Nonlinear Process Control:

Applications of Generic Model Control

Other titles already published in this Series:

Parallel Processing for Jet Engine Control

Haydn A. Thompson

Iterative Learning Control for Deterministic Systems

Kevin L. Moore

Parallel Processing in Digital Control

D. Fabian Garcia Nocetti and Peter J. Fleming

Intelligent Seam Tracking for Robotic Welding

Nitin Nayak and Asok Ray

Identification of Multivariable Industrial Processes for Simulation,

Diagnosis and Control
Yucai Zhu and Ton Backx
Peter L. Lee (Ed.)

Nonlinear
Process Control:
Applications of Generic Model Control

With 77 Figures

Springer-Verlag
London Berlin Heidelberg New York
Paris Tokyo Hong Kong
Barcelona Budapest
Peter L. Lee
Head, Department of Chemical Engineering
The University of Queensland
4072 Queensland
Australia

ISBN-13:978-1-4471-2081-0 e-ISBN-13:978-1-4471-2079-7
DOl: 10.1007/978-1-4471-2079-7

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

Apart from any fair dealing for the purposes of research or private study, or criticism or
review, as permitted under the Copyright, Designs and Patents Act 1988, this
publication may only be reproduced, stored or transmitted, in any form or by any
means, with the prior permission in writing of the publishers, or in the case of
repro graphic reproduction in accordance with the terms of licences issued by the
Copyright Licensing Authority. Enquiries concerning reproduction outside those terms
should be sent to the publishers.

© Springer-Verlag London Limited 1993

Softcover reprint ofthe hardcover 1st edition 1993

The publisher makes no representation, express or implied, with regard to the accuracy
of the information contained in this book and cannot accept any legal responsibility or
liability for any errors or omissions that may be made.

Typesetting: Camera ready by authors

69/3830-543210 Printed on acid-free paper

SERIES EDITORS' FOREWORD

The series Advances in Industrial Control aims to report and encourage

technology transfer in control engineering. The rapid development of
control technology impacts all areas of the control discipline. New
theory, new controllers, actuators, sensors, new industrial processes,
computing methods, new applications, new philosophies, ... , new
challenges. Much of this development work resides in industrial reports,
feasibility study papers and the reports of advanced collaborative
projects. The series offers an opportunity for researchers to present an
extended exposition of such new work in all aspects of industrial control
for wider and rapid dissemination.
Nonlinear control theory has a long history reaching back to the
stability theory of nonlinear systems developed by mathematicians like
A.M. Lyapunov. What has been less successful has been the application
of nonlinear methodologies in industrial applications. This volume in the
Advances in Industrial Control series is a valuable contribution to the
task of rectifying this situation. Dr Peter Lee has drawn together a
number of industrial applications of the nonlinear control technique
variously known as Generic Model Control (GMC) or Reference System
Synthesis (RSS) or Nonlinear Process Model Based Control (PMBC).
The thrust of the technique is to match the simplicity of PI control tuning
with the knowledge which resides in existing nonlinear models of
industrial processes. The contributions in the volum~ are interesting and
extremely instructive. For not only are the technical problems of using
GMC covered but the contributors cover issues like how to obtain
operator acceptance of new untried techniques, and how to persuade
management to invest in new solutions. Equally instructive are the
financial payback figures given which result from the investment in new
advanced control schemes.
We are grateful to Dr Peter Lee and his co-authors for producing such
a thought provoking contribution to the Advances in Industrial Control
series.

Michael Grimble and Michael Johnson,

Industrial Control Centre,
University of Strathclyde,
Glasgow, Scotland, U.K.
PREFACE

Much has been written and said about the "theory and practice" gap that
exists in Process Control. Academic control theory has not been well
utilized in many industries. The big strides that were made in the 1960's
in 'Modern State-Space' methods and to a lesser extent, in the 1970's on
Adaptive Control, were largely ignored by industry in general. The PID
regulator continues to meet the needs of many practitioners. However,
even though the number of difficult control loops that cannot be solved
using a PID regulator is small, the economic impact of poor control in
these loops can still be quite large.
This book clearly is aimed at practitioners. It is not meant to be a full
exposition of nonlinear control methods in all its guises. I have
attempted, rather, to gather a series of case studies that illustrate
important aspects of implementing advanced control in an industrial
setting. Only one nonlinear control method is chosen for examination -
Generic Model Control. No claim is made in this book that this technique
is the best of the available nonlinear methods (although this author
would claim it does have some considerable advantages!) No compari-
sons are made in this book of alternative control strategies for that is not
the point. It is hoped that this book will demonstrate that advanced
control methods can successfully be implemented and that the economic
returns for doing so can be substantial.
The editor, Dr Peter Lee would like to take this opportunity to thank a
number of people who have contributed to this monograph.
Firstly to the contributors who have endured my bullying and cajoling
in equal proporitions, and who really made this work possible, a big
thanks is due - "goodonya mates". To their respective companies who
granted permission for the works to be published, we are all grateful for
their contribution to the profession.
To my colleagues at the CAPE Centre, Dr Bob Newell and Dr Ian
Cameron who put up with me and carried some extra load while I was
writing, a big thank-you.
To Christine Smith who most of all endured the drafts, revisions and
vagaries of the word processing graphics systems can only go my sincere
thanks and much appreciation for her dedication to quality.
Finally to my family Janet and Geoffrey who have foregone many
hours of leisure because "Daddy is working", my love and appreciation.

Peter Lee
Brisbane, 1993
CONTENTS

1 Introduction................................. 1
1.1 References................................. 4

2 Generic Model Control - The Basics . . . . . . . . . . . . . . . . . 7

2.1 Introduction................................ 7
2.2 The Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 8
2.3 Reference Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Using Steady State Models . . . . . . . . . . . . . . . . . . . . .. 21
2.5 Constraint Handling Strategy. . . . . . . . . . . . . . . . . . . .. 25
2.6 Robust Stability and Performance Analysis ........... 31
2.7 Applications................................ 35
2.8 References................................. 37

3 Industrial Applications of Reference System Synthesis/Generic

Model Control: Wastewater pH Control . . . . . . . . . . . . . .. 43
3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Introduction................................ 43
3.3 Process Description. . . . . . . . . . . . . . . . . . . . . . . . . .. 45
3.4 Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.6 Controller Design and Tuning . . . . . . . . . . . . . . . . . . .. 52
3.7 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 57
3.9 Conclusions................................ 61
3.10 Acknowledgement. . . . . . . . . . . . . . . . . . . . . . . . . . .. 63
3.11 Nomenclature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 63
3.12 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 64

4 Using Tray-To-Tray Models for DistiUation Control ....... 67

4.1 Introduction................................ 67
4.2 Description of Columns Considered . . . . . . . . . . . . . . . . 71
4.3 Controller Models . . . . . . . . . . . . . . . . . . . . . . . : . . .. 76
4.4 Implementation Issues . . . . . . . . . . . . . . . . . . . . . . . .. 85
4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 93
4.6 Economic Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5 Automatic Moisture Control in Particulate Dryers . . . . . . . . . 105
5.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3 Rationale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4 Research and Development of the Control System ....... 112
5.4.1 Moisture Control . . . . . . . . . . . . . . . . . . . . . . . . 112
5.4.2 Conveyor Belt Speed . . . . . . . . . . . . . . . . . . . . . . 115
5.4.3 Zone 1 Temperature . . . . . . . . . . . . . . . . . . . . . . 116
5.4.4 Zone 2 Temperature . . . . . . . . . . . . . . . . . . . . . . 117
5.4.5 Zone 3 Temperature . . . . . . . . . . . . . . . . . . . . . . 117
5.4.6 Dryer Control Modifications . . . . . . . . . . . . . . . . . 118
5.5 Development of the Controller . . . . . . . . . . . . . . . . . . . 118
5.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.7 Cost Benefit Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.10 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.11 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

6 Eccentricity Control of a Cable Jacketing Line . . . . . . . . . . . 133

6.1 Introduction................................ 133
6.2 The Application .. . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.3 The Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

7 The Development of a Nonlinear Adaptive Generic Model

Controller for Chemical Reaction Quality Control . . . . . . . . . 151
7.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.2 Introduction................................ 152
7.3 Chronological Development of the Control Strategy. . . . . . 153
7.3.1 July 1990 - The Process . . . . . . . . . . . . . . . . . . . . 153
7.3.2 August 1990 - The Model . . . . . . . . . . . . . . . . . . . 155
7.3.3 October 1990 - The Controller . . . . . . . . . . . . . . . 158
7.3.4 November 1990 - Tuning of the GMC Controller ... 159
7.3.5 June 1991 - Tuning of the Adaptation Algorithm .' .. 163
7.3.6 January 1992 - Model Validation . . . . . . . . . . . . . . 165
7.3.7 March 1992 - Economic Benefits . . . . . . . . . . . . . . 168
7.3.8 Process Monitoring . . . . . . . . . . . . . . . . . . . . . . . 170
7.4 Summary.................................. 172
7.5 Issues in Nonlinear Adaptive Control . . . . . . . . . . . . . . . 173
7.6 References................................. 175
8 Blast Furnace Stove Control . . . . . . . . . . . . . . . . . . . . . . . 177
8.1 Introduction................................ 177
8.2 Blast Furnace Stoves . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.2.1 Physical Description of Full-Scale Stoves ......... 177
8.2.2 Laboratory Scale Stove System . . . . . . . . . . . . . . . 181
8.2.3 Control Objectives, Measurements and Manipulated
Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.2.4 Control Difficulties . . . . . . . . . . . . . . . . . . . . . . . 184
8.3 Model and Control Law Development. . . . . . . . . . . . . . . 188
8.3.1 Blast Air Temperature . . . . . . . . . . . . . . . . . . . . . 188
8.3.2 Blast Flow-rate GMC Control Law . . . . . . . . . . . . . 190
8.3.3 On-gas Stove Model . . . . . . . . . . . . . . . . . . . . . . 194
8.3.4 Long-term Control Algorithm . . . . . . . . . . . . . . . . 196
8.3.4.1 Limiting case operations . . . . . . . . . . . . . . 198
8.3.4.2 The optimisation problem . . . . . . . . . . . . . 200
8.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.4.1 Functional Hierarchy . . . . . . . . . . . . . . . . . . . . . . 204
8.4.2 Hardware and Software Hierarchy . . . . . . . . . . . . . 205
8.4.3 Solution of the Optimization Problem . . . . . . . . . . . 210
8.4.4 Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
8.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
8.5.1 Short-term Results . . . . . . . . . . . . . . . . . . . . . . . 214
8.5.2 Long-term Results . . . . . . . . . . . . . . . . . . . . . . . . 230
8.6 Economic Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.7 References................................. 242

9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
CONTRIBUTORS

Chapter 3: An Industrial Application of Reference System Synthesis

Generic Model Control: Wastewater pH Control
R. Donald Bartusiak,
Exxon Chemical Company, Texas, USA

Chapter 4: Using Tray-To-Tray Models for Distillation Control

James B. Riggsl, Martin Beauford2 and Jackie Watts2
lDepartment of Chemical Engineering, Texas Tech University,
Texas, USA
2Phillips 66 Company, Borger, Texas, USA

Chapter 5: Automatic Moisture Control in Particulate Dryers

D. Doerr, P.L. Douglas 1 and M.G. Whaley2
lDepartment of Chemical Engineering, University of Waterloo,
Waterloo, Canada
2Dantec Systems Corporation, Waterloo, Ontario, Canada

Chapter 6: Eccentricity Control of a Cable Jacketing

B. W. Surgenor
Department of Mechanical Engineering, Queen's University, Kingston,
Canada

Chapter 7: The Development of a Nonlinear Adaptive Generic Model

Controller for Chemical Reaction Quality Control
Barry J. Cott,
Sarnia Manufacturing Centre, Shell Canada Limited, Ontario, Canada

Chapter 8: Blast Furnace Stove Control

G.A. Labossiere and P.L. Lee
Computer Aided Process Engineering Centre, Department of Chemical
Engineering, The University of Queensland, Queensland, Australia
 CHAPTER 1
INTRODUCfION

The need for high performance control systems has accelerated over the past
decade. This need has been pushed by economic pressures related to
increased throughput, higher quality products produced more consistently,
increased utilization of raw materials and decreased energy utilization. The
rise of environmental and safety issues and a tighter material and energy
integration has also added to the demands of modern control systems. At the
same time, increased computational power through modem distributed control
systems has provided the means to implement many advanced control
algorithms at reasonable costs.

While the need for high performance control systems is well known to most
practitioners, less known is the means to quantify the benefits of
implementing Advanced Process Control (APC). The Warren Centre Study
(Marlin et al 1987) examined a series of case-studies and documented a
methodology for justifying the expenditure required to implement APC.
Despite this excellent work, one reason often cited by industrial practitioners
for not implementing APC is the inability to cite other successful
Introduction Chapter 1

implementations. It is hoped that this monograph may in some small way

alleviate this problem.

Most advanced algorithms rely, either implicitly or explicitly, on a process

model to perform the control calculations. The development of process
control algorithms has in the past taken a very different approach in using
process models than other process systems related activities, including
process design and optimization. The latter activities have made use of
nonlinear, multivariable models, often based upon a physical understanding of
the process behaviour. These models have often been collated and coded in
standard simulation libraries such as ASPEN (Evans et al 1979) and others.
Of course, these models have predominantly only considered the steady-state
behaviour of the process. In contrast, process control has relied primarily on
linear, single-input, single-output models, often derived from empirical plant
tests. That this approach has failed to yield adequate results in all cases is
hardly surprising. Most process plants behave in a nonlinear way, exhibiting
interactions between many of the process variables. Thus, increasingly there
is interest in developing methods that deal directly with these issues.

Generic Model Control (GMC)(Lee and Sullivan, 1988) was developed with
the specific objective of incorporating nonlinear, multivariable process models
directly in the control algorithm. Independently, similar approaches were
developed by Bartusiak et al (1988) called Reference System Control (RSC)
and by Balchen et al (1988) called Internal Decoupling. These approaches
are all similar and can be traced to some earlier work by Liu (1967). They
are also part of a distinct body of mathematical knowledge known as

2
Introduction Chapter I

differential geometry. One of the key differences in these approaches is the

way in which performance is defined. While attaining the required setpoint
for each controlled variable is still maintained as necessary, an additional
requirement is also added. This additional requirement specifies the rate of
approach towards the setpoint and it is this specification that differentiates
these methods from other techniques.

This monograph will introduce the required "theory" in an attempt to convey

the basic principles of GMC. This is written in a tutorial manner to help to
explain the basic concepts of the approach. Details of the underlying theory
and extensions will be left to the reader to pursue, if interested, through the
cited references. A series of case studies describing the application of GMC
to a number of processes then follows. These case studies highlight
implementation aspects as well as the benefits of applying this technique to
the particular problem described. Finally some overall conclusions are drawn
in the final chapter.

The case studies presented in this monograph have been chosen to cover a
wide range of different applications across a range of industries. The
processes chosen also exhibit a variety of characteristics that make process
control difficult. Table 1.1 summarizes the characteristics of each case study.

3
Introduction Chapter 1

Table 1.1 Case Study Characteristics

Chapter Case Study Input! Deadtimes Constraints

Output Present Present
Dimensions
3 pH Control 1x 1 NO NO
4 Distillation Control 2x2 NO YES
5 Dryer Control 1x 1 NO NO
6 Extruder Control 2x2 NO NO
7 Reactor Control 1x 1 YES NO
8 Stove Control 2x2 NO YES

In this monograph the tenns Generic Model Control (GMC), Reference

Systems Synthesis (RSS), and Process Model Based Control (PMBC) will be
used interchangeably.

1.1 REFERENCES

Balchen J.G., Lie B. and Solberg I. (1988) Internal decoupling in nonlinear

process control. Modelling, Identif. and Control 9: 13 7 -148
Bartusiak R.D., Georgakis C. and Reilly MJ. (1988) Designing nonlinear
control structures by reference system synthesis. Proc. ACC, Atlanta GA
USA, 1585-1590.
Evans L.B., Boston J.F., Britt H.I., Gallier P.W., Gupta P.K., Joseph B.,
Mahalec V., Ng E., Seider W.D. and Vagi H. (1979) ASPEN: An
advanced system for process engineering. Comput. & Chern. Eng. 3:319-
327.

4
Introduction Chapter 1

Lee P.L. and Sullivan G.R. (1988) Generic Model Control (GMC). Comput
& Chem. Eng. 12:573-580
Liu S-L. (1967) Noninteracting process control. Ind. Eng. Chem. Proc. Des.
and Dev. 6:460-468
Marlin T.E., Perkins J.D., Barton G.W. and Brisk M.L. (1987) Advanced
Process Control. The Warren Centre for Advanced Engineering.
University of Sydney, Sydney, Australia.

5
 CHAPTER 2
GENERIC MODEL CONTROL - THE BASICS

2.1 INTRODUCTION

The basic concept of Generic Model Control (GMC), (Lee and Sullivan,
1988), and its closely aligned cousins Reference System Synthesis (Bartusiak
et aI, 1988) and internal decoupling (Balchen et al, 1988) is to find values of
the manipulated inputs that force a model of the system to follow a desired
reference system or trajectory. The methods are clearly related to a body of
mathematical knowledge known as differential geometry involving exact
linearization of nonlinear mappings between the input and output variables
(lsidori, 1989). The purpose of this chapter is not to present all of the
theoretical underpinnings of the technique, but rather to provide enough of
the essential elements of the method in a tutorial style that will allow the
reader to appreciate the applications of the technique presented in each of the
subsequent case studies.
Generic Model Control - The Basics Chapter 2

2.2 THE BASICS

Consider a dynamic model of a process described by a set of differential

equations:

y = f(y,u,d,t,8) (2.1)

where y is a vector of process outputs of dimension m, u is a vector of

process inputs of dimension m, d is a vector of process disturbances of

dimension I, t is time and 8 is a vector of model parameters of dimension q.
In general f is a vector of nonlinear known functional relationships. Note
also in this simplified presentation we will only consider square systems - i.e.
where the number of inputs and outputs are the same. However, the
technique is not limited to only such systems (Lee and Sullivan, 1988).

The second component of the algorithm is to define a reference system, r(y).

This reference system defines a desirable rate of change of the output

variables,(y)*, This notation is consistent with the notation that defines the
setpoint of y as y.,

One reasonable choice of the reference system r(y) is:

r(y) =(y)* =K. (y. - y) +~ "f (y. - y)dt

o
(2.2)

where ~ is the current instant of time.

This equation expresses the desires that:

8
Generic Model Control - The Basics Chapter 2

(a) When the system is a long way from setpoint, we would like the system
to be travelling towards the setpoint quickly.
(b) We would like to ensure offset free performance and thus if the system
has been away from setpoint for some time we would like it to start
moving more quickly towards setpoint.

The values of Kl and ~ can be chosen to specify a range of different process

behaviours. Their choice will be discussed in section 2.3. Bartusiak et al
(1989) examined other possible choices for the reference system I(y).

The third and final element of the control algorithm is to ensure that the rate
of change of the outputs follows the desired reference system. Thus:
y =(Yt (2.3)

Using equations 2.1 and 2.2 yields:

f(y,u,d,t,9) "(y*-y)dt
=K.(y*-y) + Kz f
o
(2.4)

The control law to be solved at every sample time, equation 2.4, for the
manipulated inputs, u, is a set of nonlinear, algebraic equations in the
unknown variables.

A simple single-input, single-output problem may help to clarify the control

law. Consider a simple liquid level control problem shown in Figure 2.1.
The control problem is to control the liquid level by .adjusting the inlet
flowrate MI. A disturbance inflow M2 also exists.

9
Generic Model Control - The Basics Chapter 2

M1
£ M
2

i
h

Figure 2.1 Gravity Underflow Tank

A simple model for this process is

dh
dt
=M1 + M2 - c.jh (2.5a)

since

(2.5b)

where h is the height of liquid in the tank and C is a constant. Comparing

this specific model with the general case shown in equation 2.1, it is seen
that
h - y, the output variable
Ml == u, the input variable
M2 == d, the disturbance variable
and C = e, a model parameter.

10
Generic Model Control - The Basics Chapter 2

For this system r(y) is given by:

. ~
(ht =Kl (h * - h) + K2 i (h * - h)dt (2.6)
o
where h * is the setpoint for h.

Using equations 2.4, 2.5 and 2.6 yields:

~
Ml +M2 -c{h =Kl(h* -h) +K2 i(h* -h)dt (2.7)
o
or
~
Ml = -M2 + c.;h + Kl (h* - h) + K2i(h* - h)dt (2.8)
o
In this case the control law given by equation 2.8 is an explicit expression for
the manipulated input Ml and is shown in Figure 2.2. Sometimes however, it
is not possible to algebraically manipulate equations 2.4 to obtain such
explicit expressions. In these cases equation 2.4 must be solved numerically.
However, the task of solving the nonlinear simultaneous equations is often
very fast, only requiring one or two iterations of a numerical method. This is
because you can provide such a good initial guess for the solution - the
values of the manipulated inputs at the previous sample time.

The derived control law in equation 2.8 represents a feedbacklfeedforward

controller. The feedforward component would require measurement or
estimation of the disturbance flowrate M2. The control law is also dependent
upon the model parameter C. Signal and Lee (1992) show how these
disturbances and parameters may be estimated on-line when the~ values are
not known. Much of the work on GMC has been to demonstrate how

11
Generic Model Control - The Basics Chapter 2

"robust" the method is when such disturbances, parameters and other model
inaccuracies are present (Lee, 1991). An illustration of this property will be
given in an example at the end of section 2.3, and brief summary of the
major results given in section 2.6.

A final comment on the basic algorithm is required. A solution to equation

2.4 is only directly possible when the manipulated variable chosen to control
a particular output appears in the model equation for that output These types
of systems are known as relative degree one. Other approaches can be taken
when this is not the case (Lee, 1991; Henson and Seborg, 1990; Signal,
1992).

M2

h* + M1 h
GMC
<
PROCESS

Figure 2.2 Block Diagram of Control Structure

12
Generic Model Control - The Basics Chapter 2

2.3 REFERENCE SYSTEMS

As discussed in section 2.2, the reference system is one of the key elements
in the GMC control law. The form given in equation 2.2 is only one of
many possible choices (Bartusiak et al, 1989), and contains two parameters
Kt and Kz. These are diagonal matrices, where the diagonal elements are
chosen for each output independently. Thus, one can use these values to
select any "reasonable" desired response for the system. "Reasonable"
implies that the parameters are chosen in relation to the system's natural
dynamic response. How well the system matches this reference system will
be governed by how closely the chosen model matches the plant behaviour.

If Laplace transforms of equation (2.2) are taken for a single-input, single

output system, the resulting transfer function becomes:
Y(s) 2't~s +1
(2.9)
=-,.......::....---
Y'(s) 't2S2 +2't~s +1
where

1 k
't = - - and ~ =_1_
{k; 2{k;
This system does not yield the same response as a classic second-order
system (Stephanopoulos p. 188, 1984), due to the presence of the zero in the
transfer function. However, similar plots to the classic second-order response
showing the normalized response of the system yly. vs normalized time ti't
with ~ as a parameter can be produced as is shown in Figure 2.2. The design
procedure can be specified as follows:

13
Generic Model Control - The Basics Chapter 2

1. Choose ~ from Figure 2.3 to give desired shape of response.

2. Choose t from Figure 2.3 to give "appropriate" timing of response in
relation to known or estimated plant speed of response.
Calculate k t and k2 using the following equations:

k =2~ (2.10)
1 t

1 (2.11)
k =-2
2 t

An alternative approach is to examine the characteristic equation of the

closed-loop system. From equation 2.9, the characteristic equation is given
by:

S2 +ks +k2 =0 (2.12)

1

It may be desirable or more natural to specify the location of the two closed
loop poles and then solve equation 2.12 for the unknown values of k t and k 2•

To bring some of these concepts into focus consider the control of an

anaerobic digestor.

An anaerobic digestor is a process of converting organic material in a liquid

stream, by a series of biological reactions, to methane and carbon-dioxide
gas, and an effluent stream of less concentration. It is commonly used in
treating waste-water streams.

14
Generic M ode/ Control - The Basics Chapter 2

NORMALISED RESPONSE
1.4~-----------------------------------------'

1.2~----~~----~~------------------------~

o.e~~------------------------------------~

0.2~----------------------------------------~

o 2 4 e 8 10
'N ORMALISED TIME t I 'T
~ -r- '0.5' ---- '1.0' -i:r '3.0' -M- '5.0'

Figure 2.3 Reference System Curves

Consider a simplified model of anaerobic digestor described by Costello et al

(1989b):

-dS
dt
=D(S.m - S) - <Xl QCH• (2.13)

where S is the organic eftluent discharge concentration

(g COD/dm3)
Sin is the organic influent concentration (g COD/dm3)
D is the dilution rate (hr-l) (inlet flowrateivolume of reactor)

15
Generic Model Control- The Basics Chapter 2

QCH is the production rate of methane gas (dm31hr)

•
0. 1 yield parameter

The variable to be controlled is the organic eftluent concentration and the

manipulated variable is the dilution rate, D. Defining a reference trajectory
for S yields

( dS)* =Kl (S· - S) + K2 ) (S· - S)dt (2.14)

dt 0

Applying equation 2.3 to equations 2.13 and 2.14 and solving D yields
t,.
K 1(S* - S) + K2 f (S* - S)dt + 0. 1 QCH (2.15)
D = 0 •
SIN -S
This represents a nonlinear control law that will compensate for changes in
process behaviour.

The control law of equation 2.15 and the system described in equation 2.13
were simulated using MATLAB (1990). The script file for this simulation is
shown in Figure 2.4. The values of the simulation constants SIN' a, andQCH
•
were those used previously by Costello et al (1989b) and are representative
of those found in some industrial processes. The value of the GMC reference
trajectory parameter 't was chosen such that closed loop time constant should
be 9 hours (compared with the open-loop time constant of 21 hours). This
was chosen to illustrate the effective improvement that can be obtained using
the GMC algorithm. The value of the other reference trajectory parameter ~

16
Generic Model Control - The Basics Chapter 2

was chosen on the basis that some overshoot in this process is acceptable.

Figure 2.5 shows the result of controlling the process using "perfect control".
A step change in the inlet concentration, SIN, was introduced after 10 hours.
The controller, equation 2.15, knew of this change, contained no parameter
errors, and hence maintained perfect control.

Figure 2.6 shows what occurs when no feedforward action is included in the
control law. In this instance, the value of the inlet concentration, SIN' used in
the control law was always constant at 30 gCOD/dm3, while the actual inlet
concentration changed to 60 gCOD/dm3 after 10 hours.

Figure 2.7 shows what occurs when parameter mismatch exists in the control
law. The value of the parameter, a., used ip the control law, was 50% larger
than the value in the "plant". The same disturbance as was used previously
was again applied, but in this instance it was assumed that the controller did
have feedforward information. Notice also that the controller takes action
from time zero as the controller is no longer initialized correctly.

17
Generic Model Control - The Basics Chapter 2

% GMC Control of an Anaerobic Process

%
% Define Some constants
%
clear
simtim = 80.0; % Simulation Time
sin = 30.0; % Inlet Substrate Concentration
alpha 1 = 1. 0;
qch4 = 1. 0;
tau = 9.0; % Reference Trajectory Specification
xi = 1.0;
k1 = 2.0 * xi / tau;
k2 = 1.0 / tau / tau;
s = 9.0; % Initial values
d = 0.0476;
setpoint = 9.0;
esum 0.0;
splt = [splt ; s];
dplt = [dplt ; d];
%
% Main simulation loop
%
for i = l:simtim
%
% Calculate Control action
%
if i > 10
sin = 60;
end
err = setpoint - s;
esum = esum + err;
reftraj = k1 * err + k2 * esum;
d = (reftraj + alpha1 * qch4) / (sin - s);
if d < 0
d = 0.0;
end
%
% now simulate process response
%
dsdt = d*(sin - s) - alpha1*qch4;
s = s + dsdt * 1.0;
%
% form plotting arrays
%
splt = [splt ; s];
dplt = [dp1t ; d];
ds = [ds ; dsdt];
end
subplot (211), plot (splt), title('s');
subplot (212), plot (dplt), title('d');
subplot

Figure 2.4 MATLAB script file for Anaerobic System

18
Generic Model Control- The BaYics Chapter 2

:--.--:
S ICXQdm3

':I~ . . . . . . - - r
: - - :- . - - - - - - . :: - - - . - - :- . - - - , :
]

o 10 20 30 40 50 eo 70 80 110
.1 _.
D •

OA:~I'\ - - :....----.--:
: - ::---.--:~,I
o 10 20 30 40 50 eo 70 80 110
1ft.

