chain as a Dynamical System

where Xi,j = ci − cj represents the added value of each subactivity. This number
may be positive, zero, or negative and it does not depend on the path taken since
the function A(v, t) is unique. For a cyclical activity we have


0= Xi . (10.34)
Loop

This expression conveys the idea that there is no value added in a cyclical activity.
However, there is cost associated with every activity. Cyclical activities therefore add
cost but no value.
A supply chain is an integrated network of activities which transports and trans-
forms assets so that its intrinsic value [3] is maximized. An asset may be a tangible
product like a gallon of oil, a piece needed in an assembly line or an intangible
item like an order, information or intellectual property. An increase in value may
be the result of an asset been transported to a location closer to the customer, a
transformation (e.g., chemical reaction or assembly) or because time progresses
and market parameters change. The objective of this section is to describe the con-
servation laws that constrain the dynamic behavior of the supply chain.
Consider a supply chain with p distinct SKUs and n basic activities. The state of
the supply chain is then given by the vector

Z T = (Z1T , . . . , ZnT ).

The topology of the supply chain is represented by the graph G = {H, A}. Here
H represents a set of edges, along which assets are allowed to flow and A repre-
sents the activities where assets are stored, transformed, shipped, or routed. It is
useful to introduce a little more structure and distinguish among the four differ-
ent classes of activities. These include transportation, manufacture, storage, terminals
(shipping/receiving), and routing. This additional structure allows us to define the
value added network (VAN):2)

G = {H, M, S , T , R}
  
A

where elements hi ∈ H, i = 1, . . . , nh denote the transportation of assets, ele-

ments mi ∈ M, i = 1, . . . , np represent the manufacturing with assembly or dis-
assembly of chemical constituents or parts into precursors and products. Elements
si ∈ S , i = 1, . . . , ns denote the storage facilities, elements ti ∈ T , i = 1, . . . , nt de-
note terminals for receiving and shipping and lastly, elements ri ∈ R, i = 1, . . . , nr
represent points where material, energy, money, or data can be routed in different
directions in accordance with the conservation laws for inventories (10.2) and or-
ders (10.4). For a system with n activities we get n balance equations
dZi

= fr,i − fs,i + pi . (10.35)
dt
terminals

2) The notation and order has been chosen in memory of our beloved hamster TicTac.
 10.6 Flow Control and Load Balancing 331

A supply chain network of this type is called a VAN if there also is a value function
A(Z, t) which represents the value of assets at time t satisfying the conditions
introduced above.
We now have the following extremely important result for transportation, stor-
age, and production in an SCS.

Theorem 3. Consider a supply chain network with linear network operators T and L.
We have

d Z̄ T

c̄ = f¯T X̄
storage
dt
transportation




+ p̄T c̄ + − c̄rT f¯r + c̄sT f¯s ,
production receiving shipping

where the overbar denotes a deviation from a stationary reference value which we denote
with a * so that

Z̄ = Z − Z ∗ ,
c̄ = c − c∗ ,
f¯ = f − f ∗ ,
X̄ = X − X∗ ,
p̄ = p − p∗ .

Proof. See [17]

This result expresses the interesting fact that the space of inventories and costs
are orthogonal.
The problem we want to solve is how to choose control policies to stabilize the
dynamics and maximize the profit of the supply chain without using centralized
coordination of all decisions. We can formulate this problem so that


n
dP
max
fi ,pi dt
i=1

subject to Eq. (10.35).

In order to develop logistic systems and manufacturing schedules that balance
system load we need to evaluate the activity costs and their sensitivity with respect
to changes in the activity rate. The following definition is useful.

Definition 9. A transportation schedule is stable if f¯T X̄  0 and a manufacturing

process is stable if p̄T c̄  0.
332 10 The Supply Chain as a Dynamical System
Table 10.5 Market parameters for a supply chain example.

Terminal a b k

R1 0.02 2 0.075
R2 0.04 3 0.043
R3 0.06 2 0.060
R4 0.10 1 0.038
R5 0.03 6 0.043
R6 0.03 6 0.043
S1 −0.01 20 0.060
S2 −0.01 20 0.060
S3 −0.01 20 0.060

Stable transportation and manufacture ensures that the per item cost of a given
activity decreases (or at least does not increase) with increasing traffic. Stability is
related to passivity so that the pricing and control policies

f = f ∗ + Cf (X − X∗ )
c = c∗ + Cc (I − I ∗ )
are stable if the operators Ci (·), i = f, c are passive. Examples of stable policies
include the barrier function, which describes capacity constraints, gradient direc-
tions that result from optimization of convex cost functions and more generally
any cost which is monotonic in the sense that higher added value gives incentive
to larger shipments as well as many policies for mark-up and inventory control.
The pricing policies are designed to solve load balancing and resource allocation
problems since they determine how cost varies with respect to production volume
throughout the supply chain. The policies developed above will automatically give
preference to larger retailers and suppliers that can deliver FOB/COD at the lowest
cost. Without such costs load balancing is not a well-posed problem.
We now show that a stable, decentralized policy solves the optimal control prob-
lem. We proceed in two steps. We first show that the decentralized control system
is stable and converges to a unique solution provided the boundary conditions are
fixed. We then show that the stationary point is optimal.