Figure 2.5 "Perfect" Control

/---.---,
SlCXQdm3

:~f~ . . . . . . - -,
r -
- -.,- - - - - - . ,: . . . . - - - - . - - .- . -
] - - "

~I~:~--:-: :~:---.--:- -.- - ,:I

o 10 20 30 40 50 eo 70 80 110
4 1ft II
DIr

o 10 20 30 40 50 eo 70 80 110
1ft.

Figure 2.6 No Feedforward Control

19
Generic Model Control- The Basics Chapter 2

S gCCQdlll3

"r-----~--------~--------~----~--------~----~~----~--------~----_,

30 150 eo 70 eo 80

;,;:' :;1
DIII·1

·~6 o '0 20 30 40 50 eo 70 80 80
'llllltir

Figure 2.7 Parameter Error

The results of both Figures 2.6 and 2.7 clearly illustrate the robust behaviour
of GMC. Despite modelling errors (Figure 2.6) or parameter errors (Figure
2.7), the controller is able to return the process to setpoint with zero offset in
a smooth manner.

Figure 2.8 illustrates the effect of changing the reference system parameters.
In this case, the value of 't used in equation 2.11 is half that used for
obtaining the results shown in Figure 2.7. It can be seen that the speed of
response has indeed been increased, with the transient finished by 40 hours
compared to 80 hours in Figure 2.7.

20
Generic Model Control - The Basics Chapter 2

S.~3

~.:r----r--[~~._: ----.----.::----------'-....---,:1
o 10 20 30 40 50 eo 70 80 80
'BIll.

~5 o 10
: : ' :: : :1
20 30 40
'DIll II'
50 eo 70 80 80

Figure 2.8 Effect of Reference System Parameters

2.4 USING STEADY STATE MODELS

The basic algorithm as described in section two has been extended to include
the use of steady-state models, systems with dead-time, coping with model
uncertain~ and controlling systems in the presence of constraints. A brief
review of using steady-state models is given in this section, and an extended
development is to be found in Cott et al (1989).

The process model of equation 2.1 assumes an approximate dynamic model

of the process can be derived, but it is far more common to have access to
models that only contain steady-state information, i.e.

21
Generic Model Control - The Basics Chapter 2

(2.16)

where flS represents the approximate model of the steady-state behaviour.

Although these models describe the steady-state, nonlinear, interactive

behaviour of the process to a better accuracy than corresponding linear cause-
and-effect models, for control purposes, some estimate of the process
dynamic response is required. The most likely estimates available to the
designer are the average time constants of the process obtained from step
response tests. Although these estimates may be inaccurate at different
operating conditions, the degree of approximation is often sufficient to obtain
good control performance.

Assuming that the step response data can be represented by a first-order

response, a simple estimate of the time response of the output variables in
moving from one steady-state to another can be given as:
y~ rl(yu -y), (2.17)

where T is a diagonal matrix of the estimated open-loop time constants, and

Yu are the ultimate or steady-state values of the output variables if no further
control action is taken. The diagonal elements of the matrix T are
"averaged" time constants of the output variables based on step changes of all
input variables. Combining this approximate description of the dynamic
response of the process with the reference system or performance
specification of equation 2.2, the ultimate response can be calculated as:

22
Generic Model Control - The Basics Chapter 2

yu =y + T(K.(y· - y) + ~ "of (y. - y)dt) (2.18)

The control action required to achieve this performance specification can be

found by replacing y with Yu in the nonlinear, steady-state model described

in equation 2.16 and solving for the control variables u. An example will
clarify this procedure.

Consider the problem of controlling the anaerobic digestor described in the

previous section. Imagine that only a steady-state mass balance is available
to describe this system:
D(SIN - S) - a 1()
'<ell.
=0 (2.19)

This equation is in the form of equation 2.16. Applying equation 2.18 for the
output variable S yields:
t,.
Su =S + t p (K1(S· - S) + K,/ (y* - y)dt ) (2.20)
o
where tp is an estimate of the open loop time constant Using equation 2.20
to determine the ultimate value of the substrate concentration if no further
control action is taken, equation 2.19 can be used to calculate the
manipulated variable required to achieve Su by substituting Su for S:

D = 0. 1 ~. (2.21)
(SIN - Su)

Equations 2.20 and 2.21 form the complete control algorithm. Note that if:

23
Generic M ode/ Control - The Basics Chapter 2

t K\ = I
and K2 = 0
then Su = S*.

Thus the control law will result in calculating a value of D that according to
the steady-state model should give the required setpoint. Equation 2.20. will
compensate for model errors and give improved dynamic performance.

It should be pointed out that the use of equation 2.17 does not require the
steady-state model to be explicit in either output or manipulated variables,
making the availability of suitable models to develop control laws much
greater than if explicit models were required. Depending on the model
structure and number of control variables, solving for the control variables
may be straightforward, or may require numerical methods for solving
systems of nonlinear equations.

Implicit steady-state process models, such as those described in equation 2.16

do not model the dynamic effects of the disturbance variables d. A simple,
but effective method to improve the model's performance is to pass the
disturbance variables through appropriate dynamic filters (eg. first-order lags)
to compensate for the lack of dynamic structure in the process model. These
filters need to be tuned on-line to achieve good control performance. This is
similar to conventional feedforward control.

24
Generic Model Control- The Basics Chapter 2

1.S CONSTRAINT HANDLING STRATEGY

The control law derived in section 2.2 was derived by equating equations 2.1
and 2.2. Equally however, a control law could be developed by posing the
problem as an optimization problem (Lee and Sullivan, 1988). This
optimization problem would seek to minimize the difference between the
process and reference system by choosing specific values of the manipulated
variables. Of course equating is certainly minimizing. However, when
constraints on the inputs and outputs are present, a constrained optimization
approach is more suitable. A method where slack variables defining the
variables departure from the chosen reference trajectories are added to the
control law for both the control variables and constraint variables has been
proposed (Brown et al, 1989). Selecting the weighting factors on these slack
variables and defining a control objective function which is dependent on
these weighted slack variables allows the controller to achieve the desired
compromise between constraint violation and setpoint tracking. The solution
becomes a nonlinear constrained optimization problem.

Consider the process described by equation 2.1 where the number of outputs
and the number of inputs are equal.

It is desirable to have the system operate within the feasible region such that
for the q constraints:

CLi S "'i(y,u,d,t,9) S Cui; i =1 ... q (1.11)

where "'I represents some nonlinear function such that:

25
Generic Model Control - The Basics Chapter 2

As well, both input constraints and input movement constraints are defined
for the m controls, such that:
i=l ... m (2.23)

(2.24)

If the output of the process is such that it lies outside the feasible region or
that the current output variable trajectories will violate the given process
constraints, then it is desirable to operate the system such that the rate of
change of the constraint variable Ci approaches its constraint value according
to the reference trajectory:

i =1 ... q (2.25)

or
de.
- ' = K1C·(e. - CLi) ; i =1 ... q (2.26)
dt "
where K1Ci are chosen a-priori using the GMC reference trajectories.

The dynamics of the constraint variables can be found by utilizing the chain
rule on equation 2.22. Alternatively, an approximation like equation 2.17
would be to assume that the constraints will move to their constraint value in
a first order manner according to:

dC i =Tc-~ (C~M - C,.) i =1 ... q (2.27)

dt 'I

26
Generic Model Control - The Basics Chapter 2

C;"M = 'Vi (yMM,u,d,t,9) ; i =1 ... q (2.28)

f(yAlM,u,d,t,9) = 0 (2.29)

The q slack variables, lo~ and lo~, defining the variables departure from the

chosen specification curves are added to equations 2.25 and 2.26 for the
constraint variables. For the case of the known constraint dynamics,
equations 2.25 and 2.26 become:
de.
_I _ lo~ S K1Ci(CU . - e.) i =1 .•. q (2.30)
dt 1 1

and
dC i
+ loci S K1c·(e. -eu ) ; 1 =1 ... q
+ •
- (2.31)
dt 1 1

where lo~ and lo~ represent the variables departure from the chosen reference

trajectories for the upper and lower constraints respectively. Defining a

reference system for the constraint variables not only allows the system to
ensure that the constraint is not violated, but provides control over the rate of
approach to the constraint through the selection of K1Ci '

A set of slack variables can also be incorporated into the performance curves
to denote the systems efficiency in terms of setpoint tracking. If two

mutually independent slack variables, A.; and A.;, are defined to express the
systems negative offset and positive offset from the prespecified response
trajectory, the GMC control law for the system performance in equation 2.4
can be written as:

27
Generic Model Control - The Basics Chapter 2

where y." represents the process setpoints.

The solution of the constrained multivariable control problem can be solved

as a nonlinear constrained optimization problem, which minimizes a function
of the slack variables.

The overall control problem can be formulated as the following single step
nonlinear optimization problem:
NLP1:

Choose: i =1 ... q

To Minimize:

J --0) .,('1"'p11
+,2 ('1 -,2 ("A.cv
+\2 (" -,2 (A •• )2 i• =1 ... m (2.33)
PI
+0) ".'''JIII +0)",
PI CJ
+0)". A.cjI
CJ
+O)Aui UU j
J =1 ... q

where O)j and O)j are weights such that

and

28
Generic Model Control- The Basics Chapter 2

Subject to:

\
f(y,u.d,t,9) +A;-A; =K\i(Yt -Y j ) +K2j I (yt -y)dt ; i=l ...m
o

(2.34)

(2.35)

(2.36)

(2.37)

llu ~ u. ~ 1lu;;i =l ...m (2.38)

(2.39)

A- ~ O· i =l ...m (2.40)
cJ 'j=l ...q

The overall problem described above can be solved as a single time step
NLP.

The fact that separate control perfonnance curves have been defined for the
constraints as well as the controlled variables provides a great deal of
tlexibility in the controller design. The weighting terms are re$ponsible for
assigning a priority level to the various control objectives, while the

29
Generic Model Control - The Basics Chapter 2

performance parameters are able to define the trajectories of the constraint

paths. Therefore, if the set point of a given control variable is increased
above a heavily weighted upper variable constraint, then the control variable
trajectory will deviate from its pre-defined performance curve to follow the
trajectory defined for the constraint variable. It is the degree of flexibility in
establishing the proper balance between the violation of the constraint
variables and the deterioration of the control performance, coupled with the
ability to predefine the reference trajectories for both the controls and
constraints, which are the merit of this technique.

The optimization problem which arises can be solved using a nonlinear

constrained optimization algorithm. The form of the optimization problem is
well structured since a slack variable is added to each control law equation to
ensure that a solution to the set of equations does in fact exist. If the control
is implemented at a reasonable frequency, the solution of the NLP is very
fast (3 to 4 iterations) since the current control settings and slack variables
provide a good initial estimate of the solution vector.

Brown et al. (1989) successfully demonstrated the application of this

approach to the simulated forced-circulation evaporator. The simulation
study explored the effects of changes in the algorithm parameters and clearly
demonstrated the efficiency and ease of the approach.

30
Generic Model Control - The Basics Chapter 2

2.6 ROBUST STABILITY AND PERFORMANCE ANALYSIS

One of the most important properties of any system is closed-loop stability.
Unfortunately the analysis of the stability of nonlinear closed-loop feedback
systems is quite difficult. and requires advanced mathematical concepts and
tools. Secondly, when a control algorithm relies upon a mathematical model,
it is important to understand the effect of modelling errors on the closed-loop
stability and performance. This is known as Robustness Analysis, an area
also made difficult by advanced mathematical concepts.

The robust stability and performance analysis of GMC has been developed by
Zhou et al (1992b) and by Signal (1992). These works provide a detailed
analysis and only a summary of the key results will be presented here.

Both Zhou et al (1992b) and Signal (1992) show that:

1. If no model errors exist and the model is minimum phase
2. If no un-measured disturbances exist
3. If no constraints on the manipulated variables exist
4. If the sampling interval is insignificant compared to the process
dynamics
then the closed loop system will be stable for any positive values of the
reference trajectory parameters ~ and t. This is the "Nominal" stability
criterion and is quite restrictive in terms of the assumptions made. However,
the proof involved the application of Lyapunov theory and the demonstration
that a Lyapunov function could be found (Zhou et al, 1992b).

31
Generic Model Control- The Basics Chapter 2

Signal (1992) analysed the closed-loop system by linearizing the nonlinear

system about some nominal operating point. Obviously this approach cannot
then make global claims about the stability and performance of the system,
but rather claims that are valid in the region around the point of linearization.

Consider the linearized process model to be given by:

(2.41a)

y = ex (2.41b)

and the real process as:

.t=.Ax+Bu+Dd (2.42a)

y = ex (2.42b)

The analysis also considered the presence of noise and unmeasured

disturbances. The closed-loop system in the presence of modelling errors is
stable if:

(2.43)

for "structured" uncertainties where

TN = [sI + fP - Arl B(tbr' «KIs + K2) t + tAB)rl

(2.44)
fP - ArlB(tB)-l«Kls + K2)t + tAB)

s is the Laplace Operator

32
Generic Model Control- The Basics Chapter 2

KI and Kl are the GMC reference trajectory parameter values

arranged in diagonal matrices
WI and Wl are weighting matrices
and I'A(.) is the structured singular value.

Equation 2.43 (and hence 2.44) needs to be evaluated over a range of

frequencies of interest "Structured" uncertainties implies that the
uncertainties in the model can be ascribed to specific model components
rather than some overall accuracy statement

Performance of the controller when modelling errors are present is also

important. For any unmeasured disturbance D' that satisfies:

(2.45)

the weighted Integral Square Error of the outputs is guaranteed to be less

than or equal to J where J is:

and 0* (d) ~ 1

where

A represents the structured model uncertainty

and 0'0(.) is the maximum singular value

33
Generic Model Control - The Basics Chapter 2

J is a function of the nominal model, its uncertainty description, the GMC

reference trajectory parameters and two weighting matrices WI and W0- Wo
weights different outputs according to their importance in controlling the
plant. WI ensures that condition 2.45 holds.

Condition 2.46 can be used to determine the 'best' values of the GMC
reference trajectory parameters. It was noted in the previous paragraph that
the robust performance J is dependent on the GMC tuning parameters KI and
Kl. This is obvious for the nominal case when the model is perfect. For
such perfect models, increasing the elements of KI and Kl will result in
improved performance. Increasing KI and Kl can be achieved by increasing
the GMC reference trajectory shape parameters ~j and decreasing the speed
parameters 't j • In the presence of model error, however, excessive demands
placed on KI and Kl may lead to inferior performance, or even instability.
It is clear then, that an optimal performance JoP! can be found which
minimises J with respect to KI and Kl. Le.:

J
opt
= Kl,K2
min J (2.48)

Jop! is a function of the nominal model, its uncertainty description, the input
weighting matrix WI and the output weighting matrix WOo

It is important therefore, that the uncertainty description is representative of

the uncertainty associated with the model. Excessive or overly conservative
descriptions will lead to lower KI and Kl values than are necessary, resulting
in sluggish control when implemented. Conversely, imprecis'e and narrow

34
Generic Model Control - The Basics Chapter 2

uncertainty descriptions will lead to higher KI and Kl values, and this may
result in inferior closed-loop performance. There is a need, therefore, to
develop accurate and reliable methods for quantifying the uncertainty
associated with various types of structural mismatch.

Examples of the use of both the robust stability and performance criteria can
be found in Signal (1992). However, the theory recently developed has not
had any impact of the applications presented in this book, or the applications
described in the next section. In the future, the theory does offer the potential
to guide model development as discussed by Signal (1992).

2.7 APPLICAnONS

GMC has been applied to a range of industrial, pilot-scale and simulated

processes. A partial list of known applications has been compiled and is
shown in the Table 2.1 (Lee, 1991). This table only lists "major" application
studies where the results have either been applied industrially to pilot-plants
or to extensive simulations involving many equations and variables. The
table classifies the applications according to the type of model and process
characteristics. In summary, these applications have all indicated the
practicality of implementing the direct approach for a wide range of
processes, and have demonstrated good control performance against the stated
control objectives.

35
Generic Model Control - The Basics Chapter 2

Table 2.1 GMC Applications

Application Reference Deadtimes Constraints Type of Type of

Present Present Model Used Environment

Anaerobic Costello et al NO NO Steady-state Simulation

Digestion (1989a,1989b,199Oa, and Dynamic
199Ob,1993)
Costello (1990)
Greenfield et al
(1990)

Activated Lee et al (1988) NO NO Dynamic Simulation

Sludge
Plant

Sugar Wilson et al (1988) NO NO Dynamic Pilot Plant

Crystal1izers Wilson (1990)

Crude Zhou and Lee YES YES Dynamic Simulation

Tower (1990)

Evaporator Lee et al (1989) NO NO Dynamic Simulation

Newell and Lee
(1989)

Blast Labossiere (1990) NO NO Steady-state Pilot Plant

Furnace and Dynamic
Stoves

Blast Furnace Zhou et al (1992a) YES NO Steady-state Industrial!

Simulation
Reactors Brown (1989) NO NO Dynamic Industrial!
Simulation
Cott and Macchieto NO NO Dynamic Simulation
(1989)
Howie (1988) NO NO Dynamic Industrial!
Simulation

Distillation Cott et al (1989) NO NO Steady-state Industrial!

Simulation
Fountain (1990) NO NO Steady-state Industrial!
Simulation
Malik (1988) NO NO Steady-state Industrial
Riggs (1989a) NO NO Steady-state Industrial
Sinha .t Riggs NO NO Steady-state Industrial
(1989)

Pneumatic Surgenor (1990) NO NO Dynamic Pilot Plant

Positioner

36
Generic Model Control - The Basics Chapter 2

Application Reference Deadtimcs Constraints Type of Type of

Present Present Model Used Environment
ThennaV Liu et al (1990) NO NO Dynamic Simulation
Hydraulic
Process

Furnace Surgenor and NO NO Dynamic Simulation

Hesketh (1989)

Surge Tank Lee et al (1991) NO YES Dynamic Simulation

2.8 REFERENCES

Balchen J.G., Lie B. and Solberg I. (1988) Internal decoupling in nonlinear

process control. Modelling, Identif and Control 9:137-148
Bartusiak R.D., Georgakis C. and Reilly M. (1989) Nonlinear
feedforwardlfeedback control structures designed by reference system
synthesis. Chem. Eng. Sci. 44:1837-1851.
Bartusiak R.D., Georgakis C. and Reilly MJ. (1988) Designing nonlinear
control structures by reference system synthesis. Proc. ACC, Atlanta,
Ca, Session TPG-4:30: 1585-1590
Brown M.W. (1989) Operational and Control Policies of an Acetylene
Hydrogenation Reactor, M.A.Sc. Thesis, University of Waterloo,
Waterloo, Ontario, Canada
Brown M.W., Lee P.L., Sullivan G.R. and Zhou W. (1989) A Constrained
Nonlinear Multivariable Control Algorithm, AIChE Annual Meeting, San
Francisco, California, November 5-10.

37
Gelleric Model Control - The Basics Chapter 2

Costello D.J., Greenfield P.F. and Lee P.L. (1989a) Cost Effective Operating
Strategies and Control of High-Rate Anaerobic Reactors. Anaerobic
Digestion Workshop, Massey University, New Zealand 13-14
November.
Costello D.J., Lee P.L. and Greenfield P.F. (l989b) Anaerobic Digestion
Control by Generic Model Control. Bioproc. Eng 4:119-122.
Costello D.J. (1990) Modelling, Optimisation and Control of High Rrate
Anaerobic Reactors, PhD Thesis, Department of Chemical Engineering,
The University of Queensland, Queensland 4072, Australia.
Costello D.J., Lee P.L. and Greenfield P.F. (1990a) Application of Generic
Model Control to the pH Control of a Single-Stage High-Rate Anaerobic
Reactor. Int. Assoc. Water Pollution and Control (IA WPRC)
Conference, Kyoto, Japan.
Costello D.J., Greenfield P.F. and Lee, P.L. (1990b) Strategies to Minimise
the Operational Costs of High-Rate Anaerobic Treatment Systems.
Regional Seminar on Management and Utilisation of Agricultural and
Industrial Wastes, Universite Malaya, Kuala Lumpur, Malaya, 21-23
March.
Costello D.J., Lee P.L. and Greenfield P.F. (1993) Strategies for the pH
Control of Two-Stage High-Rate Anaerobic Treatment Plants. Water
Research (submitted).
Cott B.J., Durham R.G., Lee P.L. and Sullivan G.R. (1989) Process model
based engineering. Comp & Chern Eng 13:973-984

38
Generic Model Control - The Basics Chapter 2

Cott BJ. and Macchieto S. (1989) Temperature control of Exothermic batch

reactors using Generic Model Control. Ind Eng Chem Res 28:1177-
1184
Fountain P. (1990) Process Model-Based Control of High Purity Distillation -
An Industrial Case Study, M.A.Sc. Thesis, University of Waterloo,
Waterloo, Ontario, Canada.
Greenfield P.F., Nalbantoglu M., Costello D., Newell R.B. and Lee P.L.
(1990) Improved Operation of Wastewater Treatment Plants: Combined
Roles of Biotechnology and Information Processing. 4th WPCF/JSWA
Joint Technical Seminar on Sewage Treatment Technology, Tokyo, May
17-19.
Henson M.A. and Seborg D.E. (1990) A critique of Differential Geometric
Control Strategies for Process Control. IFAC World Congress, Tallinn,
Estonia
Howie B.A. (1988) Control and Optimization in Catalytic Reforming,
M.A.Sc. Thesis, University of Waterloo, Waterloo, Ontario, Canada.
Isidori A. (1989) Nonlinear Control Systems: An Introduction. 2nd ed
Springer Verlag, New York.
Labossiere G.A. (1990) Modelling and Control of Blast Furnace Stoves.
M.Eng.Sci Thesis, Department of Chemical Engineering, The University
of Queensland, Queensland 4072, Australia.
Lee P.L. (1991) Direct Use of Nonlinear Models, Proc. CPC IV, South Padre
Island, Texas, USA 517-542.
Lee P.L. and Sullivan G.R. (1988) Generic Model Control (GMC). Cornput.
& Chern. Eng 12: 573-580

39
Generic Model Control - The Basics Chapter 2

Lee P.L., Newell RB. and Sullivan G.R (1989) Generic Model Control - A
Case Study. Can J Ch E, 67:478-484
Lee P.L., Zhou W., Cameron I.T., Newell RB. and Sullivan G.R (1991)
Constrained Generic Model Control of a Surge Tank. Computers &
Chern Eng.15,3:191-195
Lee P.L., Zhou W., Newell RB. and Greenfield P.F. (1988) Generic Model
Control of an activated sludge plant. First Regional Seminar on Process
Control, Kuala Lumpur, Malaysia 28-29 November 222-235.
Liu J., Pieper J. and Surgenor B. (1990) Intelligent identification and control
of a thermal/hydraulic process: SISO PSTC and MIMO GMC. Internal
Report. Manuf. Research Corp. of Ontario, Canada
Malik A. (1988) Dual Composition Control of an Industrial Propylene
Splitter. IF AC Symposium on Model Based Process Control, Atlanta
GA, USA
MATLAB (1990) Users Guide. The MathWorks Inc. Natick, MA USA
Newell RB. and Lee P.L. (1989) Applied Process Control - A Case Study.
Prentice Hall, Sydney, Australia
Riggs J.B. (1989a) Nonlinear Process Model-Based Control of a Propylene
Sidestream Draw Column. Ind Eng Chern Res. 29,11:2221-2226
Signal P.O. (1992) GMC Relevant Modelling: Tools for an Improved
Modelling Methodology, PhD Thesis, The University of Queensland,
Queensland 4072, Australia.
Signal P.O. and Lee P.L. (1992) Generic Model Adaptive Control, Chern
Eng Commun. 115:35-52

40
Generic Model Control- The Basics Chapter 2

Sinha R. and Riggs J.. (1989) High Purity Distillation Control Using
Nonlinear Process Model-Base Control. Advances in Instrumentation,
Proc ISAl89 Paper #89-0531, Vol 44, Part 2, 765·772
Stephanopoulos G. (1984) Chemical Process Control· An Introduction to
Theory and Practice. Prentice Hall, Inc., Englewood Cliffs, New Jersey.
Surgenor B.W. and Hesketh T. (1989) Multivariable control of a furnace:
Optimal LQS versus model based GMC. American Control Conf. June
21·23, Pittsburgh, PA, USA.
Surgenor B.W. and Wijesuriya E.T. (1990) Control of a Pneumatic
Positioner: Simulation results for Model-Based versus Conventional
Control. IEEE Indust Automation Conference, June 19·21, Toronto,
Canada
Wilson 0.1. (1990) Advanced control of a batch raw sugar crystallizer. PhD
Thesis, Oeparbnent of Chemical Engineering, The University of
Queensland, Queensland 4072, Australia.
Wilson 0.1., Lee P.L., White E.T. and Newell R.B. (1988) Computer control
of an industrial sugar crystallizer. Process Systems Engineering Conf,
Sydney, Australia 28 August· 2 September 26·31.
Zhou W. and Lee P.L. (1990) Model-based controller design for a heavy oil
fractionator. Control 90, 4th Australian Control Conference, 1-3 August
12·16.
Zhou W., McNamara A.R., Lee P.L., Clark C.C., Lock Lee L. and Burgess
J.M. (1992a) Simulation Studies of Model·Based Controller Design for
an Ironmaking Blast Furnace. CHEMECA 92, Canberra Australia,
September 27·30 199.1·205.1.

41
Generic Model Control - The Basics Chapter 2

Zhou W., Lee P.L. and Sullivan G.R. (1992b) Robust Stability Analysis of
Generic Model Control. Chern Eng Cornrnun. 17:41-72

42
 CHAPTER 3

AN INDUSTRIAL APPUCA TION OF REFERENCE SYSTEM

SYNTHESIS/GENERIC MODEL CONTROL: WASTEWATER pH
CONTROL

3.1 ABSTRACT

A nonlinear controller designed by reference system synthesis was applied to

control the pH of wastewater being discharged from a petrochemical plant.
A first principles model was developed based on a set of simplifying
assumptions, and validated against plant data. The controller was
implemented as a Fortran program in a PMX computer linked to TDC2000.
The results were the essentially complete elimination of the need for manual
intervention when major disturbances arose, and significantly improved
control within the regulatory consent limits for effluent pH.

3.2 INTRODUCTION

The plant was experiencing chronic problems controlling its effluent pH. The
number of incidents where the effluent exceeded the regulatory consent limits
was excessive. Neutralizing agent consumption was higher than necessary
because of controller cycling. This situation existed regardless of a well-
Wastewater pH Control Chapter 3

designed effluent treabnent system. Tuning the pH controllers was a rite of

passage for application engineers assigned to the plant. Despite the efforts of
many good people, no set of controller settings provided satisfactory
performance over all conditions. Instead of trying to tune again, we decided
to implement a nonlinear controller designed by reference system synthesis
(Bartusiak, et al., 1989).

Controller design by reference system synthesis is a three step process. First,

a dynamic model of the process must be developed. Second, the desired
closed loop behavior of the plant is specified in the form of a differential or
integro-differential equation (i.e. the reference system). Third, the control
law is derived by minimizing the difference between the process model and
the reference system. In simple cases, this minimization can be done
analytically which results in a closed form solution for a nonlinear control
law. The Reference System Synthesis method is equivalent to the Generic
Model Control (GMC) method (Lee and Sullivan, 1988).

Reference System Synthesis and Generic Model Control are but two
techniques in an active research area in chemical engineering. Bequette
(1991) has published a review which sets reference system synthesis and
generic model control within the context of this research.