Theorem 4. Consider a VAN with fixed boundary costs and stable feedback controls.
The inventories are then stable and converge to stationary values.

Proof. Details given in [17].

Theorem 5. Consider a VAN with fixed boundary costs and positive feedback controls.
The total activity cost is minimized.

Proof. Details given in [17].

 10.6 Flow Control and Load Balancing 333

R5
R1
4 S1
1 5
R2 6 S2
2 9
R3
7
3 10
R4 S3
8

R6

Fig. 10.30 Supply chain network with 10 nodes and 9 terminals.

Example 5 (Supply chain network). The distribution network shown in Fig. 10.30
consists of 10 activities, 6 receiving terminals, and 3 shipping terminals. There are
3 echelons and 10 internal transfers. We use an exponential function to define the
market price at the terminals

cT = b + aekfT ,

where a, b, k are constants. The parameters are chosen so that b gives a base cost.
The prices rise or fall at the supplier and consumer terminals, respectively, for or-
ders above a certain point related to the “activation energy” k. The parameter a is
positive or negative depending on whether the terminal is a supplier or consumer.
We also consider variable disturbances in the demand from the consumers. These
are added to the determined demand at the current price using the uniform ran-
dom number generator in Simulink, producing an additional demand between 0
and 4.
The internal transportation cost is based on the linear expression

fij = Kij (cj − ci − cT ), Kij = 100,

where cT is a constant transportation cost per unit of material. In this example,

there are no hard flow constraints as in the one node example. The only constraints
to the flow are due to changes in prices upstream and downstream with changes
in throughput levels.
The high gain allows us to approximate a switching mechanism so that goods are
bought from the suppliers with the lowest cost. Figure 10.31 shows the inventory
level in activity 7, while Fig. 10.32 shows how the controller adapts the price of
activity 7 to control the inventory.
Several of the flows within the network for the current pricing rules at the ter-
minals and transportation costs become zero, indicating that those routes are inef-
ficient. In fact, currently the coupled network splits into two disconnected chains,
one between receiving terminal R3 and shipping terminal S1 (through nodes 2,
6, and 5), and the other between receiving terminal R4 and shipping terminals S2
334 10 The Supply Chain as a Dynamical System

500.5
Inventory vs Time in Node 7
500.4

Inventory 500.3

500.2

500.1

500

499.9

499.8
0 20 40 60 80 100
Time

Fig. 10.31 Measured inventory level at the 7th node.

4.4
Price/unit vs Time in Node 7

4.2
Price/unit

3.8

3.6

3.4
0 20 40 Time 60 80 100

Fig. 10.32 Control action on the price of material in the 7th node of the network.

and S3, (through nodes 3, 7, 8, 9, and 10). The results at the final time are given in
Table 10.6.

10.7
Summary and Conclusions

In this chapter, we have presented a brief overview of the supply chain management
problem from a dynamical systems perspective. We use a combination of distrib-
uted feedback and feedforward control to address important issues in supply chain
control. We show that the problem of demand amplification can be addressed by
using frequency domain analysis. A modified version of the MIT Beer Game was
used to illustrate the bullwhip effect along with possible solutions.
 References 335

Table 10.6 Flows at the final time for supply chain system.

Flow name Final flow rate Flow name Final flow rate

fr1,1 0 f3,8 23.67

fr2,1 0 fr6,8 0
fr3,2 23.67 f4,5 0
fr4,3 47.34 f6,5 23.67
f1,4 0 f7,9 23.67
fr5,4 0 f7,10 0
f2,6 23.67 f8,10 23.67
f2,7 0 f5,s1 22.38
f3,7 23.67 f9,s2 22.37
f10,s3 22.38

One of the developments presented here is the ability to use filtering theory and
approaches to effectively develop extra storage methods so that unexpected changes
in demand and supply can be appropriately managed even in the presence of a fluc-
tuating market. We also presented the concept of a value added network. The fact
that many echelons of the supply chain such as assembly, storage, routing, process-
ing, and transportation units are all incorporated into the VAN makes this approach
a powerful one. Here, activities interact and complexity increases as the number of
business activities and links increase, resulting in a powerful building-block ap-
proach to modeling. In order to establish some theoretical results we used results
motivated by electrical circuit theory. Stability and optimality of the decentralized
decision making process have been established.
Distributed decision making in supply chain systems arises naturally in several
ways. For example, it is not unusual that the systems under consideration are dis-
tributed since processing segments, business units and enterprises are integrated
into a complex, diverse, and highly dynamic global market. Future work in this
area should focus on how to integrate distributed decision making with forecast-
ing, planning, and optimization.