The literature on pH control is itself extensive. A rough categorization of the

methods applied for pH control is the following: non-model based, nonlinear
model based, and adaptive techniques. Shinskey blazed many of the trails in
the pH control field, and his reference (1988) discusses the non-model based

44
Wastewater pH Control Chapter 3

techniques - PID controllers with nonconstant gains, such as error-squared

and three-piece nonlinear. In the same reference, he also discusses the
process design and instrumentation matters for pH control which will be most
helpful to designers and control engineers in indus1Iy.

The controller described in this chapter is an instance of the nonlinear model

based type. Other examples of nonlinear model based pH controllers are
described by Wright and Kravaris (1991), Parrish and Brosilow (1988), and
Williams, et al. (1990). Other nonlinear model based techniques and the
adaptive control approaches for pH control are surveyed in the review by
Gustafsson and Waller (1992).

In this chapter, we provide a process description, and detail the development

and validation of a simple first principles model for pH dynamics. Next, we
discuss the reference system synthesis controller derivation, tuning, and
implementation. Finally, we present plant data which illustrate the nonlinear
controller's performance.

3.3 PROCESS DESCRIPTION

The plant produces approximately 210 metric tonnes per hour (950 gpm) of
wastewater during normal operating conditions. Several of the most noxious
streams entering the effiuent system are treated at-source. However, all
effiuent from the plant is routed together for final treatment. The final
treatment process take place in a concrete-lined pit in which o~l separation,
primary pH adjustment and secondary pH adjustment are applied.

45
Wastewater pH Control Chapter 3

A schematic of the effluent treatment pit is provided in Figure 3.1. In the

figure, flow is from right to left. Rightmost in the schematic are the oil
separators. The "Coarse Pit" designator refers to the first stage of pH
adjustment in which either 96% sulfuric acid or 20% caustic solutions are
added via a compound split range valve arrangement (four valves). After the
Coarse Pit, an equalization basin is provided to dampen major pH swings.
Final pH adjustment is applied in the "Fine Pit" where, again, either 96%
sulfuric acid or 20% caustic solutions are added via a split range valve setup
(two valves). At the extreme left of Figure 3.1 is the pump bay from which
the effluent can be either discharged to the environment, or recycled to a
basin adjacent to the effluent treatment pit if the wastewater is outside of the
consent limits for pH or Total Oxygen Demand.

Duplicate pH meters are provided at the locations indicated in the Coarse and
Fine Pits. A rough estimate of effluent flow rate is measured using a V-
notch weir just upstream of the Coarse Pit. Temperature is measured in the
pipe downstream of the discharge pumps. The neutralizing agents are fed by
gravity from elevated storage tanks through the control valves. There are no
flow measurements on the acid and caustic streams.

The nonlinear controller discussed in this chapter was implemented for the
Fine Pit.

46
Wastewater pH Control Chapter 3

3.4 MODEL DEVEWPMENT

From a study of the various streams entering the eftluent system, we

concluded that the predominant acidlbase species in the wastewater were
derived from sulfuric acid and caustic soda, and that there was little buffering

2tm--~

Equalizallon Basin

Fine Pit

- 96% Sulphuric Acid .•.... 0.01 tonneJhr

-20%Causlic - 0.1 ton neJhr

Figure 3.1 Schematic of the eftluent treatment unit.

47
Wastewaler pH Control Chapter 3

potential. These and other simplifying assumptions, summarized in Table

3.1, were made as the basis for modelling.

Table 3.1 Modelling Assumptions

Modelling Assumptions

1. Sulfuric acid (H2S04) and caustic (NaOH) only acid/base species

present.
2. Insignificant buffering.
3. Complete acid/base dissociation. (K.(HS04") = 1.2xlO"2 but at the
near-neutral conditions of the setpoint, dissociation is practically
complete.)
4. Fine Pit is a constant volume CSTR.
5. Perfect mixing in CSTR.
6. Infinitely fast acid/base equilibria.
7. Linear valve characteristics, such that Flow = a x Output.

The nonlinear dynamic model is derived from the following six conditions.

• Electrical neutrality is maintained.

[H1 + [Na1 =[OHl + 2[SO:] (3.1)

• The water dissociation equilibrium fixes the relative concentration of

hydrogen and hydroxyl ions.

(3.2)

48
Wastewater pH Control Chapter 3

• The fluxes of sodium and sulfate ions through the Fine Pit are described
by material balance.

.!dt lNa1 = [Fi]lNa1- + [FC]lNa 1 _ [FoutJlNa1

V I V C V
(3.3)

• The pH definition relates the instrument reading to the hydrogen ion

concentration.

pH = -log [H1 (3.5)

• The emuent material balance, assumption 4. in Table 3.1, and an
additional assumption that the neutralizing agent flows are negligible
relative to the emuent flow yield the sixth basis condition.

-dV
dt =0 =F. I
+ F + F - Fout
a c
(3.6)

To proceed with the model derivation, we first substitute the water

equilibrium relation 3.2 into the condition of electoneutrality.

(3.7)

Differentiation of equation 3.7 with respect to time yields the hydrogen ion

49
Wastewater pH Control Chapter 3

dynamics in tenns of the sodium and sulfate dynamics.

(~[H+]{l
dt
Kw] =2~[SO·] -~[Na+]
A - [H+]2 dt dt 4
(3.8)

We differentiate the pH definition to relate the hydrogen ion dynamics to pH

dynamics.

(3.9)

Substituting the sodium and sulfate ion material balances 3.3 and 3.4 together
with equation 3.9 into equation 3.8, we derive the pH dynamics in terms of
flow rates and ionic concentrations.

~(PH)
dt
=[ -1 )h(F[SOa + F [SO;] - F[SO;])
2.303(101'H + KwlOPH)V Jl a
I a

-{F[Na1i + Fe[Nal - F[Na1)}

(3.10)
Since the effluent entering and leaving the Fine Pit must also satisfy the
electroneutrality condition, we can relate the ionic concentrations as described
in equations 3.11a and 3.11b.

(3. 11 a)

(3. 11 b)

50
Wastewater pH Control Chapter 3

Finally, by substituting equations 3.11a and 3.11b into 3.10 and simplifying,
we derive an expression for the pH dynamics in terms of values which are
known (Kw, V, [S041., [Na1e), measured (pH, F, T), or estimatable (PHi' F.,
Fe)·

~(PH)
dt
=( . -I
2.303(10"H +KwIOPH)V
1
{F(IO "Hi -K 10pHi) + 2F [so:]
w ••
-

F(IO"H -KwlOpH) - Fc[Na1e}

(3.12)

3.5 MODEL VALIDATION

To validate the model, we first developed estimates for pH;, Fa and Fe from
the available measurements. The pH of the eftluent entering the Fine Pit was
estimated by averaging five values of the Coarse Pit pH measurement delayed
over a window of time centered about 19 minutes (i.e. IS, 17, 19,21 and 23).
We arrived at this simple estimation algorithm via engineering judgement by
(I) matching the deadtime across the Equalization Basin and (2) smoothing to
model the convective/diffusive processes.

The acid and caustic addition rates were inferred from their valve outputs
with the simplifying assumption that the installed characteristics (Luyben,
1990) of the valves were linear equation 3.13. Initial estimates of the Fj,;dent
factors were obtained from a few steady state data points. The factors were
fine-tuned while the model was running on-line to fit equation 3.12 to the

51
Wastewater pH Control Chapter 3

plant data.

F. =F.. Mj,max]
_ x [OPj - OPj,min ]
x (3.13)
J [_ p.J
J.ldenl OP. - OP. .
J,max J,mm
Nominal values for the data of interest in equations 3.12 and equation 3.13
are listed in Table 3.2.

Table 3.2 Nominal values of process data

F 53.0 l/sec Ma,max"p. 0.00281 l/sec

V 17600 I Mc.max"Pc 0.00278 l/sec
[S04-]. 18.0 molesll T 25.0°C
[Na+lc 6.09 molesll

One example of the model fit is illustrated in Figure 3.2. The model tracked
the process quite well, and we proceeded to implement the controller.

3.6 CONTROLLER DESIGN AND TUNING

To design a controller by the Reference System Synthesis/Generic Model

Control method, we take (1) the process model, and (2) a specification for
the desired closed loop behavior of the process (i.e. the reference system),
and then (3) compute the control law by minimizing the difference between
the open loop and reference systems. In our case, the process model is given
by equation 3.12.

The "no frills" specification for the desired closed loop behavior is the

52
Wastewater pH Control Chapter 3

integro-differential equation 3.14 which yields proportional and integral

action in the controller. The coefficients kl and k2 are the tuning parameters
which specify the closed loop speed of response.

d t. (3.14)
-(pH)" =k1 (pH"ot -pH) +k2 J(pHsot -pH)dt
dt 0

The control law is derived by minimizing the difference between the model
and the reference system (i.e. d(PH)/dt - (d(PH)/dt)" = 0).

2 Fa [SO-] _ F C [Na1 F l'(IO~ -K 10PH) -(IOPH. -K lOPH

=y1 i) }
Y 4 a y c w w

+-2.303( 10 .... +K,.1 0"){k,(PH'" -pH) +k,!(PH- -PH)dt}

(3.15)
Note that both acid and caustic flow rates appear on the left hand side of
equation 3.15. These flows are mutually exclusive. Hence, in the controller,
we compute equation 3.15 twice -- once for Fa and once for Fe -- and only
implement the nonnegative value.

Initial values for the tuning constants kJ and k2 were determined by pole
placement. The specification for the closed loop response was that the pH
dynamics should be roughly twice as fast as the residence time of the Fine
Pit, and that the response exhibit critical damping. Consider the closed loop
characteristic equation derived from the reference system equation 3.14.

53
Wastewater pH Control Chapter 3

10·r-------------------------------------------------~1

I
:z::Q,
6

!
I
I
I

4 !
I

I
I
IIV meas model
2

42
Time (hrs)

Figure 3.2 pH model fit to plant data

54
Wastewater pH Control Chapter 3

S2 + k I s + k2 =0 (3.16)

The verbal tuning specification translates to repeated roots of 3.16 at a target

pole twice the distance from the origin in the left half plane as the open loop
pole associated with the residence time.

(3.17)

The residence time was approximately 8 minutes. Therefore, the target pole
was -0.25 min-I, and the initial values of kl and k2 were 0.5 min-I and 0.0625
min-2 respectively.

3.7 IMPLEMENTATION

The controller was implemented in a PMX process control computer above

TDC2000. Initially we coded the model in a BPL program (A BPL is a
procedural program which has the advantage of tight integration with the
process control computer but also has several disadvantages such as only
single precision arithmetic). During model validation, however, we became
concerned about the limited numerical precision by 24 bit word lengths. The
final phases of the modelling and the control algorithm were programmed in
PMX FORTRAN in which double precision floating point arithmetic was
used. The controller runs every fifteen seconds.

55
Wastewater pH Control Chapter 3

The controller was implemented as an application above the existing linear

controller. Accordingly, additional code was written to convert the results of
the control law equation 3.15 into the output of the split range controller
above the control valves. To enable fine-tuning of the model online, we
provided fudge factors Fj,fiIdae which, in essence, adjusted the identification
factors Fj,ideDtand hence the process gains in the model. These fudge factors
allowed us to compensate for the uncertainty in the actual flowrates of the
neutralizing agents relative to the valve manufacturing's maximum raings.

OPj =OPj,min +[ F.p.

J J
1 (OPj,IDIX - OPj,min) (3.18)
Fj,fiIdge Fj,ideot Mj,IDIX
The control application in PMX was built in two levels - feedback-only and
feedforwardlfeedback. The feedback-only control law was equation 3.15 with
nominal constant values for pHi' F and T. This provided the operators with
the capability to switch levels which allowed the controller to degrade
gracefully in th event of the loss of the instrument for pHi' F or T. In fact,
some time after the controller was commissioned, the Equalization Basin 'was
cleaned out after a turnaround. The cleanout changed the fluid flow dynamics
significantly which invalidated the pHi estimate described in the Modelling
section (i.e. greater attenuation of pH disturbances was achieved). The
operators were satisfied running in feedback-only mode, so we have no
incentive to revisit the pHi estimate, or to request that an additional pH
instrument be added at the inlet to the Fine Pit.

Regarding tuning, we modified the kl and k2 tuning parameters away from

their initial theoretical values during commissioning. The motivation was to

56
Wastewater pH Control Chapter 3

increase the proportional action and reduce overshoot The final values of kl
and k2 were 0.67 min-I and 0.02 min-2 respectively.

Finally, several other practical implementation matters deserve brief

comments.
• Reset windup was mitigated by restricting the absolute value of
the integral error to a user-set maximum value.
• The incremental change in the control law result could be
constrained to a user-specified step bound.
• The controller was coded with a provision for a deadband around
the pH setpoint However, there was no need to use this feature
based on the performance we achieved.

3.8 RESULTS

It is instructive to consider the performance of the linear controller before

discussing the nonlinear controller. The general strategy was to control
aggressively in the Coarse Pit and gently in the Fine Pit This strategy was
manifested both in the process design for the neutralization system, and in the
tuning constants of the linear controllers (which were continually being
readjusted by the application engineers). In practice, the Coarse Pit was
often in a limit cycle racing back and forth across the neutral zone after a
load disturbance occurred, while little automatic control was being applied in
the Fine Pit Furthermore, the Fine Pit controller would often windup (even
with anti-windup measures) and, for example, continue to add ac~d when the
pH was consistently low. The operators had to routinely put the Fine Pit

57
Wastewaler pH Control Chapter 3

controller into manual and reset the output to 50% (both acid and caustic
valves closed) to get the controller back "in the ballpark".

pH
12

10

~I
6

4 I
2
V
Row(t/h)
300

250

200

150

100

50
Controller OP (%)
100

80
Manual Intervention
}~;d
60 ~FlnePit

}~".,
40

20

ouo----3~0~~60--~90~~1~20~~1~5~0--1~8~0--~2~10~­
TIme (min)

Figure 3.3 Typical response of linear controller

58
Wastewater pH Control Chapter 3

These phenomena are illustrated in Figure 3.3. At the beginning of the plots,
the Coarse Pit is cycling; the Fine Pit emuent is acidic and being recycled
within the plant. At time 25, the Coarse Pit stops cycling with the emuent in
the alkaline range. Seeing the Fine Pit controller wound up, the operator put
the loop into manual at time 50; reset the output to 50%; and returned the
loop back to automatic. Little regulation is achieved by the Fine Pit
controller during the period from 50 to 240 minutes -- the controller output
does not move much and the pH of the emuent leaving the Fine Pit generally
follows the pHi estimate.

Now let's consider the performance of the nonlinear controller. Figure 3.4
depicts a characteristic load disturbance, i. e. a sharp acid spike. The
nonlinear controller acts vigorously when the pH is far from the neutral
range, and gently when the pH is near-neutral. During the time period 70 to
210, the controller strokes the caustic valve and then the acid valve fully
open as the acid spike comes followed by a surge of alkaline emuent. As
the pH returns to setpoint after time 240, the controller moves are small in
the region where the break in the titration curve occurs. Also note that the
nonlinear controller compensates for the difference in the ionic strength of
the neutralizing agents as evidenced by the sudden change of slope of the
output at time 370.

Despite the vigorous action of the nonlinear controller, it was not prone to
high frequency cycling across the neutral zone. In Figure 3.5, the Coarse Pit
is cycling symmetrically around a pH of 7. The Fine Pit controller continues
to respond actively to the cycling pHj estimate, but does not begin to cycle

59
Wastewater pH Control Chapter 3

itself. Figures 3.4 and 3.5 both illustrate the controller performance with
feedforward action enabled.

pH
12

10

4
I
\/........ pH(in) estimate

2
Flow (t/h)
300

50
Controller OP (%)
100 ...,- Coarse Pit

}~ld
....---:::~'A":"""'\IV'\"""-
80

60

40

20 }eau.•,
o ':'0~3'f:0-=6':-0'-9:::'0:-':-:12=0-=1:-:S=-0:-:18=0-=2:71=-0-="24;':0-=2=70::-=:30=0-=3:-:3=-0-=36!:0:-:3:-!9=0-
Time (min)

Figure 3.4 Nonlinear controller response to acid spike disturbance

60
Wastewater pH Control Chapter 3

Although the nonlinear controller's performance was significantly better than

the linear controller, it did not achieve the desired linearization as expressed
in the reference system equation 3.14. The speed of response when returning
from the alkaline region (e.g. from pH 8.5 to the setpoint at 7.6) was taster
than a return from the acidic region (e.g. from 6.5 to 7.6). Furthermore,
during times of no load disturbances, the nonlinear controller did exhibit a
low frequency limit cycle (period on the order of an hour) of modest
amplitude (PH between 7 and 8 for a setpoint at 7.6). However, from the
perspective of the operators and for environmental compliance, this limit
cycle was inconsequential.

3.9 CONCLUSIONS

The reference system synthesis/generic model control methodology was

effective for designing and implementing an industrial-quality nonlinear
controller for regulating wastewater pH. Compared to the existing linear
controller, the nonlinear controller provided aggressive regulation without
high-frequency cycling through the neutral zone, and essentially eliminated
the need for manual intervention.

The source of most of the problems during model validation and controller
commissioning was the lack of knowledge of the actual flow rates of the
neutralizing agents.

In this application, a simple strong acid/strong base model was sufficient.

Equation 3.12 is probably the least complicated model for pH dynamics one

61
Wastewater pH Control Chapter 3

pH
12

10
..~
," ,
,:
J,,
.
6

4
Coarse Pit
2
Flow (t/h)
300

250

200

150

100

50
Coarse Pit

NVVVW}AC;d
~­
40

20 }caust;c
0~0--~3~0--~6~0--~90~~1~20--~15~0---1~80----2~10---­
lime (min)

Figure 3.5 Nonlinear controller response to cycling in Coarse Pit

could derive. A rigorous model for systems of weak acids and bases with
buffering species will include terms for ionic concentrations whose values are

62
Wastewater pH Control Chapter 3

not readily measurable in industrial situations (Gustafsson and Waller, 1992).

Could we have achieved satisfactory control with non-model based nonlinear

techniques, such as error squared, three-piece nonlinear, or scheduled gains
(Shinskey, 1988)7 I have no doubt that these techniques could have made
significant improvements in performance, as Shinskey has shown, with less
engineering time invested. However, the model based controller provides
unique capabilities for controlling to pH setpoints other than 7, and for
compensating for the differing ionic strengths of the neutralizing agents. The
majority of the engineering time was spent on developing and validating the
pH model, not designing and implementing the controller. Once the
modelling methodology was specified, the engineering work proceeded in a
straightforward manner.

3.10 ACKNOWLEDGEMENT

This work was accomplished with the encouragement and assistance of Ken
Walker, Mark Harrington, Martin Kerlin and Ronnie Millet.

3.11 NOMENCLATURE

F volumetric flow rate (l/sec)

Kw(T) water dissociation constant (moles'IP)
kl,k2 tuning constants in reference system (sec·] and sec· 2 , respectively)
M mass flow rate (glsec)
OP output to valve (%)

63
Wastewater pH Control Chapter 3

s the Laplace operator

t time (sec)
T temperature (0C)
V volume (I)
[] concentration (moles/I)

Greek symbols
a. general constant
p density (g/I)
A. pole in imaginary plane (time-I)

Subscripts
a sulfuric acid stream
c caustic stream
fudge online correction factor
effluent stream into Fine Pit
ident model identification factor
j general stream index; either a for acid or c for caustic
out effluent stream out of Fine Pit

3.12 REFERENCES

Bartusiak R.D., Georgakis C. and Reilly MJ. (1989) Nonlinear

feedforwardlfeedback control structures designed by reference system
synthesis. Chern. Eng. Sci. 44(9):1837-1851.

64
Wastewater pH Control Chapter 3

Bequette B.W. (1991) Nonlinear process control: A review. Ind. Eng.

Chem. Res. 30: 1391-1413
Gustafsson T.K. and Waller K.V. (1992) Nonlinear and adaptive control of
pH. Ind. Eng. Chem. Res. 31:2681-2693.
Lee P.L. and Sullivan G.R. (1988) Generic Model Control (GMC). Comput
Chem. Eng. 12:573-580.
Luyben W.L. (1990) Process Modelling, Simulation and Control for
Chemical Engineers. McGraw-Hill, New York.
Parrish J.R. and Brosilow C.B. (1988) Nonlinear inferential control.
AIChEJ. 34 (4):633-644.
Shinskey F.G. (1988) Process Control Systems: Application, Design and
Tuning. McGraw-Hill, New York.
Williams G.L., Rhinehart R.R. and Riggs J.B. (1990) In-line process-model-
based control of wastewater pH using dual base injection. Ind. Eng.
Chem. Res. 29:1254-1259.
Wright R.A. and Kravaris C. (1991) Nonlinear control of pH processes using
the strong acid equivalent Ind. Eng. Chem. Res. 30: 1561-1572.

6S
 CBAPTER4

USING TRAY-TO-TRAY MODELS FOR DISTILlATION CONTROL

4.1 INTRODUCTION

It has been estimated that there are more than 40,000 distillation columns in
operation in the refining and petrochemical industries in the U.S. (Humphrey
et. al., 1991) which accounts for 95% of all separation process in these
industries. It is further estimated that distillation consumes approximately 3%
of the total U.S. energy consumption. More importantly, distillation
operations almost exclusively are responsible for the purity and unifonnity of
the products for the petrochemical industry.

As a result, distillation control has a significant economic impact on the

refining and chemical industries. Therefore, improved distillation control
across the board would result in major economic improvement. For example,
it has been estimated (Humphrey et. al, 1991) that there is potential for an
average 15% reduction in the energy consumption by distillation in the
refining and chemical industries in the U.S. with improved control.
Specifically, one or more of the following economic benefits can result from
improved distillation control:
Using Tray-To-Tray Models for Distillation Control Chapter 4

Reduced product variability

Reduction in the rate of production of off-specification products
Increased production rates when a column is the bottleneck in
the system
Reduced utility usage
Increase in the yield of a product through an increase in the
average impurity levels while maintaining product specification.

In the refining industry, the most important benefit is the reduced utility
usage, but when a column represents a bottleneck, larger economic savings
are possible.

For the chemical industry, the most important benefit is reduced product
variability. With the advent of product certification procedures (e.g., the ISO
9000 series), a greater emphasis is being placed upon the reduction of
product variabilities. In fact, the consumers of chemical intermediates are
beginning to request tighter and tighter specifications for the uniformity of
their feedstocks. For example, polymer manufacturers are more concerned
with receiving a uniform feedstock than the purity of the feedstock. As a
result, suppliers which may provide on the average a higher purity feedstock
are losing business to suppliers that provide the feedstock with significantly
reduced variability even with higher average impurity levels. This is true
because the polymer manufacturers makes a better product when their
feedstocks are uniform. This phenomena is spreading as, manufacturers
recognize the benefits of reduced variability feedstocks. Distillation is a

68
Using Tray-To-Tray Models/or Distillation Control Chopter4

challenging process control problem due to the inherent nonlinearity of the

process and the severe coupling. Moreover, columns are many times
subjected to large upsets in feed composition and feed flow rate. In addition,
nonstationary behaviour, such as a variation in overall tray efficiency, can
result

A typical distillation column is shown in Figure 4.1. In a typical column, the

liquid level in the reflux drum and the pressure are controlled by simple PI
controllers. The variables that most influence economic performance are the
compositions of the two product streams, one at the top of the column and
one at the bottom of the column. These can be controlled by adjusting
differing combinations of two flowrates, but include reflux ratio, reflux rate,
distillate rate, reboil rate, bottoms product rate.

Table 4.1 summarizes the advantages and disadvantages of PID, model

predictive control (MPC) (Cutler and Ramaker, 1979), artificial neural nets
(Bhat and McAvoy, 1990), and nonlinear process model based control
(nonlinear PMBC) (also known as GMC, Lee and Sullivan, 1988). Note that
while PID and MPC are generic, they both assume a linear picture of the
process. Since nonlinear PMBC uses a nonlinear process model directly for
control, it is a nonlinear controller. Artificial neural nets use empirical
nonlinear models.

Table 4.2 lists the base case conditions for a depropanizer with 39 theoretical
stages producing a high purity overhead product and a moderate purity
bottoms product Table 4.3 lists the process gains for a reboil-reflux (V-Lo)

69
Using Tray-To-Tray Models for Distillation Control Chapter 4

configuration calculated using a 50% relative change in impurity levels. Note

the nonlinearity of the gains evident from the changing process gains for step

DIstIllate

Feed

Bottoms Product

Figure 4.1 Typical Distillation Column

increases and step decreases of the manipulated variables. Also note that each

70
Using Tray-To-Tray Models for Distillation Control Chapter 4

manipulated variable affects each controlled variable, indicating a strong

coupling between both inputs and both outputs.

Table 4.4 lists the average relative changes in the gains from the base case
(Table 4.3) for a feed composition change, an operating pressure change, a
feed rate change, a change in stage efficiency and a change in feed enthalpy.
Since PID and MPC use fixed gain models, they are at an obvious
disadvantage for the control of columns with significant variation in process
gains. As a result, for such cases the PID and MPC controller must be
de-tuned to maintain reliabi1i1y at the expense of performance. The nonlinear
PMBC controller has the potential of being able to predict these various gain
changes if the process changes are measured and the model is sufficiently
accurate.

4.2 DESCRIPTION OF COLUMNS CONSIDERED

Nonlinear PMBC work on several high reflux ratio, low relative volatili1y
columns and a high relative volatili1y, multi-component column are described
in this chapter. Following is an overview description of the columns
considered in this work.

Low Relative Volatility Columns. At this point, we have applied nonlinear

PMBC to four low relative volatili1y columns. While these columns were not
strictly speaking purely binaI)' columns, they are assumed to be binaI)'
columns for purposes of the controller model development. The sum of light
key and heavy keys in the feed ranged from 80 to 90 mole percent. The

71
Using Tray-To-Tray Models for Distillation Control Chapter 4

relative volatility between the light key and heavy key was approximately
equal to 1.2 for each column. Each column had approximately 100 trays
while the reflux ratios employed ranged from 10 to 50. As a result, the open
loop time constants are estimated to range from 2 to 6 hours.

As with most distillation columns, the most difficult upset results from feed
composition upsets. Only one column had an accurate on-line estimate of
feed composition while all but one column had analyzers on both the
overhead and bottom product streams. Composition analyzer delays ranged
from 5 to 10 minutes. Since each of these columns produced a final product,
statistical process quality (SPQ) charts were kept on them using the results
from samples taken on each shift.

Table 4.1 Advantages and Disadvantages of Advanced Control Options

Advantages Disadvantages

PID Availability Nonlinear process

Generic Strong coupling
Well understood Nonstationaxy processes
MPC Based upon empirical Uses linear models
dynamic models Assumes that process is
Generic stationaxy
Provides decoupling Significant identification
Handles constraints directly process
Requires large matrix
operations

72
Using Tray-To-Tray Models for Distillation Control Chapter 4

Neural Uses nonlinear models Needs to be trained for full

Network Provides decoupling ranges of operations
Generic Not well understood
Training is CPU intensive
Nonlinear Understands the nonlinearity Requires significant process
PMBC- Provides nonlinear decoupling knowledge
GMC Feedforward compensation Somewhat "tailor-made"
Can adapt to nonstationary Neglects dynamics except in
changes tuning
Tuned like a PI controller

Table 4.2 Base Case Values of Variables for the Depropanizer

Number of stages = 39
Feed composition:
Ethane = 0.019
Propane = 0.215
Isobutane 0.084
N-Butane = 0.209
N-Pentane = 0.159
N-Hexane = 0.211
Feed rate = 0.849 lbmoles/s
Feed temperature = 165 degrees F
Feed tray location = 18
Column pressure 18.040 atm
Stage efficiency = 100 %
Composition of heavy key (isobutane)
in top product = 0.00364
Composition of light key (propane)
in bottom product = 0.02
Top product withdrawal rate = 0.274 lbmoles/s
Reflux ratio 1.8
Temperature of column bottom = 280 degrees F
Temperature of top tray = 122 degrees F
Reflux temperature = 104.9 degrees F

73
Using Tray-To-Tray Models for Distillation Control Chapter 4

The flow controllers and the level controllers are pneumatic controllers.
Composition controllers (whether PI or nonlinear PMBC) resided in a Perkin-
Elmer control computer and performed as supervisory controllers since they
selected the setpoints for the appropriate pneumatic flow controllers.