References

1 Badell, M., Romero, J., Huertas, R., http://www.berkshirehathaway.com/

Puigjaner, L., Planning, scheduling letters.html, 2002.
and budgeting value-added chains. 4 Christopher, M., Logistics and Sup-
Comput. Chem. Eng. 28 (2004), pp. ply Chain Management – Strategies for
45–61. Reducing Cost and Improving Service,
2 Beamon, B. M., Supply chain design Second edition, Prentice-Hall, Lon-
and analysis: Models and methods. don, UK, 1998.
Int. J. Prod. Econ. 55 (1998), pp. 281– 5 de Souza, R., Zice, S., Chaoyang, L.,
294. Supply chain dynamics and optimiza-
3 Buffett, W., Berkshire hath- tion. Integ. Manufac. Syst. 11 (2000),
away shareholder letters. pp. 348–364.
336 10 The Supply Chain as a Dynamical System

6 Disney, S. M., Farasyn, I., Lam- 17 Jillson, K. R., Ydstie, B. E., Process
brecht, M., Towill, D. R., Van networks with decentralized inventory
de Velde, W., Creating win-win sce- and flow control. J. Process Control 17
narios from the bullwhip problem-by (2007), pp. 399–413.
design not accident. In EUROMA 18 Lee, H. L., Padmanabhan, V.,
POMS Conference, Como Lake, Italy, Whang, S., Information distortion
16–18 June, 2003. in a supply chain: The bullwhip effect.
7 Disney, S. M., Grubbstrom, R. W., Manage. Sci. 43 (1997), pp. 546–558.
The economic consequences of a pro- 19 Lin, P., Wong, D. S., Jang, S., Shieh,
duction and inventory control policy. S., Chu, J., Controller design and re-
In 17th International Conference on duction of bullwhip for a model sup-
Production Research, Virginia, USA, ply chain system using z-transform
3–7 August, 2003. analysis. J. Process Control 14 (2004),
8 Forrester, J. W., Industrial Dynamics. pp. 487–499.
MIT Press, Cambridge MA, 1961. 20 Maravelias, C. T., Grossmann, I.
9 Garcia-Osorio, V., Ruszkowski, M., E., Simultaneous planning for new
Ydstie, B. E., Passivity based control product development and batch man-
of transport reaction systems. AICHE ufacturing facilities. Ind. Eng. Chem.
J. 51 (2005), pp. 3147–3166. Res. 40 (2001), pp. 6147–6164.
10 Geary, S., Disney, S. M., Towill, D. 21 Mele, F.D., A Multi-agent Approach for
R., On bullwhip in supply chains- Modeling and Analysis of Supply Chain
historical review, present practice and Networks. PhD thesis, Universitat Po-
expected future impact. Int. J. Prod. litecnica de Catalunya, 2005.
Econ. 101 (2006), pp. 2–18.
22 Papageorgiou, L. G., Rotstein, G.
11 Goel, V., Grossmann, I. E., A sto- E., Shah, N., Strategic supply chain
chastic programming approach to optimization for the pharmaceutical
planning of offshore gas field develop- industries. Ind. Eng. Chem. Res. 40
ments under uncertainty in reserves. (2001), pp. 275–286.
Comput. Chem. Eng. 28 (2004), pp.
23 Perea-Lopez, E., Grossmann, I. E.,
1409–1429.
Ydstie, B. E., Tahmassebi, T., Dy-
12 Goel, V., Grossmann, I. E., A class namic modeling and decentralized
of stochastic programs with decision control of supply chains. Ind. Eng.
dependent uncertainty. J. Math. Prog. Chem. Res. 40 (2001), pp. 3369–3383.
108 (2006), pp. 355–394.
24 Perea-Lopez, E., Ydstie, B. E., Gross-
13 Guillen, G., Contribution to Supply mann, I. E., A model predictive con-
Chain Management in the Batch Chem- trol strategy for supply chain opti-
ical Processing Industry. PhD thesis, mization. Comput. Chem. Eng. 27
Universitat Politecnica de Catalunya, (2003), pp. 1201–1218.
2005.
25 Perkins, J. R., Kumar, P. R.,
14 Hayes, R. H., Wheelwright, S. C., Stable, distributed, real time
Restoring Our Competitive Edge: Com- scheduling of flexible manufactur-
peting Through Manufacturing. Wiley, ing/assembly/disassembly systems.
New York, NY, 1984. IEEE Trans. Autom. Control 34 (1989),
15 Helfert, E. A., Financial Analysis 1143–1149.
Tools and Techniques. McGraw-Hill, 26 Riddalls, C. E., Bennett, S., The op-
New York, NY, 2001. timal control of batched production
16 Hugo, A., Pistikopoulos, E. N., En- and its effect on demand amplifica-
vironmentally conscious long-range tion. Int. J. Prod. Econ. 72 (2001), pp.
planning and design of supply chain 159–168.
networks. J. Cleaner Prod. 13 (2005),
pp. 1471–1491.
 References 337

27 Sasieni, M., Yaspan, A., Friedman, 36 Taylor, D., Brunt, D., Manufacturing
L., Operations Research: Methods and Operations and Supply Chain Manage-
Problems. Wiley, New York, NY, 1959. ment: The LEAN Approach. Thomson
28 Schei, A., Tuset, J. K., Tveit, H., Pro- Learning, Great Britain, 2001.
duction of High Silicon Alloys. Tapir 37 Tsiakis, P., Shah, N., Pantelides, C.
Verlag, Trondheim, Norway, 1998. C., Design of multi-echelon supply
29 Schwartz, J. D., Rivera, D. E., chain networks under demand uncer-
Control-relevant demand modeling tainty. Ind. Eng. Chem. Res. 40 (2001),
for supply chain management. In pp. 3585–3604.
Proceedings of the 14th IFAC Sympo- 38 Wang, W., Rivera, D. E., Kempf, K.
sium on System Identification (SYSID), G., A novel model predictive control
Newcastle Australia, 29–31 March, algorithm for supply chain manage-
2006. ment in semiconductor manufactur-
30 Shapiro, J. F., Challenges of strategic ing. In Proceedings of the 2005 Ameri-
supply chain planning and modeling. can Control Conference, Portland OR,
Comput. Chem. Eng. 28 (2004), pp. 8–10 June, 2005.
855–861. 39 Warburton, R. D. H., Disney, S. M.,
31 Simon, H. A., Administrative Behavior: Variance amplification: Time equiva-
A Study of Decision Making Processes lence of discrete and continuous time
in Administrative Organizations, 4th analysis. In Proceedings of the 18th In-
edition, Free Press, New York, NY, ternational Conference on Production
1997. Research, Salermo, Italy, 31 July–4
32 Singhvi, A., Madhavan, K. P., August, 2005.
Shenoy, U. V., Pinch analysis for ag- 40 Ydstie, B. E., New vistas for process
gregate production planning in supply control: Integrating physics and com-
chains. Comput. Chem. Eng. 28 (2004), munication networks. AIChE J. 48
pp. 993–999. (2002), pp. 422–426.
33 Srinivasan, V., Flow Control in Non- 41 Ydstie, B. E., Distributed decision
Continuous Chemical Plant. PhD the- making in complex organizations:
sis, University of Massachusetts, 1992. The adaptive enterprise. Comput.
34 Stadtler, H., Kilger, C., Supply Chem. Eng. 29 (2004), pp. 11–27.
Chain Management and Advanced
42 Yu, H., Cassandras, C. G., Perturba-
Planning. Springer, Berlin, Germany,
tion analysis for production control
2000.
and optimization of manufacturing
35 Sterman, J. D., Business Dynamics: systems. Automatica 40 (2004), pp.
Systems Thinking and Modeling for a 945–956.
Complex World. Irwin McGraw-Hill,
New York, NY, 2000.
 339

Index

a – simulation 12
active ingredient (AI) 2, 10, 17 – stochastic programming 6, 12
– consumption 17 – strategic 289
– content 2 – two-stage 12ff.
– sales 18 arc
activity 324f., 332f. – constraint 100
– basic 330 – directed 100
– circular 325 – undirected 100
– cyclical 330 asset 8
– processing rate for 325 automation 309f.
Adaptive Action Set (AAS) 67 Away From Home (AFH) 238
adaptive enterprise 292
aggregated time periods (ATP) 7 b
aging backorders 301, 305
– aspect 253 batch processes
– penalty cost factor 253 – multiproduct 173
– profiles 250, 263 – multipurpose 173
allocation transfer 11 Bellman 64f.
amplitude ratio (AR) 306f. – equation 65ff.
analytical hierarchy process (AHP) 21 – iteration 65
approach benchmark problems 46f., 49ff., 65
– discrete-time 173 – Solomon 47
– distribution-based 13f. Benders cut 19
– dynamic programming 12, 32 Bill Of Material (BOM) 6, 247, 273
– evolutionary random policy search blending
(ERPS) 67f. – decisions 220
– heuristic 29, 31, 46 – operations 220
– holistic 8 – optimization 220
– model predictive control (MPC) 289 – process 220
– multiobjective robust optimization – product 221
228 blends 220
– multistage stochastic programming bounds
12 – lower 314, 316, 318, 323
– optimization-based 12, 31 – upper 314, 318, 323
– parametricMILP (pMILP) 219, 221, Branch and Bound (B&B) 32, 196, 220, 224,
224f., 234 231
– real time approximate dynamic pro- Branch and Cut (B&C) 29, 32f., 45
gramming (RTADP) 61, 66ff. Branch and Price (B&P) 32f.
– scenario-based 14f. bullwhip effect 290f., 308

Supply Chain Optimization. Part II. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31906-0
340 Index

business – resource 181

– activities 4 – scheduling 275
– target 130 – sequencing 199f., 204, 206, 221
c – shipping rate 327
capacitated plant location problem (CPLP) – space 84
8 – tightening 207
capacity – time-based sequencing 43
– expansion 6, 14, 19 – time window 32ff.
– investment plans 8 – timing 176, 179
– limitations 8 – transportation 74
– under-utilize consumption
– vehicle 6, 46 – per capita 238
cargo contraction mapping 66
– initial 43 control
– minimum 43 – adaptive 296
changeovers 209, 244, 252ff. – back stepping nonlinear 296
– requirements 177 – centralized 291
– sequence-dependent 177, 181, 257, – decentralized 291
274 – feedback 285, 300, 313, 334
– time 177, 199 – flow 323
channel – hybrid 313
– delivery 5 – law 306
clinical trials 15ff. – nonlinear programming based finite
company horizon 296
– biotechnical 4 – objective 312
– centric 269 – Order-up-to (OUT) 290
competition 238 – PID 296
constraint-elimination rules 43f., 50 – policy 290f., 297, 313
constraint-violation – strictly passive 295f.
– routing-time 40 – system 304
– time window 40, 43 – theory 306
constraints converting line 243
– allocation 197, 202f., 206, 221 – process 243
– arrival-time defining 39 corporations 2
– assignment 36, 43 cost
– backlog 75 – activity 325ff.
– big-M 178f., 183, 188, 191 – allocation 7, 10
– capacity 33, 40, 43, 197, 203, 272 – backlog 272
– demand 198, 204, 221 – changeover 244, 258
– distance-based sequencing 38f., 43 – coefficient 207
– duration 199, 204 – cold storage 143
– extra tightening 201 – factor 7
– flow conservation 272 – handling 11
– hard 74, 246, 256 – holding 271
– imported material 74 – impelmentation 240
– input flow 74 – inventory 3, 11, 75, 270ff.
– inventory 195, 305 – lost sales 3
– load-based sequencing 41 – minimization of 177
– mass balance 221 – net 75
– material balance 200, 202f., 205, – products 7
221 – operating 143f., 153
– nonanticipativity 21f., 39 – processing 143
– product delivery 138 – production 7, 10f., 139, 271f.
 Index 341