For three of the columns, the light and heavy keys were adjacent components
on a relative volatility scale. But for one column, there is a component
between the light and heavy key. As a result, for the latter case the use of a
binary model was quite a bit more challenging.

Table 4.3 Base Case Gains

IlxfAV IlxIAL Ay/AV Ay/AL

increase in -.295 -.529 -.302 -.193

impurity
level
decrease in -.220 -.323 -.179 -.122
impurity
level

x - mole fraction impurity in the bottoms product

y - mole fraction impurity in the distillate product
V - vapor rate (# - moles/sec)
Lo - reflux rate (# - moles/sec)

74
Using Tray-To-Tray Models for Distillation Control Chapter 4

Table 4.4 Relative Change in Process Gain From Base Case (Table 4.3)

Change llx/aV Ilx/M. ay/aV ay/M.

+5 Mole% change in -14% -30% 26% 25%

Heavy Key
1.8 atm increase in -10% -15% -16% -15%
colunrnn pressure
10% increase in 10% -11% -11% -10%
Feedrate
10% decrease in Stage -20% -32% -50% -48%
Efficiency
20°F Increase in Feed -5% -10% -13% -11%
Temperature

Finally, three of the colunrnns have to operate against a condenser duty

constraint at times during the summer months.

mgh Relative Volatility Multicomoonent Columns. At this point, a nonlinear

PMBC controller has been developed for dual composition control on a
depropanizer and has been tested on a detailed dynamic simulator of the
depropanizer. It remains to test the controller on the actual colunrnn.

Table 4.2 contains operating conditions that are typically observed. Note the
multi-component character of the feed to the colunrnn. The column is
equipped with online product analyzers for the overhead and bottoms
products with cycle times of about 6 minutes but does not have a feed
composition analyzer. The manipulated variables are the steam flow to the
reboiler and the external reflux rate while the controlled variables are the

75
Using Tray-To-Tray Models/or Distillation Control Chapter 4

propane concentration in the bottoms product and the iso-butane

concentration in the distillate product. The accumulator level is controlled by
the distillate flow rate and the reboiler level is controlled by the bottoms
product flow rate.

At times during the summer, the column also encounters a condenser duty
constraint.

4.3 CONTROLLER MODELS

Dynamic vs Sready S1a1e Models. With regard to the selection of a controller

model for distillation column control, the first decision that must be made is
whether to use dynamic or steady-state models. The application of GMC or
other nonlinear PMBC controllers (e.g., nonlinear inferential model control,
Economoric et. al. , 1985; nonlinear predictive model control, Parrish and
Brosilow, 1988) to distillation control using a dynamic column model
requires (strictly speaking) measurements of the internal states (i.e., all tray
compositions and liquid flow rates) which is not practical. Moreover, finding
the model inverse of a dynamic distillation model is clearly computationally
expensive. But certain assumptions can be made about the unmeasured states
(e.g., initialized with steady-state results) allowing dynamic distillation
models to be used in a nonlinear PMBC framework. Results reported by
Pandit (1991) showed that while the dynamic model was computationally
much more expensive than the steady-state model, the control results using
both types of models were effectively equivalent.

76
Using Tray-To-Tray Models/or Distillation Control Chapter 4

GMC Control Law using Steady-S1* Models. Cott and Sullivan (1987)
showed that by assuming first order composition dynamics, the GMC control
law can be applied using steady-state model equations. For a distillation
column, target product composition levels, Xu and yII' are calculated as
follows:

(4.1a)

tt
(4.lb)
Y. = Yo + Ku (yap - YJ + ~ ~ (yap - y)dt

Then the values of yII and Xu are used in the steady-state model to calculate
the control action. For example,

[ Xss
YII
1 -+ [ Column
Steady -State ]
Model
-+ [VL ]

Figure 4.2 shows a schematic of the GMC control law. While the PI action
is not exactly a classical PI controller it is quite similar. If one considers the
steady-state model inverse with the process, the resulting system will be
linear on a steady-state basis if the model is perfect. Then the PI-type
controller will have a linear system to control which it might cOJ,ltrol quite
well. Therefore, in the limit of a perfect model this arrangement is expected
to work well as long as neither excessive measurements deadtime nor
excessive inverse action are present.

Note that when the proportional gain (dimensionless) terms (i.e., KlI and K 12)

77
Using Tray-To-Tray Models for Distillation Control Chapter 4

are equal to 1.0, the target values become the setpoint values plus any
integral correction. Under these conditions, the controller is behaving as
largely a feedforward controller with no proportional feedback.

Consider the case for which KlI is equal to 2, x,p is equal to 1%, and Xo is
0.5%. Neglecting the integral term, x.. is 1.5%. This means that if you want
to move from too high a purity (0.5%) to your setpoint of 1%, you
implement control that at steady-state (assuming a perfect model) would yield
a value of x of 1.5%. Of course, we do not leave the process alone until
steady-state. We re-evaluate the control action at each sample time. Since
our control intervals are generally quite small compared to the time constant
of an industrial column, the procedure is effective.

r------------------------------
I
I
I
Proc8Mode1 V
@--
PI xssl x
COOIrOI CoUm
Irw8IS8
.
I
law Yss I
I
L Y
I
I
~-----------------------------

Figure 4.2 Schematic of the GMC control law

In effect, we are saying that the gain prediction and steady-state decoupling
of the steady-state model is more important in distillation control than the

78
Using Tray-To-Tray Models for Distillation Control Chapter 4

dynamic behavior when the control interval is small compared to the process
time constant. It will be discussed later how dynamic differences can be
absorbed by controller tuning and implementation choices.

Criteria for Model Selection. The following factors are identified as

controller model selection criteria:

Accuracy of gain prediction

Accuracy of decoupling and feedforward compensation
Computational efficiency
Computational reliability

Nonlinear Algebraic Models. The first model used in GMC to control a

distillation column was an algebraic design type model (Cott and Sullivan,
1987). In fact, Cott et.a1.(1987) compared several algebraic models for
distillation control and found that the Smith-Brinkley (Smith and Brinkley,
1960) model gave the best overall performance. Cott et. a1. also examined
several other algebraic distillation models including the Jafarey-Douglas-
McAvoy model (Jafarey et. al. 1979) and the Smoker equation (Smoker,
1939).

Algebraic distillation models have the advantage that they are, in general,
simple to use. For example, the modified Smoker equation (Riggs, 1988b)
and the Jafarey-Douglas-McAvoy model can be solved analytically for
control action or for parameterization. For systems for which the
development of a detailed model is not practical, an algebraic distillation

79
Using Tray-To-Tray Models/or Distillation Control Chapter 4

model can be an effective alternative.

Each of the algebraic models assume that the feed tray has the same
composition as the feed. This is rarely the case, and as a result, these models
do not usually accurately predict the effect of feed flow rate and feed
composition changes. In addition, these models assume equimolar overflow
and a constant relative volatility each of which mayor may not be
appropriate to a particular case. In addition, all the algebraic models are
based upon a binary system except the Smith-Brinkley model.

We have also found that the application of algebraic models using the
relative volatility of the system predicts process gains that are significantly
different than the gains exhibited by the process. This can be corrected, at
least for a specific operating case, by adjusting the relative volatility used by
the model to improve the gain prediction.

For the first application of nonlinear PMBC for distillation control, we used a
modified Smoker equation model. We found that the model parameter (the
number of theoretical stages) would change over a relative large range of
±2S% from day to day. (This relatively large model parameter variation is a
direct indication that the model is not very accurate.) The model parameter
variations caused by anellor coupled with feed composition changes caused
reliability problems for this controller. It required regular controller re-tuning
and periodically had to be taken off-control by the operators. The control
performance was much improved but the attention required and the reliability
were less than desirable.

80
Using Tray-To-Tray Models/or Distillation Control Chapter 4

By an analysis of the assumptions that our algebraic model was based upon
and the record of process model/mismatch, we determined that we needed a
more accurate column model. This led us to consider how we could use a
tray-to-tray binary model as our controller model for distillation column
control.

Biruuy Tray-to-Tray Controller Model. A binary tray-to-tray model was

selected because it was expected to overcome the limitations of the algebraic
distillation models. The gain prediction accuracy was expected to be
improved because of the tray-to-tray modelling and because the feed tray
composition is calculated explicitly. It should be noted that the binary model
can be adjusted to consider variations in the relative volatility and to consider
non-equimolar overflow. In fact, Pandit et.al. (1992) presented laboratOlY
results for nonlinear PMBC applied to a methanol/water column that has both
non-equimolar overflow and a factor of six variation in the relative volatility.

In applying the GMC control law, x.. and yIS are determined and the model is
used to calculate the boilup rate and the distillate rate, for example. From a
degree-of-freedom analysis on a binary column (Luyben, 1989), if the feed
composition, feed flow rate, overhead and bottoms products are specified,
there is one degree-of-freedom left. As a result, if the boilup rate is
calculated from the model, the distillate rate can be calculated by the material
balances and then the reflux rate is determined using the equimolar overflow
assumption.

81
Using Tray-To-Tray Models for Distillation Control Chapter 4

The solution procedure for the tray-to-tray binary model is as follows. For a
specified value of the boilup rate, distillate composition (yIS) and bottoms
composition (XI.)' the bottoms and distillate product flow rates are calculated
by simultaneous solution of the overall and component material balance.
Using the equimolar overflow assumption, the reflux rate is equal to the
boilup rate minus the distillate rate for a saturated liquid feed. Now the
component material balances are solved starting with the reboiler and
working up the column to the stage just below the feed stage. That is,
knowing the liquid and vapor rates and the composition on a stage and below
allows for the direct determination of the composition of the liquid on the
stage above. Therefore, starting on the reb oiler, one can calculate the liquid
compositions up the column.

In a similar manner, the material balances can be used to calculate stage

compositions from the top of the column down. That is, knowing the liquid
and vapor rates, and the composition of a stage and the stages above allows
for the direct determination of the composition of the vapor leaving the stage
below. Then using the relative volatility, the liquid composition on the stage
can be determined.

For the proper solution, the composition of an internal stage should be the
same whether calculated from the bottom upward or from the top downward.
Choosing a matching stage near the feed tray results in a well behaved
"black-box" function that can be solved iteratively for the boilup rate using
the secant method (Riggs, 1988a).

82
Using Tray-To-Tray Models/or Distillation Control Chapter 4

The model parameter for the binary model is the stage efficiency. For the
stripping section, the stage efficiency is defined as

For the rectifying section

where
Yi - is the calculated vapor composition of stage i
Yi* - is the equilibrium vapor composition of stage i
~ - is the liquid composition of stage i
1'\ - is the stage efficiency
~* - is the equilibrium liquid composition for stage i

One value of 1'\ is used for all stages. Stage efficiencies have been observed
to vary ±3% over a few days period.

The binary tray-to-tray has proved to be both computationally efficient and

highly reliable. Convergence for a 100 tray column on a Perkin-Elmer
control computer takes consistently less than one second CPU. In addition,
we have demonstrated over 7 column-years experience with this binary
tray-to-tray controller model with an outstanding reliability record.

Multi-Component Tray-to-Tray Controller Model. We recognized that the

binary model would not be useful for certain multi-component columns. This
was due to the multi-component effects upon the composition profiles

83
Using Tray-To-Tray Models/or Distillation Control Chapter 4

through the column and to the multi-component effect upon the vaporlliquid
traffic through the column. First we tried a multi-component model with
fixed relative volatilities. The advantage of this approach is that one does not
have to perform a bubble point calculation. Unfortunately, the gain tests
against a computer aided design package indicated a poor gain prediction
capability. hi addition, the numerical solution procedure was not always
reliable.

Next we went to the more rigorous multi-component steady-state theta

method presented by Holland (1981). In this procedure, there is an inner
loop that uses approximate thermodynamic correlations for a simultaneous
solution of the material and energy balances for all the stages in the column.
The outer loop involves updating the empirical thermodynamic correlations
using the rigorous vaporlliquid calculations.

The inner loop is based upon the theta method (Holland, 1981). In the theta
method, by neglecting the effect of composition on component K-values, the
molar flow rates of each component are determined by the solution of a
tri-diagonal set of linear equations. The value of theta is calculated based
upon satisfying the overall component material balances. Then a correction
factor, which is determined from the value of theta, is applied to all the vapor
and liquid compositions. It was found that the theta correction tended to
over-correct the stage composition estimates particularly for the stages around
the feed stage. A small amount of filtering on the corrected stage
compositions improved convergence speed and convergence reliability.

84
Using Tray-To-Tray Models for Distillation Control Chapter 4

Since the GMC control law provides x.. and y.., the steady-state multi-
component column model must be solved several times (usually 4 to 8 times)
in order to determine the boilup rate that satisfies x's and y...

The model parameter is the Murphree stage efficiency (Van Winkle, 1967).
Some modification of the equations used to implement the theta method is
necessary to include the Murphree stage efficiency. One value of the
Murphree efficiency is applied to all stages and all components.

The multi-component controller model has proven so far to be

computationally reliable and reasonably computationally efficient. We have
checked convergence of the multi-component model over a wide range of
operating conditions and as a part of the controller tested on a dynamic
simulator of a depropanizer and in both cases have found it to be highly
reliable. Controller implementation requires four to eight steady-state
solutions which requires about 5 to 10 seconds CPU on a 33MHz 486 PC.

4.4 IMPLEMENTAnON ISSUES

Dynamic Column Modeling. For each of the nonlinear PMBC distillation

control projects, we have developed a dynamic model of the column and used
the dynamic model to test the nonlinear PMBC controller. This greatly
reduces the amount of on-process development time. That is, a careful job of
dynamic modeling results in a dynamic model that has much of the true
dynamic character of the real process and as a result presents many of the
control challenges the real process offers. In addition, if unexpected

85
Using Tray-To-Tray Models for Distillation Control Chapter 4

problems develop (e.g., identification of an unexpected operational

constraint), the dynamic column model allows one to quickly analyze the
problem, test possible solutions, and arrive at a workable solution. Finally,
the tuning of the controller used on the dynamic simulator provides good
estimates of the tuning parameters for the actual column.

A major decision in the development of a dynamic column simulator is what

type of vaporlliquid equilibrium (VLE) model to use. For the binary model,
the VLE is conveniently represented using a relative volatility. For the
binary cases that we considered, a constant relative volatility was a
reasonable assumption. For the multicomponent model, we used the
Souave-Redlick-Kwong (SRK) (Holland,1981) model. The SRK model is
computationally intensive requiring two iterative searches for each K-value
evaluation. As a result, we found it to be computationally advantageous to
use the KB model (Holland, 1981) for bubble point calculation and to
calculate the KB model parameters using the SRK algorithm. Specifically, KB
models for all stages were reparameterized each 30 seconds of simulation and
the correlation for a tray was parameterized before 30 seconds of simulation
time if the temperature of that tray changed by more than I°C since it was
last parameterized. This provides a major speedup of the dynamic model
resulting in a dynamic model that was 60 times faster than real time for a 39
stage, seven component column model run on a 33Hz 486PC.

The following items pertain to both binary and mUlti-component dynamic

column modeling.

86
Using Tmy-To-Tmy Models/or Distillation Control Chapter 4

Effective molar holdup of each s_e. A tray in a distillation

column is a distributed parameter problem yet it is usually
modelled as a perfectly well mixed vessel. As a result of the
transport time for the liquid through the downcomer and the
transport across the tray, the effective molar holdup will be at
least as large as the holdup of liquid on the tray and the liquid
holdup in the downcomer. The more the flow through the
downcomer and the flow across the tray approaches plug flow,
the larger the effective molar holdup will be. As an upper limit,
one can expect an effective molar holdup as much as twice the
actual sum of the downcomer and tray holdup if perfect plug
flow existed. Since the effective molar holdup has a significant
effect upon the predicted composition dynamics of a dynamic
column simulator, when one underestimates the effective molar
holdup of the tray, the simulator will predict faster composition
dynamics than would be expected on the real process. Also the
molar holdup in the accumulator and the reboiler should be
accurately estimated.
Level control dynamics. The dynamics of level control can have
a very significant effect upon the dynamics of a column. For
example, very loosely tuned level controls can result in
composition dynamics that are inverse acting. As a result, it is
important to accurately model the level controls of a column.
The level indication of 0 to 100 percent on an accumulator may
mean that it is empty when 0 and full at 100 percent or it could
mean that it is 40% full at 0 percent level and 60% full at 100

87
Using Tray-To-Tray Models/or Distillation Control Chapter 4

percent Therefore care should be taken to model the level

control process accurately.
Analyzer delays. The total product composition and feed
composition delays can have a dominant effect upon the dynamic
behavior of a column particularly under feedback control. The
total analyzer delay is the transport delay from where the sample
is taken until it arrives at the analyzer plus the analyzer update
frequency.
Valve dynamics and heat tnmsfer dynamics. For relatively fast
acting columns, valve dynamics and heat transfer dynamics
should be modelled. Even though the time constant is only of
the order of three to five seconds, it can have a significant effect
upon controller tuning results.
Inferendal esdmates of product composidon. When tray
temperatures are going to be used to estimate product
composition, the inferential approach should be implemented on
the simulator as well.
Constraint modeling. The various constraints on two product
distillation columns result in a limitation on the boilup rate. The
trick is to use process data to correlate the constraint to a
limitation in boilup rate. In this manner, constraints can be
directly added to the dynamic column simulator.

Constraint Handling. As indicated earlier, three of the four high reflux ratio
columns experience condenser duty constraints during the hot summer days.
During these periods, the energy input to the column is set in order to

88
Using Tray-To-Tray Models/or Distillation Control Chapter 4

maintain the reflux temperature at a maximum level. Since the energy input
is set independently. the only degree-of-freedom left to manipulate is the
relative production rates of the products. Because the distillate product is a
saleable product with hard constraints on the product purity. the distillate rate
was adjusted to maintain the overhead product purity while the bottom
product purity was allowed to float.

The identification of the constraint is based upon reflux temperature while a

comparison of the constraint control selected energy input with the energy
input calculated from the dual composition control nonlinear PMBC control
is used to determine when one can return to unconstrained control. If care is
not taken when a column is operating near the constraint, it can cycle into
and out of constraint control over a period of time. This problem can be
overcome by making it harder to leave constraint control than to enter it.

Bumpless Transfer and Rate of Change Limi1s. Bumpless transfer allows for
a smooth transition into the nonlinear PMBC controller while rate of change
limits prevent the controller from making excessively large changes in the
manipulated variables. There are at least a couple of ways to approach
bumpless transfer. For the industrial column considered here, the objective
of bumpless transfer is to be able to turn on the nonlinear PMBC controller
during a major upset and have the controller stabilize the process and drive it
to its respective setpoints. Fortunately, the controller models are quite good
the old model parameters are usually sufficiently accurate to allow the
controller to take over "cold". In addition, placing rate of chan,ge limits on
the manipulated variables (e.g., steam flow rate and distillate rate) is a major

89
Using Tray-To-Tray Models for Distillation Control Chapter 4

factor in the stable operation of this approach to bumpless transfer.

When the model parameters are likely to have changed significantly since the
last time it was used, a different approach can be used. For this case, it is
important that the column is relatively stabilized but not necessarily at the
desired setpoints. Then when the controller is turned on, the controller model
is parameterized to match the current operation and the setpoints used by the
controller are set equal to the current measured values of the product
compositions. In this manner, the manipulated variable values remain
unchanged for the first control interval. Then the controller setpoints are
ramped to the actual setpoints at a preselected moderate rate.

Rate-of-change-limits restrict the movement in the manipulated variables.

This is certainly desirable during analyzer failure, erratic behavior of the
analyzer, or if the operator fails to turn off the controller while the analyzer
is being calibrated. In addition, the effects of certain operational problems or
conditions can be negated by the use of rate of change limits.

Overall, well conceived bumpless transfer and rate of change limits can
enhance the reliability of a nonlinear PMBC controller and likewise enhance
its acceptance by operations.

Operator Acceptance. A major factor in any successful advanced process

control project is operator acceptance. No matter how technically successful
a control project may be, it will not be an economic success unless the plant
operators are willing to use it. Even if improved control is demonstrated

90
Using Tray-To-Tray Models for Distillation Control Chapter 4

which subsequently makes their job easier, they may not want to use the
advanced control because of one or more of the following concerns:
They are afraid that advanced computer-based control will cost
them their jobs. After all, a large part of their job is "riding" the
column through tough times and if new controls eliminate or
greatly reduce that need, the company might not need them
anymore.
Improved control will just lead to tighter product specification
which may even result in more difficult work for them.
The controller may make changes that are counter to their
experience. For example, an operator will typically increase the
reflux rate to reduce the impurity in the overhead product
(one-dimensional approach) while the nonlinear PMBC controller
will usually make changes in both the boilup rate and the reflux
rate even if the bottoms product is on specification (a two
dimensional approach). This difference will worry some
operators and cause them to question the reliability of the
controller. In addition, the nonlinear PMBC controller will
typically make sharper and larger changes in the manipulated
variable than the operators are used to.
They do not usually like new ways of doing things. It's human
nature to resist change that is imposed upon you.

In order to overcome these and other problems associated with operator

acceptance, communication lines between the operators and t;he process
control engineers on the project must be opened early and maintained even

91
Using Tray-To-Tray Modelsjor Distillation Control Chapter 4

after the advanced controller has been commissioned and turned over to
operations.

Moreover, it is essential to get as much input from the operators as possible

before the controller is implemented. The operators are a valuable source of
operating history and can provide information about operational objectives,
process constraints, and abnormal operations that is essential to the
formulation of a reliable and effective controller. If the controller is not
designed to handle their real problems, it will be of little use to operations.

It will be much easier to achieve operator acceptance if they feel that they
are part of the project and that it is not being imposed upon them. Asking
for their input before the project begins will help in this area. Also,
explaining the economic incentive of improved control can help. Explaining
in their terms how the controller works will help alleviate some of their fears.
For example, they should be made aware that the controller is likely to make
changes that are sharper and different from what they are used to. They will
not, in general, trust a "black box" controller.

In summary, developing a personal relationship with each operator based

upon communication should enhance your chances of having them feel that
they are a significant part of an advanced control project. Then when
improved control can be demonstrated, it should make their job easier while
allowing them to operate their unit more economically efficient. In a truly
successful control project, the operators, the control engineers, and
management should each feel that they have made a contribution to the

92
Using Tray-To-Tray Models for Distillation Control Chapter 4

project and that the project has been a success from their point of view.
Along this line, it is essential that management, operations, and the control
personnel agree at the beginning to the operational objectives of the project.

4.5 RESULTS

Tuning Procedures. Tuning a nonlinear PMBC controller is similar to field

tuning a PI controller but since the proportional gains are dimensionless,
initial estimates of tuning parameter values are easier to make since we had
experience tuning the controller for the dynamic simulators. By running
simulations, we have found that the greater the analyzer dead-time, the less
proportional gain one should use. Also, in general, the larger the process
time constant, the larger the proportional gain that can be used.

For each industrial case, the nonlinear PMBC controller was tested on a
dynamic simulator of the column. As a result, we have a good idea of the
tuning parameters before application of nonlinear PMBC to the actual column
has been started. From experience we have found that proportional tuning
parameters (Kll and K1J that work best on the actual column are about 20%
lower than the optimal tuning parameters developed on the dynamic
simulator.

For the low relative volatility columns, proportional gain tuning parameters
ranged from 2.0 to 6.0 with the larger time constant process requiring the
larger proportional gains. In each case, the tuning parameters are fine tuned
by observing the measured variables and the manipulated variables action.

93
Using Tray-To-Tray Models for Distillation Control Chapter 4

Fractionator "A" SPQ Data

Before and After Model-Based Control
% Purity of Bottoms
87r-----------------~~------------------~87

86
,
,,

85 ~----------_r~~~~~~=±~----------~ 85
I-.- Before -t)-
\
After - Target I
I

Figure 4.3 Fractionator "A" SPC Data

That is, if the process is slow responding to an upset, then the proportional
gain is increased. Or if ringing in the controlled variable or manipulated
variable is observed, then the proportional gain is decreased.

94
Using Tray-To-Tray Models for Distillation Control Chapter 4

Control Perfonnance for Low Relative Volatility Columns. At this point,

nonlinear PMBC has been applied to four high reflux ratio binary-type
columns with a total of over 7 column-years of total operating experience.
Figures 4.3 and 4.4 show week duration SPQ charts of product uniformity for

Fractionator "8" SPQ Data

Before and After Model-Based Control

% Purity of Distillate

97.6 97.6

97

96.4 96.4
I--- Before + After -
L------r========:::::;------~
Target I

Figure 4.4 Fractionator "B" SPC Data

95
Using Tray-To-Tray Modelsjor Distillation Control Chapter 4

comparison between the nonlinear PMBC controllers and the previous single
loop PI controllers. Each column is responsible for producing one final
product. In each case, the results were chosen to be representative of normal
operation of both controllers. Note that the variability in the products
produced by these columns is reduced by a factor of 2 to 5 compared with
the PI controller results.

Figure 4.5 shows a comparison of on-control time versus off- control time for
nonlinear PMBC controllers and the remaining PI distillation controllers.
Note that the off-control time is reduced by a factor of 4. It should be noted
that the off-control time includes downtimes caused by mechanical failures
and analyzer failures. It has been estimated that the nonlinear PMBC
controllers are being used greater than 95% of the time that they could be
used.

In addition, the previous PI controllers on the four columns considered here

required almost daily tuning and a high level of operator attention while the
nonlinear PMBC controllers are rarely re-tuned after commissioning and
require relatively little operator attention.

Control Penonnance for 1he Multi-Component Column. Figure 4.6 shows the
impurity levels for the top and bottom products for the nonlinear PMBC
applied to the dynamic simulator of the depropanizer for a ramp change in
feed composition. For this case, the mole fraction of the light key in the feed
(propane) was decreased by 5 mole % while at the same ~e the mole
fraction of the heavy key in the feed (iso-butane) was increased by 5 mole %

96
Using Tray-To-Tray Models for Distil/ation Control Chapter -I

Model-Based Used More

Than Other Controls
On Control On Control
88% 47%

Off Control Off Control

12% 53%
Model-Based Other

Figure 4.5 Percentage time on-control

over a 30 minute period.

Figure 4.7 shows the impurity levels for the top and bottom products for a
20% step increase in feed flow rate to the column. The nonlinear PMBC

97
Using Tray-To-Tray Models for Distillation Control Chapter 4

controller produced results on the column simulator for feed flow rate and

2.SE-02
.-.-.-11-.-.-.--11-.-..-..-._,._._._._._.•_._.-l1-li-.-11-
~ 4~--------------------------------------~
~c 1 &
0
2.3E-02 i
j 31
.!!
2.1E-02
~
3
~ ~- ... ...--........,-..---....'*-~~-.l 1
C 2 - ...-----..-...--..-.....-~ ~
1.9E-02 ?

1
3
0
CD
1 1.7E-02
i
Iis 0 L-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ -:-:~-_:_:: 1.SE-02
~
;:)

o ~ ~ 00 ~ 100 1~
Time, mitlutes
I.' Dist. IC4 ....... Bot. c31

Figure 4.6 Model-Based Control with Feed Composition Change

feed composition upsets that are far superior to the control performance of
the existing PI controls on the actual depropanizer.

98
Using Tray-To-Tray Models for Distillation Control Chapter 4

4.6 ECONOMIC BENEFITS

The improved control of each of the four low relative volatility columns has
resulted in significant economic benefit and very short project pay-back
periods.

For the first column that nonlinear PMBC was applied to, approximately 30%
of the product previously produced by this column had been off-specification
thus requiring re-processing. After installation of the nonlinear PMBC
controller, the production of off-specification product was virtually
eliminated. As a result, approximately 30% less feed had to be fed to this
column and the proceding column in order to meet production quotas.
Therefore, the reboiler steam requirements were reduced by 30% resulting in
an energy savings of about $US250,OOO per year. In addition, an increase in
production capacity of 30% is also available if needed. The second and third
columns that had nonlinear PMBC applied to them were in similar service.
For both cases the improved control made it possible to reduce the amount of
over-refluxing while producing a more uniform product. The energy savings
amounted to a 20% reduction in steam usage yielding an annual savings of
about $US450,OOO. In addition, greater processing capacity is also available
as a result of improved control.