– purchase 140 – long-term strategy 6

– raw material 11, 139, 153 delivery
– routing 37f., 43 – just-in-time 238
– setup 272 – locations 30
– storage 75 – problems 30
– tax 11 demand 29
– total 3 – forecast 11, 14ff.
– transportation 1f., 7, 10, 140, 153 – high 5
– working capital 3 – levels 5
crude 91 – low 5
– arrival data 92 – multiperiod 10
– destillation units (CDU) 92f., 97ff. – pattern 177
– operationg schedule 92 depot 35f.
curse of dimensionality (COD) 61, 63, 66ff. design
customers – problem 19
– active ingredients (AI) 2 – supply-chain 11f.
– areas 6ff. destination 30
– cluster 46 dial-a-ride problem (DARP) 31ff.
– database 5 disinvestment 6
– demands 292 disruption 97f.
– final 1, 5 – management system 94
– first-tier 269 distribution 1
– high-volume 5 – centers (DCs) 2, 7, 293, 303f., 306f.
– inconvenience 31, 34, 43 – channels 7
– request 29, 31, 33, 38f. – decisions 6
– Third Party (TPCs) 126, 128f., 131ff. – facilities 8
– flexibility 7
d – network 8
decisions drugs 3
– dynamic 61 – discovery 4
– first stage 15f., 25 – substitute 3
– here-and-now 12, 20, 22 dynamic programming (DP) 61f., 64ff.
– long-term 269 – approximate (ADP) 62f., 66, 69
– sales 16 – framework 6, 16
– scheduling 16, 21 – stochastic 67
– second stage 16
– short-term 269 e
– strategic planning 288 e-business 296
– support system 90 echelons 8
– variables 75f. economic-lot-sizing problem 20
– wait-and-see 12, 20, 22 end-product 240f., 243
decomposition – quality 263
– algorithms 18f. energy 289
– bi-level 8, 19 energy prices 238
– Dantzig-Wolfe 33 enterprise systems (ES) 240
– Lagrangean 20f. enterprise-wide optimization (EWO) 2
– method 22, 24, 221 environmental
– spatial 20f. – impact 238
– strategies 211 environmentally
– temporal 20ff. – friendly 241
De-Inked (DI) 244 Euclidean distance 46, 81
– pulp 262f. evolutionary random policy search (ERPS)
delineation 67f.
342 Index

exclusivity period 3 hierachical algorithms 18ff.

– of new products 3 Holand Tunnel effect 318f.
hybrid
f – methods 20
facility – systems 309
– Cold Storage (CSF) 122ff.
– multiproduct 241 i
farms industry
– production 139, 163 – agrochemicals 1
– Third Party (TPF) 126, 132, 134ff. – chemical 89, 91, 244, 269, 273
fiber yield 242 – food 124, 244
flow control 285f. – forest-based 237f.
flow-shop type 241 – paper 241
formulation 2 – pip fruit 121f., 125, 278
– sites 2 – manufacturing 239
– stage 2 – pharmaceutical 1, 3f., 8f., 244, 278,
fruit 288
– balance 133ff. – process 1, 3, 6
– packed 133f., 137, 144, 147f. – R&D-based 3f.
– processing 132, 136f. – sheet-metal 278
– purchase 139 – tissue paper 237f.
– reception 132 instance
– waste 132, 134ff. – negotiation 121, 123f., 126, 129,
153f.
g – operation 121, 123f.
Gantt chart 50f., 56, 186, 188, 211ff. – time 178
– continuous-time 253 Integrated Pollution Prevention and Control
general and administrative (G&A) expendi- (IPPC) 238
ture 4 intellectual property 2
generic intermediates 6, 91, 188, 198, 206, 242, 244,
genetic algorithms (GA) 16, 21 252ff.
– manufacturers 3 inventory
– terms 1 – balances 22, 293, 305, 327
geographic area – carryover 6
– secondary 10f. – control 285f., 303, 312, 334
Gibbs-Duhem equation 326 – level 3, 301f., 308, 329
global – maximum 304, 307
– competition 238 – minimum 304, 307
– continuous optimization 220 – profiles 10
– discrete optimization 220, 251 – steady-state 305
globalization 89 – work-in-progress 238
good chains 123 investment 4
graph – budget 7
– composite-operations (CO) 100ff. – contemplating 6
– rectifications- 104 – growth 4
– scheduled-operations (Scope-graph) – return 13
100, 103ff. IT systems
Graphical User Interface (GUI) 246ff. – analytical 273
– transactional 273
h
handicapped person transportation problem k
(HTTP) 31 Kalman filter 289
harvest period (HP) 131
 Index 343

l – decision processes (MDP) 13, 61,

Large Neighborhood Search (NLS) 32 63f.
Latin hypercubes (LH) 67 – process 76
lead time 5 – property 65
load balancing 323 mark-up 326ff.
location 4 mass
– delivery 6, 34 – balance 14, 175
– distribution sites 11 – unit 1
– pickup 6 master production schedule (MPS) 248,
– primary 10 273
– secondary 10 material
lot-sizing 273f. – balance 70f., 73, 197, 199, 206
– flow 10, 238, 291
m – functional 2
maintenance – third party 137
– policy 279 mathematical programming
– stops 250, 266 – based production planning 239
makespan 177ff. – based production scheduling 239
make-to-order (MTO) mergers and acquisitions (M&A) 4
– scheme 5 method
make-to-stock (MTS) – e-constraint 223
– policy 5 – full space 276f.
manufacture – hierachical 276f.
– lean 323 – iterative 276f.
– primary 1 – Lagrange multipliers 317
– secondary 1 – multiobjective approaches-weighting
manufacturers 4 223
– active ingredient (AI) 9 mixing 207
– tissue paper 237f. models
manufacturing 2 – campaign-based algorithmic 210
– contracts 4 – continuous-time 173ff., 252
– execution system (MES) 248 – deterministic 11f., 289
– integrated global 6 – differential equation 290
– network redesign 4 – discrete-time 173f., 252f.
– plants 2, 6 – Eco-Indicator 99 288
– primary 10 – economic-biophysical 125, 289
– process 5 – global event-based continuous-time
– resources 6f. 174f., 177, 181, 183, 188, 191
– sites 5 – Ierapetritou and Floudas (I&F) 178,
– stages 2 186, 188ff.
market – integrated 21
– area 9, 11 – intertemporal linear programming
– business-to-business 238 124
– energy 237 – large mixed integer linear program-
– local 122 ming 1ff.
– overseas 122, 128 – lot-sizing 274
– parameters 330, 332 – low-order 289
– pulp 261f., 264 – maintenance policy 279
– regional 2, 122 – Maravelias and Grossmann (M&G)
– share 89, 238 186, 188ff.
Markov – Markov 80
– chains 61, 64ff. – mass balance 124
344 Index

– mathematical programming 130, multiperiod optimization 8

154 multiplant 8
– mid-term planning 13 multiple time grids 252
– mixed integer nonlinear program- multiproduct assets 2
ming (MILP) 14ff. multipurpose 8
– mixed-time 252 multisite systems 29
– multiobjective robust optimization multistage
223 – planning 61
– multiperiod blending 220 – stochastic problem 61, 86
– multiperiod mixed integer nonlinear – stochastic programming 16
programming (MINLP) 92 myoptic reward 64
– multiperiod planning 6, 8, 14, 21,
125 n
– operation 19 net present value (NPV) 11, 15, 17, 19, 24f.
– optimization 5 – expected (ENPV) 289
– parametrc mixed integer linear pro- network
gramming (pMILP) 219, 224, 234 – distribution 9
– plant hardware 97 – global 2
– predictive control 124 – multiechelon distribution 291
– probabilistic 76 – multiproduct 291
– robust optimization 219f., 232 – neural 7, 96
– robustness 223 – formulation 2
– scenario-based 8 – packaging site 2
– simulation 5, 289 – oil flow 229f.
– single period planning 13 – resource-task (RTN) 175f., 179,
– single site planning 13 195f., 220
– slot-based continuous-time 174f., – route- 35
177, 183, 188, 196 – state-task (STN) 175f., 185f., 188f.,
– statistics 193f. 191, 207
– steady state 6, 14 nodes 6, 17, 328f.
– strategic planning 8 – flow 100
– stochastic 11, 15, 22ff. – incompatible 44
– stochastic fluid (SFM) 291f. – pickup 30, 34ff.
– Sundaramoorthy and Karimi (S&K) – pure delivery 31, 33ff.
186, 188ff. – state 100
– surrogate 276 – total-flow-in 103f.
– tactical 125, 129, 154 – total-flow-out 100, 102
– two-index mathematical 56 – transfer 100
– unit-specific event-based continuous- normal boundary intersection (NBI) 219,
time 174f., 177f., 188, 191, 196f. 223, 226f., 234
modules 102
– CDUs 103 o
– data management 21 objective function 11f., 15, 17, 35, 75, 77,
– library 103 145
– model management 21 – multicriterion 21
– planning 21 – problem 43
– tank 103 operator
Monte Carlo – expectation 68
– sampling 19, 61, 68, 79 optimal control 13, 63
– scheme 67 optimization
– simulation (MCS) 131 – methods 32
multinational enterprises 11 – robust (RO) 222
multiobjective Markov decision 16 – stage 5
 Index 345

organic growth 4 – Cider (CP) 122ff.