The fourth column has been identified as a bottleneck in the production of a

product for which the refiner can sell all that it produces. The improved
control on this column has resulted in a 15% increase in product rate which
corresponds to a $US2,OOO,OOO increase in profits annually.

99
Using Tray-To-Tray Models for Distillation Control Chopter4

~
~-
3
1
..........
. ......... .........................--.........
4r-----------------------------------------,
i .-.-. -
2.5E.Q2

2.3E.Q2
Ii!
i 2.1E-D2
fJ
i 21 1.9E-02
i
o~
i
1 1.7E-D2
J
~ OL---------------------------------------~ 1.5E-D2
o 50 200
100 150
Time, minutes
1·.Dist. IC4 *' Bot. c31

Figure 4.7 Model-Based Control with Feed Rate Change

100
Using Tray-To-Tray Models/or Distillation Control Chapter 4

4.7 CONCLUSIONS

The use of nonlinear models for distillation columns compensates directly for
the inherent nonlinearity of columns and provides accurate decoupling and
feedforward compensation. An accurate controller model can predict the
nonstationary behavior of the process gains as the result of changes in
operating conditions.

As a result, nonlinear PMBC has been successfully applied to the control of

four low relative volatility industrial columns. Factors of 2 to 5 reduction in
the product variability have been demonstrated while the nonlinear PMBC
controllers have a service factor in excess of 95% and require much less
operator and control engineer maintenance and tuning time. Since the
projects reported here require only software changes, the payout periods
have been one month or less in each case.

4.8 REFERENCES

Bhat N. and McAvoy T.1. (1990) Use of Neural Nets for Dynamic Modeling
and Control of Chemical Process Systems. Computer Chem Engr
4:517-532
Cott B.1. and Sullivan O.R. (1987) Process Model Based Control and
Optimization of a Binary Distillation Column, presented at the Spring
National AIChE Meeting, Houston, TX
Cott B.1., Reilly P.M. and Sullivan O.R. (1987) Selection Techniques for
Process Model Based Controller, presented at the Spring National
AIChE Meeting, Houston, TX

101
Using Tray-To-Tray Models for Distillation Control Chapter 4

Cutler C.R. and Ramaker B.L. (1979) Dynamic Matrix Control-A Computer
Control Algorithm. AIChE 86th National Meeting, also in Joint
Automatic Control Conf. Proceed, San Francisco, CA
Economoric C.G., Morari M. and Piasson B.O. (1985) Internal Model
Control 5 Extension to Nonlinear Systems. Ind Eng Chem Process Des
Dev 25:408-418
Holland C.D. (1981) Fundamentals of Multicomponent Distillation, McGraw-
Hill
Humphrey 1L., Seibert A.F. and Koort R.R. (1991) Separation
Technologies-Advances and Priorities, US DOE Report, DOEIl0/12920-
1.
Jafarey A., Douglas 1M. and McAvoy TJ. (1979) Short Cut Techniques for
Distillation Column Design and Control. Ind Eng Chem Process Des
Dev 18:197
Lee P.L. and Sullivan G.R (1988) Generic Model Control. Compo and
Chem. Eng. 12:573-583
Luyben W.L. (1989) Process Modeling Simulation and Control for Chemical
Engineers. 2nd edition, McGraw-Hill
Pandit H.G. (1991) Experimental Demonstration of Nonlinear Model Based
Control Techniques on a Lab-Scale Distillation Column, PhD
Dissertation, Texas Tech University, Lubbock, TX.
Pandit H.G., Rhinehart RR and Riggs J.B. (1992) Demonstration of
Constrained Process Model Based Control of a Nonideal Distillation
Column, Proceed of the 1992 American Control Conf. Chicago, IL.
Parrish J.R and Brosilow C.B. (1988) Nonlinear Inferential Control AIChE
J 34:633

102
Using Tray-To-Tray Models for Distillation Control Chapter 4

Riggs J.B. (1988a) An Introduction to Numerical Methods for Chemical

Engineers. Texas Tech University Press
Riggs J.B. (1988b) Nonlinear Process Model Based Control of a Propylene
Sidestream Draw Column. I&ECR 29.
Smith B.D. and Brinkley W.K. (1960) General Short Cut Equation for
Equilibrium Stage Processes. AIChEJ 6:446
Smoker E.J. (1939) Analytic Detennination of Plates in Fractionating
Columns. Trans AIChE 34:165
Van Winkle M. (1967) Distillation. McGraw-Hill, New York USA

103
 CHAPTERS
AUTOMATIC MOISTURE CONTROL IN PARTICULATE DRYERS

5.1 ABSTRACf

Although drying is one of the oldest and most common unit operations in the
process industries little work has been done on the automatic control of
process dryers. This is largely because of the complex nature of drying
processes and lack of appropriate process control software and hardware
including moisture sensors.

This chapter presents the state-of-the-art of process control of particulate

dryers including the incentives for process control, the required sensors and
computer control hardware and software. A case study of a continuous
horizontal conveyor belt dryer for drying pet food is used as an illustration.

The work included the development of a moisture sensing system capable of

measuring the full range of products produced in a typical pet food plant, the
GMC controller and implementation on an industrial dryer. The control
system manipulated the temperature of the final two drying zones to control
the drying rate, in order to maintain the moisture content at the desired level.
Automatic Moisture Control in Particulate Dryers Chapter 5

The system demonstrated the ability to reduce overdrying associated with

manual control, resulting in a decrease in the energy consumption and an
increase in production yield. The actual economic benefits depend on dryer
characteristics and product throughput.

Conclusions of this case study and many others like it show that, although
there are significant economic benefits from the computer control of dryers,
the non-linearities in the process as well as the difficulty in process
measurements make this a challenging class of processes to control.

5.2 INTRODUCTION

Moisture control during drying is important because it relates directly to the

economic viability of the drying operation. If the product to be dried fails to
meet moisture specifications several consequences may result depending on
the product and the reasons for drying. For example in the case of grains
and food stuffs underdrying may result in spoilage and overdrying results in
increased energy costs and reduced yields. In the case of plastics
underdrying may result in a loss of product performance and will reduce the
processability in subsequent processing steps.

DIY pet food is produced by extruding a mixture of grains, meat and meat
by-products with water. Steam is added to the barrel of the extruder to adjust
the expansion of the pellets as they leave the extruder. This allows the
operators to control the bulk density of the product. This is important in a
manufacturing environment when a certain weight of product must fill a

106
A utomatic Moisture Control in Particulate Dryers Chapter 5

specific sized bag. The wet product is fed evenly onto the conveyor belt by
an oscillating spreader. The belt consists of perforated metal plates which
allow air circulation through the product. The myer normally consists of at
least two heating zones. In a heating zone, air is circulated through the
product layer on the belt. A fraction of the air is vented to the discharge
system. The remaining air flow is mixed with fresh air and passed by a
heater (typically steam coils or a natural gas burner) in order to maintain the
desired air temperature. A centrifugal fan is used to circulate the air through
the layer of product. The last zone in the myer may be used for product
cooling or as a final mying zone. Variations of this design Include multiple
conveyors in the myer which have the effect of increasing the myer residence
time without increasing the myer size.

The extruded pellets have an initial moisture in the range of 20 to 25% w.b.
The moisture level must be lowered below 12% w.b. to ensure that spoilage
will not occur in the finished bagged product. Typical product flows through
a myer are 5 tonnes/hour. This results in an evaporation load of about 0.67
tonnes/hour. These myers typically operate at about 50% efficiency. This
means each myer will require 12.2 GJ/h to meet the energy requirements.
The flowrate of natural gas consumption to provide this energy would be
about 360 m3/h.

A schematic diagram of a conveyor belt pet food myer is shown in Figure

5.1. The moisture content of the outlet product is often difficult to control
because of the many changes that occur to the inlet product and the mying
conditions with time. For example, the moisture content of the inlet may

107
A utomatic Moisture Control in Particulate Dryers ChapterS

vary considerably with time as the operators adjust extruder conditions to

maintain product size, shape and bulk density. The drying rate can also
change with time as well as the ambient conditions such as temperature and
relative humidity. These changes make it difficult to consistently maintain
the outlet moisture content to a specific value.

Previous methods to control moisture relied on operator experience to guide

the operation of dryers because on-line sensors were not available to
continuously measure moisture content On-line moisture sensors are now
available, (Carr-Brion (1986), NIR Systems (1989», and have given the dryer
operator the potential to control the outlet moisture much closer to target
The performance of these systems, however, needs to be examined closely.
For example, it is possible for an operator to actually reduce the performance
of a dryer by the inappropriate adjustment of throughput or temperature in
response to changes in inlet moisture compared to the case where the
operator makes no changes at all. This is also possible with computer control
if care is not taken in the tuning of the controller. Also, because these
sensors are not always accurate, frequent calibration may be required to
maintain reliable comparison with manual reference meters (Reilly et ai,
1988).

Dryers in the pet food industry are normally controlled based on results of
hourly manual samples. Drying air temperatures are then adjusted based on
these results.

The operators control the dryer by taking hourly samples and analysing for

108
A utomalic Moisture Control in Particulate Dryers Chapter 5

Three Zone Conveyor Dryer

Exhaust Air

Recy, ReCY~le 1
Wet Dried
Feed Zone 1 Zone 2 Zone 3 Product

Heated Make-up and Recycled Air

ITIC I Temperature Indicator and Controller

Figure 5.1 Schematic Diagram of a Conveyer Belt Dryer

moisture content in the laboratOlY. The drying air temperatures may then be
adjusted as required. Operators tend to overdry due to a lack of on-line
moisture feedback from the dryer and typically average between 7 and 8%

109
Automatic M oistUTe Control in Particulate Dryers Chapter 5

w.b.

Since operators generally have no on-line feedback from the results of these
changes, a large safety margin of overdrying is usually maintained. By
continuously monitoring the moisture content and automatically controlling
the dryer, this overdrying safety margin can be reduced. This results in
energy savings and an increase in production yield.

The purpose of this work was to develop a control system that could monitor
and control the product moisture on a continuous basis. The system would
reduce overdrying and therefore save energy. The rationale for controlling
dryers is due to the significant fluctuations in moisture content entering the
dryers, and the subsequent affect on the moisture content of the product. In
many cases operators wish to produce a product which has the highest
possible moisture content without violating moisture constraints.

This work involved the development of an on-line moisture sensor and

monitoring dryer operations in order to develop an optimal control strategy.
A computer based system was developed to display process information to
the operator and perform the control calculations. For the test site, this
system demonstrated an energy savings of 5.6% and a production yield
improvement of 1%. This resulted in an economic benefit of SCDN123,600
per year based on the energy savings, yield improvement and continuous
monitoring benefits.

110
A utomatic Moisture Control in Particulate Dryers Chapter 5

5.3 RATIONALE

Pet food dryers are typically poorly instrumented for continuous moisture
control. Operators have access only to drying zone temperature readings as
shown in Figure 5.1. In order to obtain information about the product,
operators must physically obtain a sample and analyse it for moisture content.
Due to this lack of information, operators tend to overdry the product.
Product that is too wet will be rejected because the chance of spoilage is
high. Bulk density, an important product property, is controlled by adjusting
water and steam injection flows in the extruder barrel. This results in
changes in the inlet moisture of the feed to the dryer which create
disturbances in the drying process and another level of uncertainty for the
operator.

An automated control system will provide continuous on-line monitoring of

product moisture from the dryer and will be able to quickly adjust the dryer
operation to compensate for changes in outlet moisture content. The control
system would also display all the measured process parameters. This
provides the operator with information so that the process can be operated
more effectively. Since product moisture levels will be controlled, the
average moisture content can be controlled closer to the maximum allowable
limit. The higher moisture obtained will result in improved product yield,
better product uniformity, and save energy.

Pet food drying was an attractive market to produce a control ,system for
since there are no known computer control systems commercially available

111
Automatic Moisture Control in Particulate Dryers Chapter 5

and the pet food industry in North America has a market value in excess of
seven billion dollars u.s. per year. Several on-line moisture sensors are
available but they tend to be too expensive or do not provide the user with
enough calibration curves required for the large number of products in a pet
food plant.

5.4 RESEARCH AND DEVELOPMENT OF THE CONTROL SYSTEM

5.4.1 Moisture Control

Several dryer manufacturers supply control systems based on product

temperature or exhaust air temperature (Behlen (1986), LAW (1986)). The
operator must select an exhaust air temperature or grain bed temperature
resulting in a desired outlet moisture content. This may be difficult and will
change with grain variety, inlet moisture content and atmospheric conditions.
Success of these controllers is limited since they have no feedforward
capabilities and the relationship between temperature and outlet moisture is
not constant. The relationship between outlet moisture and outlet temperature
is a function of the inlet moisture, inlet air temperature, humidity and
physical properties of the product. In general, the higher the air temperature
is, the dryer the product will be. However, because this relationship is
complex it is difficult to use in a feedback or feedforward controller.

Control systems which monitor both the inlet and outlet moisture levels
should provide superior control of the drying process. A control system
developed by Shivvers, was tested by Michigan State University (Bakker-

112
A utomatic Moisture Control in Particulate Dryers Chapter 5

Arkema et al, 1988). This system measured the moisture content of the inlet
and outlet grain using a discrete sampling system with a sampling period of
3-10 minutes. The control algorithm was a feedforward type with feedback
correction and dynamic compensation in which the throughput rate was
manipulated to control the outlet moisture (Moreira and Bakker-Arkema,
1990).

Model-based control has become an area of much interest in the last decade.
Lee and Sullivan (1988) review the current state of the art in this area as well
as present a new framework for model-based control call Generic Model
Control (GMC). The basic concept behind model-based control is that a
model of the process is used as a guide to controlling the actual process.
With a knowledge of the inlet moisture, outlet moisture target and a model of
the dryer the throughput rate can be adjusted accordingly to control the outlet
moisture.

Figure 5.2 shows the outlet moisture distributions from a pet food conveyor
belt dryer using two different control strategies. In all cases the outlet
moisture was controlled by adjusting the dryer temperature by manipulating
the valve position on the natural gas line to the burner. The moisture content
at the highest frequency is the mean moisture and the spread or width of the
distribution represents the standard deviation or variation in the outlet
moisture. The vertical line at 11.5% indicates the maximum desirable
moisture content. Good moisture control is represented by a narrow
distribution about the mean close to, but lower than 11.5%. Poor control, on
the other hand, is represented by a broad outlet moisture distribution either

113
Automatic Moisture Control in Particulate Dryers Chapter 5

1.6~-r-------------------r-----.

1.4 ··············································'GMC······...........................................
1.2 ................................................ .....................................................

I I I I I I
8 8.5 9 9.5 10 10.5 11.5 12
Moisture, %w.b.

Figure 5.2 Moisture Distributions for Various Control Strategies

above or below the target of 11.5%.

If one assumes that manual control represents the base-line operating

conditions then the standard PID (Proportional Integral Derivative) control
algorithm results in an improvement over manual control. The GMC model-
based controller in turn results in an improvement over PID control. This is
expected because an explicit model of the dIyer is built into the GMC
controller. The model-based control strategy with model update has been
incorporated into the Dryer Master (DM) moisture control system, (Dantec,

114
Automatic Moisture Control in Parliculate Dryers Chapter 5

1992).

With the development of a moisture metering system to measure the trends in

product moisture content, it was then possible to develop a control strategy.
For the dryer in the study, there were four possible manipulated variables; the
conveyor belt speed and the temperature in each of the three drying zones.
The effect of changes to each of the four variables was studied by conducting
open loop tests during dryer operation. The results were used to configure a
simulation of the dryer in order to fully understand the dynamics of the
drying process in this case.

5.4.2 Conveyor Belt Speed

The conveyor belt speed determines the residence time in the dryer and
therefore could be used to control the moisture. During the open loop tests
however, changes in the conveyor belt speed seemed to have unpredictable
results. For example, a relatively small increase in speed may result in a
decrease in moisture content while a larger increase in speed may show an
increase in the moisture content. These observations were simulated using a
rigorous simulation program and the results indicated that as the conveyor
belt is slowly increased, the bed height drops and the air recirculation rate
increases, increasing the drying rate. This increase in drying rate initially
increases proportionally faster than the decrease in residence time of the
product in the dryer. This results in the observed decrease in moisture. A
critical point is reached where the drying rate stops increasing and the
residence time determines the moisture content and the moisture therefore

115
Automatic Moisture Control in Particulate Dryers ChapterS

begins to increase. Similar observations were made for decreases in the

conveyor belt speeds.

Since it would be difficult to determine during actual operation where the

c:byer operation was relative to the two critical points, the conveyor belt speed
is not suitable for accurate moisture control.

5.4.3 Zone 1 Tempenttwe

In the first temperature zone, the product contains mostly free moisture since
it is the c:byer section immediately after the extruder. The c:bying rate in this
zone is determined by the maximum rate of evaporation from the surface and
is not influenced by small changes in zone I temperature. Increases in
temperature large enough to influence the rate would create a c:by crust on the
outside of the particle which would affect further c:bying and processing, such
as fat absorption. Decreases in the temperature of zone I large enough to
slow the evaporation rate will allow the product to cool and therefore slow
the rate of evaporation in the other zones since the water transport
mechanisms to the surface would be slowed until the particles are reheated.

Manipulation of the zone 1 temperature is not suitable for moisture control

and it will be left to the operators to determine a suitable constant operating
temperature, based on previous c:byer operation and the current outlet
moisture reading.

116
Automatic Moisture Control in Particulate Dryers Chapter 5

5.4.4 Zone 2 Temperature

Since the free moisture is largely evaporated in zone 1, the dIying rate in
zone 2 is detennined by the surface evaporation rate and internal moisture
transport mechanisms. The air temperature in zone 2 will detennine the
overall dIying rate since it will control both the surface evaporation rate and
the temperature of the product. The product temperature will directly control
the internal water transport mechanisms.

Although the results of changes to the air temperature of zone 2 can be

predicted the accuracy is not sufficient for control since the moisture content
entering zone 3 can influence the dIying rate in that zone and therefore the
final product moisture content.

5.4.5 Zone 3 Temperature

The dIying rate for zone 3 is detennined in the same way as zone 2. The
effect of changes in zone 3 temperature can be predicted more accurately
than for zone 2 since there is no further processing between zone 3 and the
outlet moisture meter. For example, a 15°C change in zone 3 will result in
about a 1% change in product moisture content.

Since the results of changes in zone 3 resulted in the most easily predictable
change in moisture content, it was decided to manipulate the air temperature
of zone 3 to control the dIying rate in order to achieve the desired moisture
levels.

117
A utomatic Moisture Control in ParliculaJe Dryers ChapterS

5.4.6 Dryer Control Modificdons

In order for the computer control system to manipulate the air temperature of
zone 3, it must be able to communicate with the single loop controllers used
for each temperature zone. Two options are available for such
communications; digital (RS422 or RS232) and analog (current or voltage).
The digital communications protocol will vary among manufacturers of single
loop controllers. Therefore, for this system two current loop communications
were specified. The temperature controller for each zone will send a current
signal proportional to air temperature to the control system. The control
system will in tum send a current signal proportional to air temperature
setpoint as requested by the operator or calculated by the control algorithm
back to the temperature controller. A schematic showing all instrumentation
locations and analog signals is shown in Figure 5.3.

5.5 DEVELOPMENT OF THE CONTROLLER

A DIyer Master system has been developed to control the drying of pet food
in a conveyor dryer. The control system consists of a centralised computer
system which receives incoming data, presents data to the operator, performs
control calculations and sends set points to existing single loop temperature
controllers.

A radio-frequency, capacitance type moisture meter monitors the product

exiting the dryer. A fraction of the total product flow is diverted into a
containment vessel. The moisture meter is mounted in this vessel along with

118
A utomatic Moisture Control in Particulate Dryers ChapterS

Rec~ ReCy~ Re~let

Wet
Feed
I
Zone 1 I Zone 2 IZone 3 ~
Dried
Product

M On-line

~~
Moisture
Sensor

Measured
Disturbance
rr==
I
I~_I
r--- - I
I
I
.
"","< III L Controller
VarIables Model Based -M"
e, P"doo' ---I~~!!!::~~
Conveyor Speed ~ Model Parameter
UPdate

ITlcl Temperature Indicator and Controller

Figure 5.3 P & ID of a Typical Petfood Dtyer

a temperature detecting device. The product flow through the vessel is

controlled by a variable speed positive displacement device (rotary valve,
auger or metering roll) in order to maintain a constant velocity 'through the

119
A utomatic MOisture Control in Particulate Dryers ChapterS

meter. The meter sends an analog signal proportional to the product moisture
content to the Dryer Master.

The general form of the dynamic model used for the dtyer is given by
equation 5.1.

~ = f(d,x,u,t) (5.1)

where: d is a set of design variables

x is a set of state variables
u is a set of manipulated variables
t is time

~ is the predicted outlet moisture of the petfood product

Since dtyers tend to exhibit long process dead time and significant process
response time it was necessaty to add a dead time compensator as shown
below, Lee et al (1990). Equation 5.2 is the predictor corrector term. This
term is filtered by an exponential filter of the process time constant.

(5.2)

A predicted trajectory of the actual controlled variable, M, is given by

equation 5.3.

(5.3)

where M, is the actual measured moisture.

Equation 5.4 calculates the error term based on the desired setpoint, M*, and

120
Automatic Moisture Control in Parliculate Dryers Chapter 5

the predicted process output based on previous control action.

(5.4)

Equation 5.5 calculates the change in the error term from the previous control
interval.

(5.5)

The change in the temperature output, equation 5.6, is calculated by solving

the GMC control equations using the model of dIyer, equation 5.1.

where ~ is a model parameter.

The first two terms of equation 5.6 result from the dIying model, while the
last two terms are derived from the GMC reference trajectory.

The control system is also capable of monitoring other variables in the dIyer
such as product temperatures during and after drying, conveyor belt(s) speed,
product height on the conveyor belt and exit air humidities. All
instrumentation is interfaced to the computer system using a modular I/O
system. This allows the system to expand and include as much
instrumentation as is necessary for further applications. Since specific pet
food products have different bulk density specifications, each product requires
its own moisture calibration curve. The calibrations are stored and updated
on-line by the computer system whenever the operator provides reference
moisture readings. The computer system software has been designed so that

121
Automatic Moisture Control in Parliculate Dryers Chapter 5

its functionality is controlled by a configuration file. This allows software

features to be changed easily or customised for a particular site application.

The display and keypad are contained in a NEMA 4 enclosure which will
allow installation of the Dryer Master display on the plant floor if required.
A photograph of a typical system for a desk mount installation is shown in
Figure 5.4. The operator interface would be the same for the plant floor
installation. All measured parameters can be alarmed and defined as critical

~aDRVER MASTER

Figure 5.4 Dryer Master

to automatic control. This means that parameter values can cause alarms if
out of a nonnal operating range but will not defeat automatic ,control unless
the parameter has been defined as critical to automatic control or proper
dryer operation.

122
A utomatie Moisture Control in Particulate Dryers Chapter 5

The Dryer Master provides two options for obtaining a permanent record of
the drying operation. A printer can be attached to the computer system to
provide a semi-graphical record of dryer operation either continuously or the
last 24 hours on demand. The second option allows communications directly
with the Dryer Master computer using another computer. This will allow
direct access to the dryer operating history for additional manipulation, or
customised analysis and plotting.

In standard pet food applications, the Dryer Master will manipulate the
temperature of the final drying zone to a specified target, and save energy
due to the reduction of overdrying. The control system provides a real time
display of all measured operating parameters. The printed copy of the dryer's
operating history will allow quality control to pinpoint any product not
meeting specifications and allow operating staff to document operations.

5.6 RESULTS

The operation of this dryer has been monitored continuously for a period of
about one year. Because there was no fuel meter on the dryer, decreases in
fuel consumption will be estimates only. The comparison between operating
runs can be difficult since the ability to dry a given product can vary with
factors such as ambient air temperature and relative humidity, efficiency of
the air preheater and how the product is processed before the dryer. For the
purpose of comparison between manual and automated control, two
production periods for the same product will be compared. The product
moisture and dryer temperatures are shown in Figure 5.5 for a manually
controlled production period in June 1989. In July 1989, the same product

123
A utomatic Moisture Control in Particulate Dryers ChapterS

was produced using automated dryer control. The product moisture and dryer
temperatures are shown in Figure S.6.

In comparing control performance, the mean and standard deviation of the

product moisture will be considered. For the manually controlled production
run, the moisture content averaged 9.5% w.b. with a standard deviation of
0.833. This product had a maximum allowable moisture content of 11.5%
w.b. Due to a level of uncertainty in the dryer operation, an average safety
margin of 2% w.b. moisture was maintained by the operator. For the
automated production run, the average moisture content was the same as the
operator specified setpoint of 10.0% w.b. The standard deviation was
reduced to 0.306.

These two production runs also demonstrate the difficulty in comparing

production runs on the basis of energy reduction. During the majority of the
manual run, zones 1 and 2 air temperatures were set at I1S0C and zone 3 at
12SoC. For the automated run, zones 1 and 2 were set at 100°C and zone 3
averaged about lOsoC. During the research phase it was found that a ISoC
change in zone 3 alone accounted for a 1% change in product moisture.
Therefore the full decrease in drying air temperature and the related energy
savings cannot be attributed to the automated moisture control system.

Figure S.7 shows normalised product moisture distributions for both the
manual and automated runs. The manual controlled data shows the large
safety margin of overdrying required to ensure that all the product will be
below I1.S% moisture. The automated moisture control results, in an average

124
A utomatie Moisture Control in Particulate Dryers ChapterS

Moisture, (% wb)
12~------------------------------------~

Maximum Moisture Level

11

10

8 ..................................................................... .

7 ..................................................................... .

10 11 12 1 2 3 4 5 6 7 8 9 10
pm I am
Time
Figure 5.5a Product Moisture Content versus Time, (Manual Control)

moisture level of 10.0% are equal to the operator specified setpoint. With
the improved standard deviation demonstrated here, management found that
they could raise the moisture setpoint 10.5% and still be assured of all the
product having a moisture below the maximum allowable level as shown in
Figure 5.7.

With the demonstrated increase in throughput of 1%, (10.5 - 9.5), from an

average decrease of 15°C in the operating temperature of zone 3 an estimated
savings of 16.6% of the fuel for zone 3 may be realised. Since
approximately one third of the total dryer fuel consumption is in zone 3, the
total energy savings would be about 5.6%.

125
A utomatic Moisture Control in Particulate Dryers Chapter 5

Temperature, (0 F)
130.-------------------------------------,

120 ":J.='"
~"'2h......... ..........".,. ..,:-.,.y.:.... .
'--...
~----\,~
Zone 2
~
"•

---- -- -----------~
110 ....................... /
Zone 3
.................................... : ... .

100 ................................................................... .

90 ................................................................... .

80~------------------------------------~
9 10 11 12 1 2 3 4 5 6 7 8 9 10
pm I am
Time
Figure 5.5b Zone Temperatures versus Time, (Manual Control)

5.7 COST BENEm ANALYSIS

The two easily identified economic benefits resulting from the reduction in
overdIying are energy savings and yield improvements. The potential energy
savings are 5.6% of the total dIyer energy consumption resulting in a
potential energy savings of about $CDN6, nO/year.

The improvement in yield will have a significant benefit for the pet food
producer. The potential of a 1% increase in average moisture will result in
the equivalent increase in yield since the final product is sold by weight. The
increased revenue possible is estimated to be $CDN78,000/year based on a
yearly production of 12,000 tonnes at a value of $CDN650/tonne.

126
Automatic Moisture Control in Particulate Dryers Chapter 5

Moisture, ('Yo wb)

12~--------------------------------------~

Maximum Moisture Level / ' "

11 ..................................................................... .

8 ..................................................................... .

7L-------~------------------------------~
9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4
am 1 pm I. am
Time
Figure 5.6a Product Moisture Content versus Time, (Automatic Control)

There are other benefits to installing a control system which are more
difficult to estimate. The improved ability to monitor the process normally
leads to improved quality through increased operator awareness. In this case,
by continuously monitoring the product moisture levels, product that must be
redried or returned for reworking can be virtually eliminated. These benefits
are estimated at 0.5% of total yearly product value, $CDN39,000. Therefore
the total economic benefit for the site considered in this study is estimated at
$CDN123,nO/year.