orthogonal arrays (OA) 67 – closures 6
outsource 2, 89, 117 – Concentrated Juice (CJP) 122ff.
overall routing time 40 – continuous 173, 195
overstocking 297, 301f. – goods manufacturing 207
– Milling (MP) 122ff.
p – multiproduct 173, 175, 177, 183
packaging 2 – multipurpose 173, 175, 177f., 183
packing tasks 207 – network-oriented 175
paper mills 238 – Packaging (PP) 121ff.
Pareto – primary 7
– curves 16 – secondary 7, 10
– optimal points 225, 229 – semicontinuous 173
– optimal solution 225, 232 – sequential 175
– optimal surface 223ff. policy
partial harvest 146 – Clear-a-Fraction (CAF) 320f.
patent lives 3 – heuristic 67
pickup and delivery problem (PDP) 30 – iteration (PI) 64, 67f.
– courier company (CCPDP) 31 – largest-backlog (LBP) 319, 321f.
pickup and delivery problem with time win- – maker 73
dows (PDPTW) 30ff. – optimal 81
– multi-vehicle (MVPDPTW) 30, – pricing 7
32ff. – storage 176, 181
– single-vehicle (1-PDPTW) 30, 33 – suboptimal 7
pip fruit business 121 – wraparound 317, 320
pipeline 91, 93f. – zero-wait 177, 180f.
– value 4 polymers and resins (P&R) 5
planning portfolio 3f., 10
– campaign 2 price
– capacity 244 – fresh fruit 164
– cider 121 – milled fruit 164
– corporate supply 244f. – packed fruit 165
– demand 270 – per unit 2
– enterprise resource (ERP) 240, 244, – waste 164
246ff. primary product flows 10
– inventory 237 probability
– Make-to-Stock 273 – -based approach 13, 18
– material requirement (MRP) 247f., – distribution 12, 61
273f. – functions 13
– methods 270 – space 12, 65
– mid-term 7, 270 – success 18
– multilevel 2 – transition matrix 84f.
– operative 244f., 269f. process
– production (PP) 237, 270f. – batch 2, 7, 173, 177
– scheduling 270 – complexity 239
– short-term 8, 270 – continuous 173, 195f.
– stops 250, 265 – control 285
– storage 237 – decision-making 220, 239, 291
– strategic 245, 251, 269 – flow-sheet 242
– tactical 245, 269f. – high value 2
– under uncertainty 16, 220f. – industry modeling system (PIMS)
plant 219
– batch 175, 178f., 183 – low volume 2
346 Index

– multistep 244 – schedule 6, 239ff.

– optimization 1 – season 5
– reaction 73 – secondary 10
– separation 73f. – shift 257f.
– system engineering (PSE) 269f. – short-term 6, 245
processing – site 8
– batch 177, 183 – speed 239
– downstream 198, 200 – stage 9
– equipment 175 – stops 250, 265
– modes 176 – strategy 244f., 251
– network 175 – unit 221
– rate 176, 197 – volume 5
– recipes 173 – yields 5
– sequences 175 productivity 3f.
– tasks 175ff. – crises 4
– time 222 – gains 4
– upstream 198, 200 profit
procurement 269 – after-tax 11
product – margins 4
– allocation 6f., 10 – maximization 13f., 177ff.
– chemical 91 – net 75, 121
– families 2, 11, 17 – total 145
– flow 2 profitability 16, 244, 246, 264
– high-demand 5 Program for the Endorsement of Forest Cer-
– high-end 244 tification Schemes (PEFC) 241
– high value 1f. pull systems 123
– low volume 1 purchases 14
– portfolio 2, 6, 8 – external 8
– primary 10 push systems 123
– quality 242, 250
– secondary 11 r
production 2 rate of accounting earnings 323
– active ingredients (AI) 2 raw material
– coordination 8 – basic 241
– facilities 2, 242f. – demand 246
– flexibility 7 – fake 261
– fluctations 28, 264 – feed 240
– fruit 130, 164 – mixed office waste (MOW) 241f.,
– levels 7 244
– management 261 – old newspapers (ONP) 241f., 244
– mid-term 6 – purchase 9, 14, 130
– optimization 239 – reception 135
– order-driven 245 – shortage 261
– performance 237 – stocks 5
– plan 253 – supply 1, 8, 293
– planning 5, 237ff. – type
– planning problem (PPP) 272, 275f. – virgin fiber (VF) 241f., 244
– primary 10 real time approximate dynamic program-
– profile 121, 130, 148, 153f., 252, 261 ming (RTADP) 61, 63, 65, 67f., 78f.
– program 256f., 260 – greedy 79
– -related data 248 recipe diagrams 175, 183
– requirements 315 recycled
– routes 175 – fiber 238, 241f.
 Index 347