The estimated installed cost of the control system in this study is

$CDN68,000. This results in a payback time of about 6 months.

127
A utomatic Moisture Control in Particulate Dryers ChapterS

Temperature, (OF)
130~--------------------------------1

zone 3, Automal8d

120 ............. / ............................................... .

90 ................................................................... .

80L---------------------------~------~
9 1011 12 1 2 3 4 5 6 7 8 9 1011' 12 1 2 3 4
am 1 pm I· am
Time
Figure S.6b Zone Temperatures versus Time, (Automatic Control)

5.S DISCUSSION

The two major benefits of the pet food dryer control system are the energy
savings and yield improvement. These benefits are achieved by reducing the
standard deviation of the product moisture content and then being able to
raise the average moisture content closer to the maximum allowable moisture
content. The cost benefit analysis provided in the previous section was
estimated for a particular drying system. The results of analysis for different
drying systems will depend on production throughput, the number of drying
zones and the overall dryer efficiency. The quality of manual dryer control
before the implementation of an automated control system will also affect the
magnitude of any benefit.

128
A utomatic Moisture Control in Particulate Dryers Chapter 5

1.6'T"""'"-----------------r----,
1.4 •..•....................•.......................•...........•.................................•.....••..
Improvement
1.2 ................................................ .•..• ....• •..•.•. _...............................

12
Moisture, %w.b.

Figure 5.7 Normalised Moisture Distributions (Manual versus Automatic)

The on-line metering of the product moisture will require a period of custom
calibration prior to installation. This work will be required in order to span
the moisture meter electronics for the full range of moistures and product
densities that will be encountered at a given site. During this phase the
temperature correction factors can be found which will reduce the
commissioning time required for moisture meter calibration. Since the
moisture meter will be customised for each application, the accuracy of the
meter and therefore the quali1y of control will be affected if the product
densi1y or temperature fall outside the defined range.

Because the control system can handle any number of products through
specific calibration curves, and the process control functions can be
customised, the dryer performance and product quali1y will be superior to

129
Automatic Moisture Control in Particulate Dryers Chapter 5

other less flexible moisture metering systems.

5.9 CONCLUSIONS

The overall objective of this project has been met; a computer control system
for pet food conveyor dryers was developed. A comparison between manual
and automated dryer control showed a substantial decrease in the standard
deviation of the product moisture content and the potential to increase
production yield by 1%.

Dantec Systems Corporation has gained a great deal of knowledge in the

mechanisms affecting the performance of conveyor dryers, and should be able
to apply the technology developed in this project to other conveyor drying
systems. Considering the cost benefit analysis, the system should have a
payback period of less than one year which would make the system an
attractive addition to a conveyor drying application.

Although the economic incentives for controlling moisture are significant the
non-linear nature of the process as well as the difficult moisture measurement
problems result in a vel)' difficult control problem.

5.10 ACKNOWLEDGEMENTS

We wish to acknowledge the support of the Ontario Government and the

continued interest of Dantec Systems Corporation in this research area.

130
A utomatie M oistUTe Control in Particulate Dryers Chapter 5

The technical assistance of Dr Gerald R. Sullivan in implementing the GMC

model based control technology on a conveyor belt dryer was greatly
appreciated.

S.l1 REFERENCES

Bakker-Arkema F.W., Anderson J.C. and Eltigani A.Y. (1988) Drying Fuel
Cost Control, Grain Age, March 4,8.
Behlen (1986) Owner's Manual, Model 850 Continuous Dryer.
Carr-Brion (1986) Moisture Sensors in Process Control, Elsevier Science
Publishers, New Jersey, USA.
Dantec (1992) Product Literature, DRYER MASTER, Dantec Systems
Corporation, Waterloo, Ontario.
LAW (1986) Dryer operations manual.
Lee P.L. and Sullivan G.R. (1988) Generic Model Control (GMC).
Computers in Chem. Eng. 12(6):573-580.
Lee P.L., Sullivan G.R. and Zhou W. (1990) A New Multivariable Deadtime
Control Algorithm. Chem.Eng.Commun. 91:49-63.
Moreira R.G. and Bakker-Arkema F.W. (1990) Journal of Agricultural
Engineering Research. 45:107-110.
NIR Systems (1989) Product Literature - 550 Spectrophotometer, NIR
Systems Inc., Silver Spring, MD, USA.
Reilly P.M., Sullivan G.R., Whaley M.G. and Fleming J.F. (1988)
Application of EVM for On-line Moisture Measurement in Drying
Processes, Proceedings of Sixth International Drying Symposia,
Versailles, France.

131
 CHAP1ER6

ECCENTRICITY CONTROL OF A CABLE JACKETING LINE

6.1 JNTRODUcnON

In the mid 1980's, an automatic jacket thickness and eccentricity control

system was installed on a number of polyethylene telephone cable jacketing
lines at Northern Telecom (Canada) Kingston Works Plant. At the time, it
was claimed by the supplier that by switching from a manual to an automatic
operation, material savings of 4 to 8 % could be achieved (Brunner and
Merki, 1985). A 10 % material savings on the jacket material provides a one
year payback on the installed cost of the control system.

The value of a jacket control system when performing as designed is widely

acknowledged within the industry (Boggs et ai, 1983). However, for a
number of reasons' the Kingston system has not performed "as designed" on a
reliable basis. The controllers were tuned for very sluggish response due to
perceived stability problems at the time of commissioning. Controller
response time is of the order of 5 to 10 min. This is considered acceptable
under the majority of conditions, where the goal is to prevent long-term drift
in the process. However, the presence of short-term distufbances has
Eccentricity Control of a Cable Jacketing Line Chapter 6

necessitated a thicker than optimal wall thickness to compensate for non-zero

eccentricity. This approach negates any direct economic savings due to
jacket material savings.

This is not to say the system is not considered to be of value. The current
benefit is perceived as being qualitative in nature given that an operator can
leave the line unattended for a period of time without having to wony about
significant drift in the process. However, this paper describes the preliminary
results of a project that set out to determine the achievable quantitative
performance given a properly tuned multivariable controller. The
characteristics of the process were such that a Generic Model Controller
(GMC) was well suited as the candidate multivariable control algorithm.

6.2 THE APPUCATION

The process involves passing a multipair conductor telephone cable core

through an extruder where a cylindrical tube of polyethylene is applied. The
performance goal is to maintain a uniform jacket of a specified thickness
around the cable and along the cable length. The overall process is
illustrated in Figure 6.1. Though the total length of the line is 75 m (250 ft),
the suspended length from the extruder to the end of the cooling trough is
less than 15 m (50 ft). Eccentricity is the difference in opposite wall
thicknesses as is shown in Figure 6.2. Thus, the performance goal can be
restated as the need to maintain zero eccentricity.

Jacket thickness is controlled by manipulating the speed of' the extruder.

134
Eccentricity Control of a Cable Jacketing Line Chapter 6

Cable
Core Motor 0
Extruder Quill
Motors
Capstan
Motor Jacketted
Cable

Ultrasonic
Thickness Cooling Tension Takeup Reel
Payoff Reel Tensioner Extruder Sensors Trough Sensor

Figure 6.1 Illustration of the cable jacketing line

Figure 6.2 Eccentricity of Cable Coating

Eccentricity is controlled by horizontal and vertical positioning of a movable
die (referred to as the quill) in the extruder head. The thickness (and
indirectly the eccentricity) is measured with a pair of ultrasonic 'transducers

135
Eccentricity Control of a Cable Jacketing Line Chapter 6

mounted on opposite sides of the cable as shown in Figure 6.3. The

transducers are rotated between ±90° to get the eccentricity in the horizontal
(X or L-R for Left-Right) and the vertical (Y or T-B for Top-Bottom)
directions.

A series of open loop step tests were conducted to identify a model of the
process. There was a great deal of trouble with excessive noise in the
measurements due primarily to misalignment of the ultrasonic transducers.
One of the few "noise-free" open loop results is given in Figure 6.4. As
indicated in the figure a good fit to the data was obtained with a relatively
simple linear model. Specifically, the simulation result corresponds to a first
order 2-input 2-output linear model:

dy/dt = t-l (G u - y) (6.1)

where the steady state gain matrix is given by:

G =[-{).Oll 0.009] (6.2)

0.009 -{).028
and the matrix of (dominant) time constants is:

t =[ 30o 0 ]
30
(6.3)

and the controlled and manipulated variables are:

u 1 = change in X quill position, %
u2 = change in Y quill position, %
YI = change in X (Left-Right) eccentricity, mm
Y2 = change in Y (Top-Bottom) eccentricity, mm

136
Eccentricity Control of a Cable Jacketing Line Chapter 6

-·-·····---·-··-----~raneduoer

......
.......
" ~\.

D D
\, "

~:~~ ..-................._--

Figure 6.3 Physical Location of Transducers

In reference to Figure 6.4, the fit is good up to the 67 min mark, at which
point a 0.5 mm offset in the Y-eccentricity develops. At this point there was
a SO % step reduction in the line speed that effectively results in a "drooping"
of the cable in the vertical plane. The horizontal or X-eccentricity is seen to
be unaffected. This is considered to be the major disturbance and as such
was used to benchmark controller perfonnance.

The process is predominantly a mechanical one with the positioning system

driven by stepper motors. The first order response is attributed to the
dynamic characteristic of the motors, together with the lag between the
movement of the quill and the movement of the polyethylene. Deadtime (due

137
Eccentricity Control of a Cable Jacketing Line Chapter 6

to backlash in the drive system) was found to be neglible.

The aspects of the system that make control difficult are (in order of
decreasing importance):

1) unless the sensors are properly aligned the high degree of noise in the
measured signal negates the possibility of any effective control,
regardless of the control algorithm in use,
2) there is interaction between the pair of controlled and manipulated
variables, and
3) practically speaking quill moves are limited to between 20 % and 80 %.
The degree of interaction is best illustrated by calculating the Relative
Gain Array (RGA) of Bristol (1966). For the steady gain matrix given
by equation 6.2, the RGA is:

RGA = [
1.36 -0.36] (6.4)
-0.36 1.36

138
Eccentricity Control of Q Cable Jacketing Line Chapter 6

I
.!I
I
oc
....: 1
---.'''''''~" -- ~"M_
.0

~
-
.!I
I
~
: :
eo
T11ft8- 'n
...
"'0 ...., 1-
..
I
I
oc :W
.0
LJ
••
:
• 0

,
TIm_
,
•• ",'n
,
"'0 ....: I
ao

:~ ~
M

~
I
I
~

.0
,
•• - ••
""m_ ",'n
"'0 ....:
Figure 6.4 Open loop test result. Note 1) offset in Y-ecc after 67 min due
aD

to speed change, and 2) controllers turned on at 73 min mark

One notes that the diagonal elements match the natural pairing for the
controllers, namely U1 to YI and U2 to Y2. The non-zero off-diagonal element
(-0.36) quantifies the magnitude and direction of the coupling. The
observation is that moving the X-quill affects both the X and Y-eccentricities,
and moving the Y-quill affects both the X and Y-eccentricities. Thus, for
optimal performance one would expect to require a formal multivariable
controller that coordinates the moves of the X and Y-quills. The controller
designs considered, together with summarized performance results, will be
given in the next section.

139
Eccentricity Control of a Cable Jacketing Line Chapter 6

One final comment needs to be made about Figure 6.4. At the 73 min mark
the X and Y-quill controllers were turned on. One observes a very gradual
return of the eccentricities to zero. For an initial eccentricity of 0.1 mm, the
time to return to 0 mm is seen to be on the order of 10 min with the
controller gains at their original "sluggish" settings.

6.3 THE CONTROLLERS

The original configuration for eccentricity control involved two Proportional

Integral (PI) controllers:

(6.5)

(6.6)

where Kpx and Kpy are the proportional gains and Kil< and Kiy are the integral
gains. The feedforward (FF) gains Kfx and Kfy were not originally
implemented by the plant. However, given the results of the RGA analysis,
some form of decoupling compensation appears to be called for.

The closed loop models formed by the above PI controllers in combination

with the process models of equation 6.1 are second order. On this basis, the
gains can be calculated to provide a specified overshoot and settling time in
response to a change in setpoint for the individual loops. Simulation tests
indicated that a specification for a 20 % overshoot with a 60 sec settling time
provided adequate regulatory action, given the restriction of keeping quill

140
Eccentricity Control of a Cable Jacketing Line Chapter 6

moves between 20 and 80 % for the benchmark disturbance. This required

that Kpx = 68, K" = 28, ~x = 15 and ~y = 6. For ideal feedforward action,
the proportional gains combine with the steady-steady process gains to give
Kfx = 0.8 and Key = 0.3.

As a fonnal multivariable algorithm, a Generic Model Controller (GMC) was

adopted. Details of the GMC approach can be found in Lee and Sullivan
(1988). For a linear first order model of the fonn given by equation 6.1, the
GMC control law is:

U =0< (. [~'Y -K,!Y $] 'y)'u. (6.7)

where Uo gives the initial positions of the quill. By comparison with

equations 6.5 and 6.6 one notes that GMC is essentially a PI controller with
fonnal decoupling provided by the inverse of the steady-state gain matrix. In
a fashion similar to that adopted for the PI design, the gains K} and K2 are
selected to provide a desired trajectOIy in response to a setpoint change. In
this application it was found the GMC gains were best taken as:

(6.8)

(6.9)

For servo action, this corresponds to an underdamped response with a 20%

overshoot and tr as the rise time. This design trajectory was found to provide
adequate integral action for the regulation problem encountered in this
application. In addition, it matched the design approach used' for the PI

141
Eccentricity Control of a Cable Jacketing Line Chapter 6

controller and consequently would provide a basis for fair comparison.

Simulation tests indicated that the minimum rise time was 10 sec if quill
moves were to be restricted to between 20 and 80 % for the benchmark
disturbance. Thus, the adopted GMC gains were Kl = diag[O.1 0.1] and K2 =
diag[O.OI 0.01].

Hanlware Details

The quill positions are measured by LVDT's mounted on the quill itself. The
wall thicknesses are measured by 5 MHz spherical ultrasonic tranducers
mounted on a rotating rocking head placed in a water bath. The controller is
a Texas Instruments (TI) 560 PLC operating with a 1 sec sampling time.
The TI PLC is capable of controlling 64 PID loops with full programming
capability in a BASIC-like language. Data acquisition is performed by a
microV AX 4000 operating in a supervisory role and operating with a 5 sec
sampling time.

6.4 RESULTS

A series of simulation tests were conducted to assist in the design of the PI

and GMC controllers, as well as to determine their relative performance prior
to actual implementation. Given experience with the open loop field tests, it
was felt that the benchmark disturbance would be the equivalent of +0.5 and
-0.5 mm near steps in the V-eccentricity at 5 min intervals.

142
Eccentricity Control of a Cable Jacketing Line Chapter 6

Perfonnanee Goal

The performance goal is benchmarked against 25 mm nominal diameter cable

with a wall thickness of 1.4 mm. The acceptable tolerance for eccentricity
(thickness) is ±O.05 mm (±3.5 %). The sensors are technically capable of
measuring to ±O.003 mm (±O.l thou). In open loop, the observed "normal"
noise was ±0.01 mm (see Figure 6.4).

The major disturbance is a ±SO % near step change in the line speed that
occurs whenever the operator has to work on the line as it runs (for example,
a tape changeover and weld operation). Minor disturbances include: 1) x-v
movement of the cable due to changes in tension, 2) X-v movement of the
sensor heads as they are continuously rotated between ±90°, and 3) minor
fluctuations in cable speed. It was felt that regulation on the order of ±0.02
mm (:=ttl thou) was achievable and this was set as the performance goal.

Simulation Resul1s

The nature of the benchmark disturbance, together with individual results for
a PI and GMC test, is indicated in Figures 6.5 and 6.6. The horizontal lines
on the eccentricity plots correspond to the ±0.02 mm performance goal. In
both figures, the "blips" in the V-eccentricity correspond to the timing of the
disturbance. In Figure 6.5, one notes that although the quill moves stay
within the 20 and 80 % guideline, both the X and Y-eccentricity exceed the
±0.02 mm goal. On the other hand, response is quite rapid with recovery in
under 1 min. Figure 6.6 illustrates that the GMC response is even more

143
Eccentricity Control of a Cable Jacketing Line Chapter 6

rapid. In addition, effective decoupling keeps the X-eccentricity within the

±0.02 mm goal.

Table 6.1 summarizes the simulation results for the PI (Kr gains zero), PI+FF
(Kr gains non-zero) and GMC designs. The Integrated Absolute Errors (JAE)
of the X and· Y eccentricities are given as quantitative performance measures
for a 20 min test. Each design was tested against a "good" and "poor"
model. For the good model, the process gains were those given by equation
6.2 and match those used in the actual design of the controllers.

For the poor model, the process gains were given as:

=[-0.011 0.009] (6.10)

G 0.004 -0.053
with the altered gains underlined. The performance with the poor model
indicates the degree of robustness of the controller designs. It was noted that
in certain open loop field tests the process gains for the Y-eccentricity could
range by the amount given by equation 6.10.

The second last column of Table 6.1 gives the total IAE score normalized to
the "worst" PI performance. The last column states whether the X (L-R)
eccentricity was maintained between the ±0.02 mm goal. One notes that PI
performance actually improves with the poor model, mainly because the
effect of the gains given by equation 6.10 is to reduce the degree of
interaction. For this same reason, the addition of feedforward action with the
PI+FF controller does not significantly improve overall performance. By

144
Eccentricity Control of a Cable Jacketing Line Chapter 6

contrast, the GMC performance is consistently better by a factor of ~2.5 (0.28

versus 0.69).

Table 6.1. Summary of simulation results.

Control Model IAEx IAEy IAEx+y Normal L-ROK

PI good 9.4 21.1 30.5 1.00 no

poor 3.8 17.1 21.0 0.69 yes

PI+FF good 5.0 19.2 24.3 0.80 yes

poor 5.6 17.1 22.8 0.75 no

GMC good 0.0 8.3 8.3 0.27 yes

poor 1.8 6.7 8.5 0.28 yes

6.5 CONCLUSIONS

The simulation results indicated that both the PI+FF and GMC designs could
provide significantly improved performance over the existing (badly tuned) PI
controller. The process model used to properly design the PI+FF controller
was also required as the basis for the GMC algorithm. Thus, the design
effort for the controllers was judged to be the same. On the other band, it
appeared that for the benchmark disturbance, GMC could provide a factor of
two improvement in regulation. On this basis, the GMC design was
implemented on the line with the least noise. Figure 6.7 gives a GMC closed
loop test result with a speed change.

145
Eccentricity Control of a Cable Jacketing Line Chapter 6

i] o
~
• .. •
~-
• 10
: ~-'
,. ,..
:<r:
,. ,.
I
80

!] v: : 1\,: : f\: : V":

Tim_ ,..1"

I
o a 4 • • 10 '2 ,.. ,. ,. 20
Tim_ ""In

f:E;; o a ..
I~:
• •
Tim_
:'=: '-
10
,"'n
,a ,.. ,. ,.
I
20

1':5: , /: : ~ : >,
o 2 ... • •
TI",,_
'0
'"'"
'2 ,... ,. 1.
I
20

Figure 6.S PI simulation result with "good" design. Note 20 to 80% quill
range, and both X and Y-ecc exceed +0.02 mm goal.

146
Eccentricity Control of a Cable Jacketing Line Chapter 6

i ]~,=======~~~~~~~~=====t~I
·0~--~.--~4~--~.--~.--~'0~~'~.--~'4~~'~.--~,~.~~.o

i:I
TI..... '"'It

i] y: o :I
:
..
~
•
:
• 10
TIm_ ",1"
: k:
1.
:
,... ,. ,. ao

• 10 ,. ,... ,. ,. 80

1':5; : I
TIm_ ml"

J : : r' : 'L :
o 2. ... • • 10 1a ,... ,. ,. ao
Tim_ min

Figure 6.6 GMC simulation result with "poor" design. Note 20 to 80%
quill range, and only Y-ecc exceeds +0.02 mm goal.

The X-eccentricity is maintained within the :1:0.02 mm goal and the

Y-eccentricity briefly exceeds the goal at the time of the speed change. On
this basis GMC is judged to be successful. Further field trials are being
conducted before a decision is taken to reduce the jacket thickness setpoint.
This has to be coordinated with a revised speed control scheme to mitigate
the severity of the Y-eccentricity disturbance. The overall result should be
an improvement in the performance of the eccentricity control system, in
terms of both its reliability and its ability to deliver the minimum optimal
jacket thickness.

147
Eccentricity Control of a Cable Jacketing Line Chapter 6

i~E,. :
,
,

,.: ... .. : .. -I
-,
TI .... ...."•• ••
I o.a~
.!!
I
o 2.

,.
:
:~ ,
: :5 ~

aD .a
TI .... ...."
II. II. II• lIO

':~~ ~
M
3
I... ,. ,.

:,. ..:
....
liD aa
TI .... "
II. •• a. lIO

'}~ :~
~
M
3 -F
I
~

,. ao
TI.... ...."
II. a. a.
1
lIO

Figure 6.7 GMC closed loop plant test result with speed change at 20 min.
Note only Y-ecc exceeds +0.02 due to speed change.

6.6 ACKNOWLEDGEMENTS

This work was supported by a Grant from Northern Telecom Canada

Limited. The assistance of Joe Oresti and personnel from the
Communication Cable Division (Kingston Works) is gratefully acknowledged.

The technical assistance of Reza Neshat during his work term at Queen's
University is also acknowledged.

148
Eccentricity Control of a Cable Jacketing Line Chapter 6

6.7 REFERENCES

Boggs L.M., Isley A.M. and Levenhood J.W. (1983) Ultrasonic Jacket
Thickness and Eccentricity Monitor and Control System. Proceedings
32nd International Wire and Cable Symposium, Atlanta 359-362
Bristol E.H. (1966) On a New Measure of Interaction for Multivariable
Process Control. IEEE Trans. Auto Control. Vol.AC-ll, 1:133-134
Brunner M. and Merki H.A. (1985) On-Line Monitoring and Control of
Jacket Thickness and of Eccentricity. Zumbach Electronics Corp., Mt.
Kisco, NY.
Lee P.L. and Sullivan G.R. (1988) Generic Model Control (GMC). Compo
and Chem. Eng. 12, 6:573-580

149
 CHAPTER 7

THE DEVELOPMENT OF A NONLINEAR ADAPTIVE GENERIC

MODEL CONTROLLER FOR CHEMICAL REACTION QUALIfY
CONTROL

7.1 ABSTRACT

The purpose of this paper is to outline the evolution of a model-based control

strategy to control the output quality variable of a chemical reaction process.
Achieving high performance control of this quality variable was a challenging
task, as the process has many characteristics that make traditional control
techniques difficult to apply successfully. These characteristics include
severe nonlinearities, slow and time - variant process dynamics, and
infrequent and noisy process measurements. In the end, the loop was
successfully closed by incorporating a nonlinear process model into a
modified Generic Model Control (GMC) framework. An adaptive algorithm
was used to update a single parameter in the nonlinear model in order to
ensure model prediction accuracy. Although the control strategy required
approximately 2 man-years of effort before it performed up to our
expectations, the strategy now yields economic benefits in the range of SCDN
2,OOO,OOO/year by reducing off-spec product.
Adoptive GMC for Chemical Reaction Quality Control Chapter 7

7.2 INTRODUcnON

The use of computer-based control for chemical processes has rapidly

expanded in the last ten years. The use of linear model-based controllers
such as Dynamic Matrix Control and Internal Model Control has permitted
the control engineer to solve multivariable control problems that would have
been impossible to solve in the early 1980's. Further extensions of these
technologies now allows the incorporation of process constraints and
optimization concerns into the controller design, giving the control engineer
much more flexibility.

Yet, there are still many processes for which high performance process
control is difficult These are processes with strong input-output
nonlinearities, and complex, time-variant process dynamics. In chemical
systems, these characteristics are usually found in and around the reaction
systems.

Until recently, we have not had the control technology available to deal with
these problems. However with the development of practical nonlinear control
algorithms in the late 1980's, we can now begin to address these problems.

This paper describes the development of one such control strategy, which
uses the nonlinear form of Generic Model Control algorithm as its basis. Our
reason for attempting to close this loop is purely an economic one: our
customers are demanding products with smaller and smaller variances in
quality so that they may optimize their production systems. In· addition, to

152
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

Shell Canada, any reductions in off-spec production translates into increased

production capacity.

There are many ways in which the development of the control strategy could
be presented. I have chosen a chronological approach because I feel it best
describes the challenges and hurdles that must be overcome in successfully
closing such a loop.

7.3 CHRONOLOGICAL DEVEWPMENT OF THE CONTROL

STRATEGY

7.3.1 ....y 1990 - The Process

Due to the proprietary nature of the process under examination, we cannot

discuss the process in detail, other than to say that we are concerned with
controlling a product quality variable resulting from a chemical reaction
process.

The process has several characteristics which make high performance control
difficult.

1) The process is highly nonlinear with respect to the manipulated

and disturbance variables.
2) The process dynamics between the manipulated and controlled
variables are both long and complex. There exists approximately
two hours of deadtime and a dominant time constant of about 4

153
Adaptive GMC for Chemical Reaction Qualify Control Chapter 7

hours in the process. In addition, the dynamics of the process

vruy with the throughput of the plant, and the throughput can
change frequently.
3) The measurement of the controlled variable is infrequent, only
once an hour, due to the testing procedure. In addition, the
measurement is subject to a fairly high level of noise about 10%
of its mean value.

The only positive process characteristic is that because there is only one
controlled variable and one manipulated variable, the system can be designed
as a single input / single output (SISO) system.

The process performance of the quality variable has traditionally been

reviewed on an informal basis, but increasing customer specifications in 1989
and early 1990 forced a formal performance review. At this time, the quality
variable was controlled by operator intervention assisted by a manual
Statistical Process Control (SPC) system.

Evaluation of the process performance during this period clearly

demonstrated that the process was not performing up to our expectations.
Figures 7.1 and 7.2 present time series and autocorrelation plots of the typical
process performance. The autocorrelation of the time series is presented so
the controller performance monitoring techniques of Harris (1989) can be
used. Harris indicates that a process that is achieving minimum variance
control will not have any significant autocorrelations past the number of lags
equal to the deadtime of the process. Given that the process has two hours

154
Adaptive GMCfor Chemical Reaction Quality Control Chapter 7

of deadtime, or two lags, it can be seen that the process is nowhere near
minimum variance control, and there is much room for improvement.

1.4

f r\ A
I ' ~v'vt7 ""e
1.2

~~A,,j') \~ lr
I O.B

~ .
o 50 100 150 200
Sample Number

Figure 7.1 Product Quality Time Series for July 1990

7.3.2 August 1990 - The Model

As with all process control, good control performance begins with a good
process model to describe the key process relationships. Due to the
importance of the quality variable on plant operation, model development was
a joint effort between the Process Control group and the Process Engineering
group. This made the introduction of the controller into the process operation
much easier as we now had two groups supporting its implementation.

155
Adaptive GMC for Chemical Reaction QUality Control Chapter 7

r--
r--
I--
0.5

h
•c
t
j II U-J
LLt
0

.i
-0.5

-1
o 10 15 20
Lag

Figure 7.2 Product Quality Autocorrelation for July 1990

After a survey of existing work, three models of increasing complexity were

available for use:
1) A relatively simple nonlinear regression model;
2) A model based on the simplification of a complex differential
equation model;
3) A complex differential equation model based on detailed mass
balances around the reaction system.

After a detailed review of each model, we finally selected the second model.
Our reasoning was as follows: we wanted the model to be based on theory,
which eliminated the first model, while at the same time, we wanted a model
based on a minimum number of inputs. The third model would require the
use of many process inputs whose instrumentation was considered as
secondary instrumentation, and not up to the same standards as other key

156
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

process variables.

Our intention was to have the process model run on-line as frequently as
possible, using the infrequent, off-line data to check the validity of the model.
From our experience, we knew that the process was subjected to many and
frequent unmeasured disturbances. Therefore, a key question was how were
we going to handle the process/model mismatch. Do we use an adaptive
method in order to update an appropriate model parameter, or do we let the
controller contain the mismatch error?

In the end, we decided on an adaptive approach. Our main reason was that
we had to sell the effectiveness of the model to our operations staff, so we
had to have our model match the process output One specific model
parameter of model #2 was chosen to be updated. Initially, the values of all
other model parameters were set from literature values.

The adaptation algorithm used was exponential data weighting with a variable
forgetting factor (Goodwin and Sin (1984». This technique was chosen due
to the previous success of the technique in real-world adaptive control
problems (Dumont. (1982». For the SISO case, implementation of the
adaptation algorithm was straightforward. Unfortunately, there was still much
trial and error tuning of the adaptation algorithm to be done, as will be
presented later.

The adaptation of the model occurs by having the operations staff enter the
off-line quality measurement into a special program on the computer system,

157
Adaptive GMC for Chemical Reaction QUality Control Chapter 7

which retrieves the model output for the same time period and runs the
adaptation algorithm. Clearly, the success of the controller is based on the
timely and frequent entry of the off-line measurement. Fortunately, the
operations staff quickly realized this fact, and the off-line entry had not been
a problem.

7.3.3 October 1990 - The Controller

The implementation of the control strategy was aided by the fact that the
process model obtained using model #2 could be partitioned into two sets of
equations. The first and very small set describes the effect of the manipulated
variable on the controlled variable. The standard Generic Model Control
algorithm could be directly applied to this set of equations. The second set
of equations was only required for the off-line update algorithm. They
generate the model output at the same point where the off-line quality
measurement is made. The entire system, the model, the controller, and the
adaptation algorithm, were all implemented in FORTRAN. No special
numerical routines were required, other than a 4th-order Runge-Kutta method
to solve the ordinary differential equations present in the full model.

Given the high performance requirements of the controller, the GMC tuning
constants were set very aggressively with an ~ of 3.0 and a t of 2.0 time
units. The loop was first closed in October 1990. Figures 7.3 and 7.4 give
the initial performance of the controller.

As it can be seen, the initial performance of the controller has improved over

158
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

manual operation, as the overall variance of the quality has been reduced.
This can be seen by comparing Figure 7.4 to Figure 7.2. Note, though, that
the autocorrelations are still significant after lag 2, indicating that we have
not yet achieved minimum variance control.

1.4

~
11
a
1i
]
..,.J!a..
~ O.B

0.6

0 50 100 150 200

Sample Number

Figure 7.3 Product Quality Time Series for October 1990

7.3.4 November 1990 - Tuning of 1he GMC Controller

Unfortunately, less than three weeks later, the following performance as

shown in Figures 7.5 and 7.6 was observed with the same controller.

Note that the autocorrelations for November do not die out very quickly.
This indicates that the process has some very slow dynamics which are
dominating the overall performance of the control strategy. Several reasons
for this poor performance were hypothesized:

159
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

-
I--
I--
0.5

.
n--n--.
I--
1-0
c

I
0

J ~J IIIlJ
-0.5

-1
o 5 10 15 20
Lag

Figure 7.4 Product Quality Autocorrelation for October 1990

1) overly-aggressive GMC tuning

2) poor adaptation tuning
3) severe process/model mismatch

It took us a great deal of time to resolve all these issues. We first addressed
the easiest topic: GMC tuning. The tuning of the GMC controller was found
to be very aggressive, given the prospect of plant/model mismatch. The
controller as initially tuned often created a severe overshoot of the quality
variable to get onto setpoint quickly. The problem with this approach is that
the quality variable really does not blend; a blend of material above setpoint
with material below does not have the same quality as material produced at
setpoint.

To this end, we revamped the structure of our GMC controller. Firstly,

160
Adaptive GMC for Chemical Reaction QUality Control Chapter 7

f
I
I
0.6

o 50 100 150 200

Sample Number

Figure 7.S Product Quality Time Series for November 1990

r--
-
0.5
- r--
I-.

r Il--h
t· -0.5
~J IIIlJ

-1
o 5 10 15 20
La;

Figure 7.6 Product Quality Autocorrelation for November 1990

examination of the control moves indicated that the integral teon of the GMC
control algorithm was not contributing anything significant to the controller

161
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

output So, our first change was to eliminate the integral tenn, thereby
eliminating one tuning constant As simultaneously proven by Kozub and
MacGregor (1990), the integral term is not required when the process model
is adapted in order to eliminate any process / model mismatch. Kozub and
MacGregor's modification yields a single tunable constant, the closed loop
time constant:

(YIP -y)/tc =dy/dt (7.1)

Secondly, we found difficulties with the standard GMC structure in that it

forced the same closed-loop time constant on the controlled variable, no
matter what the current process dynamics were. At half the plant throughput,
the control actions doubled to keep the dynamics at the same closed loop
rate. This also caused an overshoot of the quality variable. We modified
equation 7.1 to overcome this. Assuming that the derivative of the controlled
variable with time has first-order dynamics, we can rewrite equation 7.1 as:

(YIP -y)/tc =(f(y,x,u,t,d) -y)/tp (7.2)

where f(y,x,u,t,d) is the forcing function describing the nonlinear steady state
behavior of the process. Now, solving for f(y ,x,t,d) yields:

t
f(y,x,u,t,d) =..!(YIP -y) +y (7.3)
tc

Equation 7.3 shows that the control moves made by the modified GMC
controller are dependent on the ratio of tp to tc' We chose to replace the
closed-loop time constant, t c' in equation 7.3 with:

(7.4)

162
Adoptive GMC for Chemical Reaction Qualtty Control Chapter 7

Here, a is a tuning constant relating the closed-loop time constant to the

open-loop time constant By setting a equal to one, the steady-state
controller move to make the quality variable equal to its setpoint is enforced
at all times. We have found this modification and an a value equal to 1 has
steadied the control dramatically by minimizing the controller's desire to
overshoot the final setpoint in order to get the process reacting faster.

7.3.5 .ime 1991 - Tuning of the Adaptation Algoridun

The GMC algorithm changes were implemented at the end of November,

1990, yet we still had occasional problems with the quality variable cycling
with a frequency of about 24 hours. This was not resolved until June 1991
when we first used power spectrum techniques to analyze the controller
performance. Figures 7.7 through 7.8 present the time series, autocorrelation
and power spectrum for the June 1991 dataset

Note the power spectrum shows that the quality variable is only really
affected at cycle times greater than 5 hours. The then-current tuning of the
adaptation algorithm assumed much higher frequency disturbances in the
quality variable. The minimum exponential filter initially used was 2 hours.
Therefore, the adaptation algorithm was permitting high frequency noise
(between 2 and 5 hours per cycle) to pass into the updated parameter and
therefore into the manipulated variable moves. This, coupled with the known
amount of process model/mismatch, caused the plant to cycle. Retuning the
adaptation algorithm so that the minimum exponential filter used was 4 hours

163
Adaptive GMC for Chemical Reaction QUality Control Chapter 7

1.4

:l?
D
<5
1i
l
1
Ji 0.8

0.6

0 50 100 150 200

Sample Number

Figure 7.7 Product Quality Time Series for June 1991

-
r--
I--

.
0.5
...,
~
c:
0
'a

1
.i!
0
LU
-0.5

-1
o 5 10 15 20
Lag

Figure 7.8 Product Quality Autocorrelation for June 1991

further steadied the control perfonnance and eliminated any low frequency
cycling.

164
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

p"" - X Power Spectral Denelty

10-1~--------~--__----~--__~--~--~~--~----~--~

10~L---~----~--~----~--~~--~----~--~----~--~
o 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fr-equency

Figure 7.9 Product Quality Power Spectrum for June 1991

7.3.6 Janwuy 1992 - Model Validation

Even with this retuning, our desired level of perfonnance had still not been
achieved. In July 1991, we felt it necessary to do some closed-loop plant
tests in order to verify the accuracy of our process models. Several key
disturbance variables were identified and pseudo-random binary signal
(PRBS) tests of these variables were run in the plant under closed-loop
conditions. We then used time series analysis to check to see if there were
any unmodelled effects in our model predictions.

We found several effects that were not fully accounted for by finding cross
correlations between the key disturbance variables perturbed and the model
residuals. After careful analysis, we summarized the reasons for prediction

165
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

errors as follows, assuming the structure of our model was essentially correct:

1) poor model parameter values

2) measurement biases

We expected some problems of the first type because, for some effects, our
literature search revealed several possible parameter values. The PRBS test
results confirmed the appropriate value.

The effect of the second problem on the quality of model predictions took
much longer to understand. The issue here is what happens to the model
predictions when the value of the variables in the model are set from
instrumentation systems which yield biased signals. An example is useful
here.

Assume there exists a chemical reaction system in which the reaction rate
may be increased by the addition of a reaction promotion agent, P.
Furthermore, assume that the following model describes the theoretical
behavior of the promoter:

r overall =rbase +k[P]2 (7.5)

where rovcrall is the overall reaction rate, rbase is the reaction rate with no
promoter present, and [P] is the measured concentration of the promoter in
the reaction phase. Finally, assume the values of rbaso and k are known from
laboratory experiments. Can this equation describing the promoter behavior
be added directly into a process model?

166
Adoptive GMC for Chemical Reaction Quality Control Chapter 7

It all depends on the accuracy of the measurement of the promoter

concentration. If the measurement is exact, then as expected, there will not
be problems. But what if the measurement has a bias that makes the actual
concentration higher than what is measured? Clearly, the process model will
consistently underpredict the reaction rate, forcing any adaptation routine to
account for the process/model mismatch whenever the promoter concentration
is varied. This can result in poor controller performance if the promoter
effect is large compared to the base reaction rate, or if the promoter
concentration is raised to a high value where the nonlinearity begins to grow.

We suspected after awhile that these effects were present in our model. The
first attempt we made to overcome this problem was to refit the value of the
adjustable parameter in the model, here in the example, k. Clearly, given
measured values of the overall reaction rate and the promoter concentration,
we can recompute the value of k to ensure the correct value of the overall
reaction rate at a particular promoter concentration given the biases of the
measurements. But all this does is to produce a local fit of the underlying
process in which the point value is correct but the slope of the function is
not. This is due to the fact that the promoter model still does not account for
the concentration measurement bias.

Clearly, it is necessary to add adjustable parameters into the model in order

to account for these bias problems. This goes against the philosophy often
seen in linear model development of minimizing the number of adjustable
parameters. However, where linear models require only the slope of a
function to be specified, nonlinear models require both the slope and value to

167
Adaptive GMC for Chemica! Reaction Quality Control Chapter 7

be correct. The reaction rate model in equation 7.5 could be modified to the
following form:

roverau = rbue + k * ([P] - [P]bil.)2 (7.6)

in order to permit better model predictions. The value of [P]bias can be

computed from a PRBS test on the promoter concentration. Our chosen
model had two major effects that required the addition of bias parameters.

The effects of correctly tuning the adaptation algorithm and accounting for
measurement bias on the control performance can be seen in Figures 7.10 and
7.11, which display the controller performance for January 1992. We are
now at minimum variance control, which means that the controller is moving
the manipulated variable as hard as it can. At minimum variance, further
improvements in controller performance will only be realized by reducing the
disturbance variances.

7.3.7 Mareh 1992 - Economic Benefi1s

By March 1992, we had checked and verified the majority of effects in the
model, and felt a full review of the controller performance was necessary.
Up to this point, the controller performance had been evaluated by analyzing
the quality of the line samples. While it is indicative of performance, the
plant measures its performance in terms of on-spec and off-spec product
batches. Batches of product are made by segregating and blending every 4
hours of plant production. Final analysis of the product is then done on a

168
Adoptive GMC for Chemical Reaction Quality Control Chapter 7

1.4

0.6

o 50 100 150 200

Sample Number

Figure 7.10 Product Quality Time Series for January 1992

composite sample of the batch.

For each month since January 1990, the percentage of on-spec batches to all
batches produced in that month was computed. These percentages are plotted
in Figure 7.12.

It is clear that significant improvements have been made in the plant

performance since the beginning of the project The first improvement
occurred when the controller was first put on-line in October, 1990, when the
percentage on-spec increased on average from 75% to 80%. The second
improvement occurred when the adaptation mechanism was tuned properly in
July 1990. Then the performance moved up to 90% on-spec.

This improvement in on-spec percentage translates into less off-spec material

169
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

- -

n-
O.S

•c:
0
11 r-r--1-r--,
1
r- h r--1
0
'-
jL-lLJ
~

-i
-0.5

-1
o 5 10 15 20
Lag

Figure 7.11 Product Quality Autocorrelation for January 1992

that must be sold at a lower price. Another way to look at the improvement
in on-spec percentage is that it decreases the production time required on the
process to make a given amount of product.

Given the differential in selling price between on-spec and off-spec product
and the plant throughput, this improvement in performance is worth in the
region of $CDN2,000,000/year, which more than pays for the 2 man-years
development time.

7.3.8 Process Monitoring

From our model validation work, we began to realize that we could recover a
great deal of information about our process/model mismatch by analyzing the
time series of our adaptive model parameter. For example, it had always

170
Adaptive GMC for Chemical Reaction Qualify Control Chapter 7

0.95

2co. 0.9
1/1
I
c
0
c O.BS
:§::>
O.B
e
~

Q.

'a 0.75
,§u
0.7
~
0.65

0.8
0 5 10 15 20 25 30

Month. Since January 1990

0.95

1
1/1
0.9
I
c5 0.B5
c
:B
g I

\j
Co trcf'r\nnn ..
O.B
1
r-/\
V
Q.

V
'a 0.75
c
:8u /
0.7
~
0.65

0.6
0 5 10 15 20 25 30
Monthe Sine .. January 1990

Figure 7.12 Long-Tenn Product Quality Control Perfonnance

been suspected that changing levels of impurities in the reactant streams

could cause the final product quality to change dramatically. We were finally
able to prove this by correlating changes in the update parameter to changes
in reactant impurities.

171
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

This analysis, done off-line initially, proved so successful that we decided to

automate the analysis. An on-line program was developed which uses a
Statistical Process Control moving range analysis to detect changes in the
variance of the adaptive model parameter. If a statistically significant change
has occurred, the current process dynamics are estimated to arrive at a time
window in which the cause of the significant event should have occurred.
Significant changes in process variables during this time window, such as
changes in feedstream sources, setpoints and disturbances, are then identified
and logged.

The program has been most useful. In one instance, it confirmed our doubts
about the validity of a specific static model parameter. Whenever a
disturbance moved the average value of the variable associated with the
parameter, we would get a significant event report. In another instance, the
report helped us to identifY an effect we were not initially aware of.

It is clear that this type of analysis would only be necessary for monitoring
high performance control loops. But for loops with significant economics
such as this one, this type of analysis continues to pay for itself.

7.4 SUMMARY

It can be seen that the journey from original concept to actual achievement of
the desired performance level has been a long and often difficult one. If one
reason had to be singled out for the long development time, it would simply
be the lack of experience in implementing such strategies in an industrial

172
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

environment and on such a complex process. There are few published results
on such issues as nonlinear model selection, propagation of measurement
errors and biases through nonlinear models, and tuning rules for adaptive
controllers of this form.

Having said that, we at Shell Canada believe the results were worth the
effort. Certainly, the economics benefits of the control strategy have paid
back the time and money expended in its development. Knowing that we are
now at minimum variance control has changed our philosophy when
investigating poor process performance. Before we would assume that our
control moves were not appropriate for the disturbance. Now, we know that
an increase in quality variable variance is due to an increase in disturbance
variable variances. As a result, we have become very sophisticated in our
knowledge of the disturbance variables affecting the product quality variable.
We now know what our main unmeasured disturbances are, how often and
how large these disturbances are, and how they affect the product quality
variable. Weare now in a position to prioritize these effects for
incorporation into our on-line model.

7.5 ISSUES IN NONLINEAR ADAPTIVE CONTROL

There are many issues around the development of such systems that still need
to be addressed. In summary, I would like to present the following
comments:
I) The key issue in any nonlinear controller development is the
choice of process model, and not necessary the choice of

173
Adaptive GMC for Chemical Reaction QUality Control Chapter 7

controller algorithm. Given that the model structure is

appropriate for GMC implementation, I have no concerns about
the GMC algorithm itself. It has proven to be stable, robust and
easily implemented.
2) Process model selection is a key issue. One must strike a
balance between detail and accuracy in the model and the quality
of instrumentation that will provide the inputs into the model. In
order to capture important effects in the process, the
instrumentation measuring the effects must be good.
3) We have had great success in collecting any process / model
mismatch in an adjustable model parameter rather than letting the
error accumulate in a controller integral term. Adapting a model
parameter permits constant monitoring of the model performance,
and can suggest areas in which the model may be improved.
4) On-line implementation of nonlinear models will often require
the addition of extra static parameters in order to properly
account for measurement biases and errors.
5) Assuming adaptive methods are used, the GMC algorithm can be
modified by eliminating the integral term. This simplifies the
tuning of the controller; the control engineer need only now
specify the closed loop time constant of the system.
6) While the GMC tuning is often straightforward, tuning of the
adaptive parameter update can be difficult Although the theory
is well developed, we lack documented real-world experience in
tuning these systems. We found that the use of power spectrum
techniques greatly assisted in the tuning of the exponential data

174
Adaptive GMC for Chemical Reaction Quality Control Chapter 7

weighting algorithm.
7) For key effects in the model, we have had great success with
performing closed-loop PRBS experiments on the process to
verifY model parameter values. Often, the data quality of
happenstance process upsets is not sufficient for this task.
8) For any high performance control loop, the controller
performance should be constantly measured in order to permit
comparisons before and after changes to the control system.
New monitoring techniques such as the autocorrelation method
of Harris (1989) provide mathematically strong indicators that
can be used along with company-specific performance measures.

7.6 REFERENCES

Dumont G.A. (1982) Self-tuning Control of a Chip Refiner Motor Load.

Automatica Vol 18, 13:307.
Goodwin G.C. and Sin K.S. (1984) Adaptive Filtering, Prediction and
Control. Prentice-Halllnc., Englewood Cliffs, New Jersey.
Harris TJ. (1989) Assessment of Control Loop Performance. Can J Chem
Eng 67:856
Kozub D.J. and MacGregor J.F. (1990) Feedback Control of Polymer
Quality in Semi-Batch Copolymerization Reactors. McMaster Advanced
Control Consortium Report, McMaster University, Hamilton, Canada

175
 CHAPTERS
BLAST FURNACE STOVE CONTROL

8.1 INTRODUCTION

Control in the minerals and metals processing industries presents many

challenges to modern control algorithms. These processes have often been
ignored by control engineers, and yet the incentives to improve the operation
of these plants are quite large. This case study examines one such
application.

8.1 BLAST FURNACE STOVES

8.1.1 Physical Description of Full-Scale Sroves

In steel production, iron ore is first reduced to iron in a blast furnace. The
solid ingredients of iron ore, coke and limestone, are poured into the top of
the furnace and enter the tuyeres region, into which large volumes of high
temperature, high pressure blast air are blown. The coke bums in this hot air
blast, generating temperatures exceeding 2000°C and producing carbon
monoxide gas. At such high temperatures the coke and gas react with the
Blast Fumace Stove Control Chopter8

OHEQUERWORIC

HOT
.LMT

OO...UftION +--
AIR

MaT.

==>
~
eM
L ~ ~(
eM ---+ .""
I
(.) On-G. . Ph. . . (b) On-BI. .t Ph•••

Figure 8.1 Blast furnace stove operation; a, on-gas phase; b, on-blast phase

iron oxide in the ore, reducing it to molten iron.

To ensure efficient furnace operation, the required flow-rate and temperature

of the blast air must be maintained. Typical industrial flows are 3000-10000
Normal m3min-1 at temperatures of 1000-1300°C, depending on the size and
type of furnace· (Beets et al 1977). This hot air provides up to 40% of the
blast furnace sensible heat requirement (Nose et al 1984). It in tum utilizes
approximately 30% of the total fuel consumption in the steel making process
(Mitter et al 1981). From both operational and economic points of view

178
Blast Fumoce Stove Control Chopter8

therefore, a reliable and controllable blast air delivery system is required.

The hot blast requirement is met through use of regenerative heat exchangers
known as blast furnace stoves. These units consist essentially of a

TO FURNACE

STOVE 1 STOVE 2

COLD BLOW HOT BLOW

(ON-BLAST) (ON-GAS)
(a)

F~A~E~
~ ~
(STOVE 1)
LHOTBLOW
i )
o P 2P TIME

F~A~El ~
(STOVE 2)

o
LHcrrn~\~
P 2P
) TIME
(b)
Figure 8.2 Bypass main stove configuration for two stoves

179
Blast Fumoce Stove Control Chapter 8

chequerwork of refractory brick, which performs the heat exchange function

of the stove. As a regenerative heat exchanger, a stove's operation consists
of two distinct phases: a heating phase to store heat in the stove, followed
by a cooling phase, in which the stored heat is released to another fluid.

During the heating phase, often termed the 'on-gas' phase (Figure 8. 1a),
combustion gas is burned with air in the combustion chamber of the stove.
The hot gases are forced upwards and then down through the chequerwork of
refractory brick. The amount of heat stored during the 'on-gas' phase
depends on the heating period, and the rate and efficiency of firing, was well
as the mass, surface area and rate of heat absorption of the brick. When a
stove has stored sufficient heat, the gas supply is stopped.

During the subsequent cooling phase, often termed the 'on-blast' phase
(Figure 8.Ib), cold compressed air is blown through the chequerwork in the
reverse direction. The heat stored in the refractory bricks is then released to
the air passing through them. In the 'bypass main' stove configuration
(Figure 8.2) this hot air is mixed with cold bypass air to give the temperature
and flow-rate required by the blast furnace.

As a stove cools, the amount of bypass air required decreases. The stove
remains on-blast until it can no longer deliver blast air at the required
temperature and flow-rate. At such time, changeover is made to a freshly
heated stove, and the cooled stove is then reheated. In a two-stove bypass
main system, one stove is being heated (ie. on-gas) while the second is
providing blast air (ie. on-blast) (Figure 8.2).

180
Blast Furnace Stove Control Chapter 8

8.2.2 Laboratory Scale Stove System

The laboratory-scale blast furnace stove system used in this investigation is

shown schematically in Figure 8.3. It utilized two stoves in the bypass main
mode of operation.

The brass shells of the two stoves provided the thermal capacitance for the
system. A network of copper tubing and two-way solenoid valves supplied
hot on-gas air to the on-gas stove, and cold on-blast air to the on-blast stove.
The solenoid valves worked in pairs and were switched to allow changeover
of stove duties between on-gas and on-blast roles.

The stove exit temperatures were measured by three semiconductor sensor

probes. These probes were located at the stack outlet (T I)' the junction of the
on-blast stove exits (T2)' and the blast outlet, downstream of the bypass air
mixing point (T3)'

Supply air was regulated at 280 kPa gauge and distributed through a
manifold to provide compressed air for both the process on-gas/on-blast air
requirements and for pneumatic valve actuation. Air delivery to the on-blast
stove was regulated by a control valve, termed the 'main valve'. A second
control valve, called the 'bypass valve', controlled the bypass air flow.
Together the two valves regulated the overall volumetric flow-rate and
temperature of the blast air mixture. Total blast air and on-blast stove air
flow-rates were measured by two differential pressure cells. Maximum on-
blast and on-gas air flow-rates of 4000-5000 I h· 1 could be supplied.

181
Blast Furnace Stove Control Chapter 8

The hot air for the on-gas heating phase was delivered by an electric air
heater, rather than a combustive system as in the industrial case. (In this
investigation the control of the combustion process was not considered.) The

BLAST AIR

ON-QAS EXHAUST

BYPASS
l--~hAIR ,...--11

ON-GASUNE

p STOVE 1 STOVE 2 QN.GAS

e e
VAJ..VE
ROll ME ERS

S2

SUPPLY AIR

MANIFOLD

Figure 8.3 Blast furnace stove rig

182
Blmt Furnace Stove Control Chapter 8

heat input to the on-gas stove was varied by manipulation of the hot air
temperature. A local feedback control loop on the heater allowed control of
the heated air within the range of 0-100°C. Heat input to the on-gas stove
can also be varied through manipulating the tlowrate. A control valve was
installed in the on-gas air line to allow regulation of the air tlowrate. In all
subsequent studies it was decided to only use the on-gas air tlowrate as the
manipulated input, and hold the on-gas air temperature constant. This is
consistent with the aim of maximizing the thermal efficiency of the stove
operation.

The temperature and tlow-rate ranges of the experimental rig were well
below the 1300°C and 3000-10000 m3min-1 blast of an industrial stove
system. In addition, the cold air in the on-blast phase tlows co-currently with
the hot on-gas air, rather than cooling the stove by a reverse tlow of air, as in
the industrial system. The experimental system was nonetheless found to
reproduce the basic behaviour of the larger industrial system.

8.2.3 Control Objectives, Measuremen1s and Manipulared Variables

The primary objective of a blast stove control strategy is to maintain the

prescribed blast temperature and tlow demands, while maintaining as high an
overall thennal efficiency as possible. The major controlled variables are
therefore the hot blast temperature and tlow-rate. The associated manipulated
variables are the main and bypass tlow-rates in the on-blast phase, the inlet
gas flow-rate and temperature in the on-gas phase, and the period of the
operating cycle.

183
Blast Fumace Stove Control Chapter 8

The stove control task may be divided into two general areas: short and long
tenn control goals. Manipulation of the bypass and on-blast stove flow-rates
detennine the mixed temperature and flow-rate of the resulting blast air. This
represents a short tenn control goal. Short tenn control is made difficult by
significant process interactions, non-linearities, and time-varying behaviour.
This problem will be quantified in section 8.2.4.

In the long tenn, proper operation requires thenna1ly efficient control of stove
heat supply and storage, in order to meet the "hot blast" demand. This is
particularly true when heat input increases are required over several periods
to meet a load increase. However, thennal efficiency requirements conflict
with the need to maintain both cyclic stability and to obtain rapid responses
to load changes.

The level of on-gas heat applied to meet a load is instrumental in maintaining

cyclic stability. Insufficient heating for a particular blast load can bring
about cyclic collapse. Collapse results when the system cannot provide,
independent of the number of cycles, sufficient heat to meet a particular load.
Stove changeovers would become more and more frequent. If allowed to
continue unrestricted, on-blast durations in the long term would become
shorter and tend to zero, as the stove system attempts to deliver a demand it
is incapable of sustaining. The greater the heat input, the lower the risk of
cyclic collapse.

The choices made in on-gas heating are, however, also instrumental in

determining the thermal efficiency for the stove system. Blast stoves utilise

184
Blast Furnace Stove Control Chapter 8

approximately 30% of the total fuel consumption in the steel making. process
(Mitter et al 1981 ). Given the current high cost of energy, stove operation
has enormous consequences on the efficiency, and therefore profitability, of
steel production. Clearly, the minimisation of heat input in the stove
operation is sought for efficient operation.

At the same time, blast furnaces require rapid responses to load changes in
the blast air demanded. Rapid response, however, comes at the expense of
either cyclic stability or thermal efficiency, depending on how the level of
on-gas heating is manipulated. If rapid response is attempted with
insufficient heat input, cyclic collapse is risked. With high heat input,
thermal efficiency is compromised.