– paper 241 – planning 6

– raw-materials 244 – volume 5
refinery 91 scheduling 239
– and petrochemical modeling system – batch 96, 175
(RPMS) 219 – cyclic 195
– configuration 93 – feedback 310, 319
– operations 91, 93f., 219ff. – function 247
– parameters 222 – level 2, 7
– profit 219 – plans 5
– real-world 92 – policy 313
– scheduling 91f., 221 – problems 252
– system 222 – process 175
– turnover 92 – production 96
refining 91 – reactive 313
relative weight unit (RWU) 264 – robust 223
repacking 310f. – short-term 173, 175, 177ff.
request – subproblems 275
– load-balanced 55 – under uncertainty 196, 219f.
– one-to-many 55 secondary product distribution 10
– single many-to-one 55 selling profile 130, 153
rescheduling 89f., 92ff., 279 semiconductors 289
– affected operations (AOR) 95 service level 7
– batch-plant 95 simulation 82
– heuristic-based 96ff. – -based optimization 5
– methodologies 96, 99, 115f. – procedure 83
– modified affected operations (mAOR) – stage 5
95 single-vehicle pickup and delivery problem
– operations 95 with time windows (1-PDPTW) 30
– problem 92, 94 single sourcing policy 11
– refinery 95f. sites
– shop- 95f. – Fruit Reception (FRS) 122f., 127
– short-term 96 – primary 9f.
– total MILP 105, 111, 117 – secondary 9f.
research and development (R&D) 3f. solution
resource 175 – branch-to-branch 43
– external 198, 246 – mathematical programming-based
– profile 121 239
– renewable 177 – optimal 51ff.
responsiveness 3, 7 – out-of-the-box 4
retailers 1 – robust 223, 225
revenues 11 stages 1
Ricatti equations 63 stochastic dynamic programming 13, 16
risk 13 stochastic programming 12ff.
– downside 16 stock keeping unit (SKU) 298
– financial 24 stock preparation 237, 242
robustness 230 – line 240f., 244, 251, 259ff.
– process 241
s – section 245f., 260
sales 4 storage
– activities 4 – bypassing 201f., 205
– cycle 307 – capacity 127, 133, 176, 293f., 306
– income 144f. – cold 127, 129, 151ff.
– limitations 14 – facilities 1, 9
348 Index

– finite intermediate (FIS) 176, 196, – pip Fruit Industry (FISC) 122ff.
200f. – planning 1, 4, 9
– flexible finite intermediate (FIS) – real-world 89
202, 205f., 209, 211 – redesign 4
– function 294, 297 – refinery 92
– juice 136 – resilience of 117
– no intermediate (NIS) 176, 196, – responsiveness 2
200, 202, 209f. – robust 89f., 96
– Nonprocessed Fruit (NPFS) 133, – strategy 7
136 – structure 9
– Packed Fruit (PFS) 133f., 137 – system 7
– policies 197 – tactical optimization 121
– profiles 260f. – tissue paper
– requirements 314, 316, 319ff. – units 70
– setup 251f. – variables 14
– space 176 – visibility 89
– task 202, 204ff.
– third party 150 t
– time 242 tax rates 10
– units 6, 205f. technology
– unlimited intermediate (UIS) 176, – multidisciplinary 239
197, 200, 208, 210 time
– utilization profiles 241, 260 – cleanup 208f.
supplier – discretization 219
– Third Party (TPS) 128ff. – duration of 242
supply-chain (SC) – horizon 6ff.
– activities 1 – intervals 173
– agrochemical 2 – minimum cycle 313, 316
– allocation problem 24 – representation 173f.
– analysis 1 – synchronization 250
– application 61 timing aspect 240
– BTX 69, 72, 76, 86 tissue
– complexity 239, 285 – consumption 238
– control system 286 – manufacturing process 241
– costs 4f. – mill 240f.
– design 13 – production capacity 238
– disruptions 89, 91, 93 tool 240ff.
– dynamic 285f., 291 – mixed integer strategic planning
– efficiency 237 (MISPT) 237, 240f.
– expenditure 2 – paper machines 240f.
– generic 9 – production planning 239
– global 7f. – scheduling 239
– high dimensional 61, 69 trade-off 3, 7, 13, 16, 117, 270
– intraorganizational 269 – offer-demand 146
– just-in-time (JIT) 296 trade regulations 2
– management (SCM) 1, 6, 9, 89, 91 trading structures 1, 9
– modeling 1, 4 transition function 76
– multiechelon 14, 291 transportation 2
– multienterprise 291 – nonrefrigerated 151
– network 333 – request 30, 32, 35f., 40, 44ff.
– operations 89f., 92f., 96, 117 – requirements 152
– optimization 2, 4, 8, 145
– pharmaceutical 2
 Index 349

u – synchronous value iteration 66

uncertain parameters 17, 224 variance
uncertainty 5, 12, 219, 228, 230 – high 5
– analysis 222 – low 5
– demand 5 vehicle
– factor 15 – allocation of 36
– multistage 85 – assignment of 36, 38
utilization – availability condition 36
– equipment 7 – capacity 40
– pattern 7 – fixed costs 43
– rate 7 – preassigning 45f.
Very Large Crude Carriers (VLCC) 91, 93
v
value w
– asynchronous value iteration 65f. warehouse 1, 7f.
– creation 324 waste
– function 64ff. – paper 241
– iteration (VI) 64ff. wood fiber 241