The conventional industrial method of stove control is that of maintaining a

fixed period of operation. Although simple and flexible, this method has
been shown to incur considerable thermal efficiency penalties (Jeffreson
1979) and slow responses to load changes. Notably, this method relies on a
high level of residual heat in the on-blast stove at the end of a period.
To summarize the above discussion, the following statements can be made:
A. Short-teno Control:
Control Objectives: Total blast air flowrate (FT)
Blast air temperature (T3)
Manipulated Inputs: Main air flowrate over stove (F g)
Bypass air flowrate (Fb)
Measurements: Total blast air flowrate
Blast air temperature

185
Blast Furnace Stove Control Chapter 8

B. Long-1erm Control:
Control Objectives: Cyclic Stability
Thennal Efficiency
Speed of response
Manipulated Inputs: On-gas air flowrate
Blast air temperature setpoint (T3*)
Measurements: Total blast air flowrate
Blast air temperature
On-gas exit air temperature (Tg)

8.2.4 Control Difficulties

The short-term control task is made difficult by several factors. First, the
highly coupled blast temperature and flow-rate responses exhibit significant
process interactions in any traditional single-input, single-output control
scheme. The relative gain array for the experimental system is:

Fb Fg
0.61 0.39fr3
A= [
0.39 0. 61 1T
where the manipulated variables are the bypass flow, Fb, and the main stove
flow, Fg. The potential advantage of a multivariable control scheme is
therefore quite evident. In addition, the blast temperature response
demonstrates significant non-linearity and time-vatying behaviour over the
course of an on-blast period. This is demonstrated in Table 8.1 for the
experimental system. The pseudo-steady state open loop gains between blast

186
Blmt Fumace Stove Control Chapter 8

temperature and on-blast stove flow-rate are given for step increases and
decreases,. ·both early and late in the on-blast period.

Stove changeover represents a particularly disruptive disturbance in blast

stove operation. Upon changeover from a cooler stove (with zero or low
bypass flow) to a hot stove (with high bypass flow), the high level of
interaction between the blast flow and temperature loops becomes more
evident. In addition, there is a sudden change in the process dynamics. A
potentially destabilizing period of oscillation in the controlled variables may
ensue.

Table 8.1 Gains for blast temperature - stove flow response

Step in on-blast Gain (OC(1 hol)"l) x 1()l

stove flow
Start of on-blast End of on-blast
Increase 2.01 0.10
Decrease 0.73 0.48

The long term control problem is made difficult by the need to balance the
competing objectives of maximizing the thermal efficiency and the cyclic
stability, while providing rapid response to changes in the blast load-flowrate
or temperature. This type of problem is ideally suited to the application of
constrained Generic Model Control as discussed in chapter 2.

187
Blast Furnace Stove Control Chapter 8

8.3 MODEL AND CONTROL LAW DEVELOPMENT

8.3.1 Blast Air Temperatwe

A model of the blast air temperature can be developed from a simple mass
and energy balance around the mixing point of the main and bypass air
tlowrates. Thus

-~, I,{[ F,] [FT~']

dt
=-
'tT
- T2 + - - TJN - T3
FT FT
} (8.1)

where T3 is the blast air temperature (OC)

T2 is the air temperature of air leaving the stove (OC)
TJN ambient air temperature (OC)
F. is the on-blast stove tlowrate (1 hOi)
FT is the total blast air tlowrate (1 hOi)

and 'tT , is the time constant of the blast air temperature sensor (h)

Verification of this model was performed by implementing a step change in

the stove tlowrate (F.) and recording the blast air temperature T3' This is
shown in Figure 8.4. It can be seen that the model fits the recorded data
quite well.

A GMC control law can be developed from equation 8.1 and a reference
system performance trajectory:

188
Bl~t Fumace Stove Control Chapter 8

43.5 ~--~~--~-----,----~----~----~~

g 43.0
ao •
~ 42.5
LIJ
III:
42.0
~
I:L
41.5
~
~ 41.0

40.5
0 20 40 60 80 100 120
TIME (SECONDS)

Figure 8.4 Blast temperature response to an increase in stove flow

(8.2)

F
If the ratio of the flowrates -! is used as the manipulated variable, then the
FT

control law is:

F
This ratio -! is used as a setpoint R* to a low level PID controller to be
FT

discussed in section 8.4. The low level PID controller would achieve this
ratio R* by adjusting the bypass flowrate.

189
Blast Fumace Stove Control Chapter 8

The first term in equation 8.3 gives approximately the required ratio to
achieve the desired setpoint blast air temperature. The second term provides
process feedback to correct the model prediction. The pre-multiplying part of
the second term accounts for changes in process conditions.

8.3.2 Blast mow-rate GMC Control Law

The total blast flow-rate is the sum of the flows of two streams: the hot air
leaving the on-blast stove, and the cold bypass air. A dynamic material
balance over the mixing point yields:
dFT dFg dFb
=-+- (8.4)
dt dt dt

Equation 8.4 is in a convenient form for substitution into the general GMC
control law.

dFT dF dFb
- =-g +-
dt dt dt (8.5)

If the total blast flow-rate is controlled by manipulating the on-blast stove

flow-rate, Fg' then rearrangement of equation 8.5 to isolate the stove flow
term, dF/dt, gives equation 8.6.

(8.6)

190
Blast Furnace Stove Control Chapter 8

Finite differences were used to approximate the continuous derivatives in

equation 8.6. In the discrete terms of a digital sampling system. the equation
becomes: (F -F)
Bo.." I, (8.7)
T

where
k
PIF =K1.(F; - FT,> + K2, T l: (F; - FT,>
1-0
(8.8)

The bypass air flow-rate Fb, was ot measured directly on the rig system. A
suitable substitution was therefore determined. Since the temperature control
law, equation 8.3 utilized the ratio of stove flow to blast flow, it was decided
to introduce this ratio into the blast flow control law as well. Substituting the
ratio of stove flow to blast flow, R = F/FT' into a steady-state material
balance and rearranging gives the bypass flow, Fb:

(8.9)

Eliminating Fb from equation 8.7 then yields equation 8.10.

FBo.~, = R(k+l){F T, + TPIF} (8.10)

Equation 8.10 represents a control law for the blast flow-rate. It determines
the stove flow-rate, Fg' necessary to achieve the blast flow-rate set-point.
This stove flow-rate would serve as a suitable set-point, FgO, for a low level
PID controller manipulating the main valve.

191
Blast Furnace Stove Control Chapter 8

Control actions of the blast temperature GMC controller, equation 8.3, would
cause disturbances in the blast flow-rate. Feedforward compensation for
these disturbances can be made in the flow control law, equation 8.10, by
incorporating the flow ratio set-point, R*, of the temperature control law. In
the calculation procedure, therefore, the temperature control law would be
computed first to yield R*. Then R· could be substituted into equation 8.10
for ~l).

The blast flow control law now appears as:

F; =R* {For + TPIF} (8.11)

•

However, it is not possible that the flow ratio set-point, R·, can be achieved
by the low level PID controller within a single sampling interval. It is
reasonable to assume that the actual flow ratio at the end of the next
sampling interval will lie between the current flow ratio, Rv and the new
ratio setpoint, R*. A filtering of the desired ratio, R*, with the cUITent flow
ratio, Rt, could be performed to estimate a more practical ratio for
feedforward use. This ratio, Rru" is given by:

Rnlt =aR* + (1 - a)~ (8.12)

where a is the filter factor.

The blast flow-rate GMC control law now becomes:

192
Blost Fumace Stove Control Chapter 8

(8.13)

For most of the stove's operation the controller's task will be solely that of
disturbance rejection. The set-point, R·, would therefore not differ
significantly from one sample time to the next, and would likewise be nearly
equal to Rtt at all times. After a stove changeover or a load change, the set-
point ratio R· might change markedly from one sample time to the next. In
addition, it would differ, as described above, from the actual flow ratio, Rtt.
These differences, however, would diminish quickly as blast set-points were
attained, and soon R· and Rtt would again be nearly equal. The
approximation made in equation 8.12 is therefore, a reasonable measure to
account for lag in the low level controllers during instances of significant
changes in R·.

Experiments were performed with the GMC controller over the full range, 0
to 1, of the filter factor, a. A value of 0.75 was found to yield the best
controller response.

The control action obtained from the PIF term in equation 8.l3 would be
significantly affected by the magnitude of the flow ratio, Red.. This would be
particularly true near the beginning of an on-blast stage, when the flow ratio
would be very low. At such a time, the control action given by the PIF term
would be considerably reduced. As a result, the control law given by
equation 8.13 would exhibit very poor robust properties. It would rely
largely on the Rm.For model term for its control action. ' Controller
•

193
Blast Furnace Stove Control Chapter 8

performance would therefore be very sensitive to model errors, as the

correction through the error feedback term F; - F,[, in the PIF term, would
•
be reduced. An alternative formulation of the control law was therefore
developed:

(8.14)

8.3.3 On-gas S10ve Model

To enable the development of the long term controller using constrained

Generic Model Control, a model of the gas exit temperature, T l' from the
stove on-gas was required. This model was developed by considering an
energy balance on the stove and resulted in:

where ~+l is the predicted value of the sensor reading of the on-gas

stove exit air temperature one time step in the future eC)
~ is the current temperature sensor reading of the on-gas

stove exit air temperature eC)

~-l is the actual air temperature from the exit of the on-gas

stove one sample time ago eC)

~~ is the inlet air temperature to the on-gas stove one sample

194
Blost Fumace Stove Control Chapter 8

time ago (0C)

K, t,. are physical constants determined by experiment
I

T is the controller sampling time

The initial conditions for equation 8.15 are:

r.-o = Til
(S.168)

and
(S.16b)

where T.. is the theoretical temperature of first stove exit gas after

start of on-gas period.

Tt .. is the temperature of the exit gas as recorded by the sensor

at the end of the previous cycle.

The value of Tg can be predicted from the energy balance on the stove as:

r. =(1 -!),-:-t + k~ ~~ (S.17)

Figure 8.5 shows a plot of the model prediction versus experimental data. As
can be seen, a good agreement is obtained between the model described by
equations 8.15 to 8.17 and the experimentally recorded data.

An even more telling test of model accuracy is the ability to predict the cycle
length over many cycles. This is shown in Figure 8.6 over IS periods. Very
good agreement is displayed throughout, even following the load increase in
period 5, and the subsequent transient. This was essential to the correct
functioning of the optimization procedure to be discussed in the following
section.

195
Blast Fumace Stove Control Chapter 8

8.3.4 Long-ienn Control Algoddun

As was discussed in section 8.2.3, the long-term control of the blast furnace
stoves involves a trade-off between the thermal efficiency, cyclic stability and
response to load changes. This obviously suggests that the problem should
be formulated as an optimization problem, and in this work the framework of

45
..................................................
....0 40 .., ..... ..
01
II
. .'
0
'OJ
35
.-
I-
Z
0
30 .
~
~
0
25

20

0 500 1000 1500 2000 2500 3000 3500

TIME (Seconds)

-- EXPERIMENTAL -- MODEL
FIgure 8.5 On-gas Model vs Expenmental Data

196
Blast Furnace Stove Control Chapter 8

the constrained version of Generic Model Control was used.

1200

1100

1000

...... 900
..!!!.
c
Q
ffi
a.. 800
~
III
I
~ 700

600

500

400
0 2 4 6 8 10 12 14
ON-BLAST PERIOD NUMBER
o EXPERIMENTAL v MODEL PREDICTION

Figure 8.6 Dynamic Model - Experimental Agreement for On-blast Period

Duration over a 15 Period Optimal Control Run

197
Blast Fumace Stove Control Chapter 8

8.3.4.1 Umidng case operation

Simulations were performed to examine the stove rig used in this study. Its
thermal efficiency was determined with respect to several variables. The
maximum rig thermal efficiency was found to occur for maximum load at
both minimum on-gas inlet temperature and on-gas flow-rate, for constant
blast and on-gas flow-rates. This result is consistent with intuition and
reflects the significance of ambient thermal losses at higher heat input levels.
An inlet heat setpoint of 90°C was used for all experiments.

Further examination of rig thermal efficiency was performed by simulating

the rig operation under various on-gas inlet air flow-rates and blast
temperature loads. The on-gas inlet temperature and blast flow-rate were
held constant. The resulting contour and mesh plots for both equilibrium
thermal efficiency and equilibrium period are presented in Figures 8.7 and
8.8.

Two things are readily apparent from these figures. Firstly, the thermal
efficiency surface reaches a maximum for lowest on-gas flow and maximum
blast temperature load. This result would be intuitively expected. Secondly,
the maximum thermal efficiency presents itself as a ridge, below which the
thermal efficiency falls to zero. This corresponds to the verge of cyclic
collapse. As the thermal efficiency increases in Figure 8.7, a corresponding
decrease in period is shown in Figure 8.8. Where the thermal efficiency falls
to zero is the point where the period falls to zero - namely, the point of
cyclic collapse. The on-gas heat input has been reduced to the point of a

198
Blast Fumace Stove Control Chapter 8

Cl.
::!:
40
a 44.29

I
~
15
w
t- i=
t-
35
~
m

2.76
2000 4000 6000
ON-GAS F1 OW
Figure 8.7 Thermal Efficiency over a range of inlet On-gas Flow-rates and
Blast Temperature Loads

a 592
40
S
a.
::!:
it!
w
t-
IJi 35
:5
m

253
2000 4000 6000
ON-GAS FLOW

Figure 8.8 Equilibrium Period over a range of inlet On-gas Flow-rates and
Blast Temperature Loads

zero-period steady state heat exchanger. This limit was described by

Jeffreson (1979).

199
Blast Fumace Stove Control Chapter 8

S.3.4.2 The optimisdon problem

The application of constrained GMC method can be made directly to the load
control response. If f represents the blast heating load. then a constrained
GMC control law for blast air delivery could be:

df + 'A.; - 'A.; = K1(r - t) + K2 i(r - t)dt (S.lS)

dt 0

where ( is the required blast load setpoint

A few modifications can be made to equation 8.18 for the rig system.
Firstly, the long term rig operation can be regarded in discrete elements of
on-blast period. Hence, dfldt may be replaced by M, the change in load
delivery between two consecutive on-blast periods.
(S.19)

Secondly, both slack variables 'A.; and 'A.; are not required. Only 'A.;, which

defines positive deviations from the GMC trajectOlY, is needed. Negative

deviations will not occur. This situation exists because the stove rig can
always immediately accommodate a decrease in load. An increase, however,
may not always be immediately possible - particularly given the conflicting
goals of thennal efficiency and cyclic stabi1iW also being considered.
Therefore, only a single slack variable defining positive deviations is
necessaty.

The right hand side of equation 8.18 was simplified to give a control

200
Blaft Fumace Stove Control Chapter 8

trajectory proportional to the current deviation from the desired blast load. In
discretised fonn equation 8.18 becomes:

(8.20)

where 1'+1 is the load to be delivered in the next on-blast period. Kr was
specified to 1.0 so that the delivery of the full desired load is sought.

The maximum value of the slack variable ')..; is the departure from the GMC

trajectory which occurs when no move is made towards a new setpoint, ie.
1'+1 = fk and Mk+l = O. For this instance, define:

')..;_ =K r (fO - fk) (8.21)

The system's departure from the GMC trajectory may now be nonnalised by:

(8.22)

Any minimisation of ir will improve the system's response to a load change.

With a control law developed for the blast load control response,
consideration will now be given to the control objectives of stability and
thennal efficiency.

Firstly, given the importance of avoiding cyclic collapse for the blast rig

201
Blast Fumace Stove Control Chapter 8

system, it would be of practical interest to know the stability margin which

exists for any choice of operating variables. Equilibrium period was a good
indication of system proximity to collapse. The smaller the equilibrium
period, the closer the system was to the point of cyclic collapse. The
absolute point of collapse was the limiting or maximum thermal efficiency
situation, as described above.

Let the equilibrium period for any particular set of operating conditions be
defined by p... Let the equilibrium period for the same load, but at the heat
input of the limiting/maximum thermal efficiency case, be PL. A normalised
period can be defined:

(8.23)

Any attempt to minimise P, ie. maximise p.., will increase the stability of

the blast stove rig system.

Analogously, the rig's thermal efficiency may also be related to the limiting
case/maximum thermal efficiency case. Let 11 be the thermal efficiency
associated with a particular set of operating conditions. Let 11L be the
limiting case thermal efficiency - the maximum thermal efficiency obtainable
for the same load. Then, the normalised thennal efficiency is:

11L (8.24)
11 - -
11

202
Blast Fumoce Stove Control Chapter 8

Therefore, any minimisation of ii will maximise the thermal efficiency.

A formal description of the blast stove rig optimisation problem may now be
presented.

Blast S10ve Rig Optimisation Problem

Given: Desired Blast Air Load

Choose: Deliverable Blast Air Load
On-gas Heat Input
To Minimise:

~ 2
J =Wr . ("'r) - 2
+W•. (P) +Wt •
-)2
(11 (8.25)
where:

Mk+l + Ar = Kr(f* - fk) (8.26)

Ar_ =Kr (f* - fk) (8.27)

(8.28)

~<u<11u (8.29)

Selection of the weighting parameters, W, will allow trade-offs between

response speed, stability, and thermal efficiency considerations. Quadratic
functions are well known to be better suited for optimisation studies, and as
such the terms of the objective function have each been squared.

203
Blast Furnace Stove Control Chapter 8

8.4 IMPLEMENTATION

Implementation of both the short and long term controllers involved both a
hierarchy of functions and hardware and software constructions.

8.4.1 Functional merarehy

The functional hierarchy is shown in Figure 8.9. The lowest level of this
hierarchy is the blast furnace stoves themselves. Directly above the physical
equipment, including the sensors and valves, are two "low level" PID

OPTIMAL
CONTROL

... MUII .... NT. T 1.·TpO,NT.

GMC REGULATORY
CONTROL

... MUII .... NT. ..TpOINT.

i...

LOW-LEVEL PID
FLOWCONTROLLER8

"
.. IEMU II IE .. lEN T.
.. WOWI! pO.ITION •

BLAST FURNACE 8TOVE RIG

Figure 8.9 General Schematic of the Control Strategy

204
Blast Fumace Stove Control Chapter 8

controllers. One of these PID loops controls the stove tlowrate by adjusting
the main control valve, while the other loop controls the ratio between the
stove and bypass tlowrates by adjusting the bypass valve.

Low level PID control loops were implemented to increase the reliabili1;Y of
the overall control structure. Should any of the higher level functions fail to
operate, the low level PID loops would continue to maintain the last setpoint.

The setpoints for the PID loops were calculated from the GMC control laws
developed in sections 8.3.1 and 8.3.2. This is shown in Figure 8.10. Finally,
the long-term controller adjusts the setpoints for the GMC controllers and the
setpoint of a low level controller controlling the on-gas tlowrate.

8.4.2 Hanlware and Software memrchy

The computer control hardware used to implement both the short and long
term control is shown in Figure 8.11. The functions of each hardware
component is shown in Figure 8.12.

The functions of each computer facili1;Y is as follows:

QNX Data Logging Server: The primary purpose of the QNX network and
two PC's is to accept the sensor readings from the physical equipment
and return signals to adjust the valve settings. The QNX server system
is equipped with analog-to-digital as well as digital-to-analog converters,
and digital input and output facilities. It communicates via a serial

205
Blast Furnace Stove Control Chapter 8

aLAeT "LOW aLAIIT T .....

T.-
• • T ..OINT • • T ..OINT
".,.-
I I
.J. L
R-
aLAIIT "LOW aLAIIT T .....
. . .0 ~". ~ • .. 0
OONTROL
L~
r:..,. T~ OONTROL
L~

p.-
• R-

.TOV. "LOW "LOW RATIO

..ID ... D

MAl.
• W.
OONTROL

...._.
P ....
p.
OONTROL

.YOW•
P ....
p.
.LAeY
_....
.W•
P ....
P9

I .LAIIT "URNAO• •TOV. R ••

I
Figure 8.10 Signal flow diagram for GMC control of the blast stove rig

communication line to the TACTICIAN control system.

TACTICIAN Control System: The TACTICIAN Process Control System

(Turnbull Control Systems, 1988) is a real-time personal computer based
system. It allows the design and execution of process control strategies,
simulations, and data acquisition.

The software is based on a graphic, mouse-driven user .interface that

allows the design of executable control strategies and operator interface

206
Blast Furnace Stove Control Chapter 8

~ooo
FORTRAN
PROGRAMS
EIhemet
TCP/IP - -- - - r - - - - - - - --L-- -
PClAT
TACTICIAN
SOFTWARE

PClAT
QNX
SOFTWARE

Figure 8.11 Computer Hardware Implementation

graphics. A libraI)' of commonly used "user configurable" blocks is

available. These encompass real-time input/output signalling, control
(eg. PID), signal conditioning, timing, basic logic, recipe loading, and
data collection and storage. These blocks may be linked as desired to
form the required process control configuration. For more sophisticated
functions than those provided for in the basic block libraI)', the
developer may write C-Ianguage programs in seven "user blocks" and a
single "user task". For instance, tasks in this investigation such as
setpoint sequencing, tailored data acquisition, complex process
simulators, and the model-based controllers required these programs.
Such programs are linked into the control configuration, and are
executed once each scan interval during TACTICIAN applications.

One particular program, the "USERCOMM" program, contains the C-

207
Blast Fumace Stove Control Chapter 8

. ...1_.
...."

OI'T11IAL CIOIITIIOL ALeO.ITHIl

_.11e . . .,.. ....., . .

TI 1
L;.
....... I
• ...,.,.H•••T .IIT....

'Ii
T.
~
HTI'CII.T
HLMT .....
••TI'CII.T
eLMTu.,
HT....T
,,: 1':., T..
•
'llaTIDIAII CIOIITIIOL
eLMT,.... •• ....."u.,
...
•••
. . . TIIOL
l:!-
t., ~
T~
•••
...TIIOL
L•

"'. ,•. ••
.........
,.. ... .... .".., ........
,
PlO

-- - .- _.
. . .THOL . . .TIOL . .TIOL

=='
'. .&.aIr
••
L~ . . . . LIMI•••
I
II 1 1 t
•...." 'u.... .,..,. III.
nrr TIl
Figure 8.12 Schematic for the Optimal Control of the Blast Stove Rig

language instructions for reading and writing TACTICIAN requests and

data to the serial port of the TACTICIAN computer. This accomplishes
the TACTICIAN-side of communication with the QNX data logging
server. It should be noted that the functions of the QNX data logging
server and the TACTICIAN control system could have been
implemented all within the TACTICIAN system itself. Reasons outside
this single project dictated this architecture.

208
Blast Furnace Stove Control Chopter8

Figure 8.13 - Flowchart of Optimisation Program

209
Blast Furnace Stove Control Chapter 8

Apollo Computer System: The long term control and optimization programs
were too numerically intensive to implement within the TACTICIAN
control system. A UNIX-based Apollo DN 10000 was used to
implement these functions. This was interfaced to the TACTICIAN
control system via an Ethernet connection. The interface software
allowed· the TACTICIAN system to read and write data files on the
Apollo workstation. Software interfacing of the control and optimization
functions was achieved by way of intermediate data files. The
optimization algorithm was a FORTRAN program which ran on the
Apollo workstation.

8.4.3 Solution of 1he Optimization Problem

The optimisation problem requires solution for the optimum blast load and
on-gas heat input This is in fact a solution set of four variables: the blast
load entails blast temperature and flow-rate; while the heat input also entails
on-gas inlet temperature and flow-rate.

The choice was made above to utilise 90°C for the on-gas inlet temperature.
Hence, the choice of on-gas heat input for a particular load reduces to the
selection of the on-gas flow-rate.

The constant supply of the required blast flow-rate is particularly crucial to

the operation of the furnace, and its supply is also more straightforward. The
temperature, however, is more difficult to address as it is determined by the
heat storage condition of the blast stoves. Therefore, the blast flow-rate was

210
Blast Fumace Stove Control Chapter 8

always delivered, and the variable to be controlled was its temperature.

The optimisation problem therefore reduces to the selection of two

independent variables: on-gas flow-rate and blast supply temperature.

Load changes were only implemented at the start of an on-blast period.

Therefore, the optimisation strategy would sample the stove conditions at the
end of a period and proceed to determine optimal settings for the new period.
The new setpoints were then downloaded to the regulatory controllers. In
addition, only load increases were considered of interest in this study. Load
increases placed serious consequences on response speed, thermal efficiency,
and cyclic stability. Load decreases were considered trivial as they could be
met immediately and completely by the stove system at any time.

The optimisation strategy employed the dynamic models of the stove rig to
simulate operation of the blast rig. The rig's operation was simulated for
each set of independent variables trialled in the optimisation procedure. The
initial rig conditions were specified to be those existing at the end of the
preceding period, and simulation would continue forward until a cyclic
equilibrium or collapse occurred. The objective function and all associated
equations presented above were evaluated at the predicted eventual system
conditions. Based on the successive evaluations of the objective function, the
search strategy would revise its solution set and continue searching until the
specified tolerance was met. Figure 8.13 presents a flowchart of the
optimisation procedure.

211
Blast Furnace Stove Control Chapter 8

The optimisation algorithm employed for rig optimal control was the Box
Complex Method (Box 1965). This is a direct-search, non-gradient based
algorithm. The Box Complex method was found to be particularly robust for
this optimal control problem. This method has been successfully used for
other similar industrial optimisation problems, including the optimisation of a
copper anode furnace operation by Bateman (1992).

The optimisation program was executed once every period, immediately after
changeover and after the process measurements from the end of the preceding
period were read. Clearly, no new optimal settings could be down-loaded to
the regulatory controllers until the optimisation was complete.

To ensure that the optimisation did not take an excessively long time, and
consequently an overly large proportion of the actual on-blast period, a
maximum was placed on the number of iterations the search could make.
The maximum value used was 15 iterations, and this typically entailed a
computational waiting time of 30 seconds.

This limit on the number of search iterations was generally sufficient to

determine the optimum for most optimisation searches for the stove rig. In
the initial instance of a load increase the search would occasionally not reach
all the distance to the true optimum. However, it would closely approach it.
The optimisation search of the ensuing period would then carry the rig
system the rest of the way to the optimum.

Figure 8.14 shows the approximate typical scheduling for the optimal

212
Blast Fumace Stove Control Chapter 8

CHANC'lEOVER

PRE-CHANC'lEOVER STOVE CONDiTION8

WRITTEN TO FILE BY TACTICIAN U8ER-TAIK

OPTIMISATION FORTRAN TASK READS FILE

OPTIMISATION RE8ULT8

I
WRITTEN TO FILE
_OPTIMISATION 8EARCH- TACTICIAN U8ER-TASK READS
OPTIMAL IETPOINTI AND
DOWNLQAD8 TO
REC'lULATOAY CONTROLLER8

o --:;0 20 30
ON-BLAST TIME (SECONDS)
4h $Ho
Figure 8.14 Typical Scheduling for Online Optimal Control

controller.
8.4.4 Tuning

(a) PID Loops

Open-loop step tests were perfonned on all PID loops, and the resultant
transient was fitted to a first-order plus deadtime model. Miniminn Integral

213
BllBt Fumace Stove Control Chapter 8

Error Tuning formula for setpoint changes (Smith and Corripio, 1985) were
then used to calculate appropriate controller constants. Some trial and error
adjustment of these parameters were necessary to ensure stable control over
the entire range of operating conditions.

(b) GMC Reference System Parameters

The selection of the GMC reference system parameters t and ~ were made on
the basis of known speed of response of the process and confidence in the
process model. Values used were:

Blast Flow Controller Blast Temperature Controller

~ 10 1
t 15 20

8.S RESULTS

8.S.1 Short-renn Results

To compare the results obtained using Generic Model Control, two PID
controllers were also implemented. These controllers controlled the blast
tlowrate by manipulating the main valve, and the blast temperature by
manipulating the bypass valve. This was the recommended pairing as
suggested by the relative gain array analysis in section 8.2.4. These
controllers were tuned by the techniques discussed in the previous section,
and implemented on the TACTICIAN control system.

214
Blast Fumace Stove Control Chapter 8

Sequences of step changes in blast temperature and flow-rate set-points were

used to test the different control methods. The integral of the time-weighted
absolute error (ITAE) for each step change response was used to compare

BLAST TEMPERATURE (deg C) BLAST FLOW-RATE (Uh)

42 8&00

40 I
.• --
1-00
38 -
88 2&00

84 - - 2000

82
1&00
aD
I I I I
28 1000
o 180 380 &40 720 800 1080 1280 1440 1820
ON-BLAST TIME (Seconds)
- BLAST TEMP SETPOINT -+- BLAST FLOW SETPOINT

Figure 8.15 Step test sequence 1, load changes in blast temperature

quantitatively the controller performances. A lower IT AE value was the

indication of better control. The behaviour of the controllers over the stove
changeover transient was also examined.

215
Blaft Fumace Stove Control Chapter 8

The step changes in the set-points for the blast temperature and the blast
flow-rate were chosen to be approximately 1001S% of the operating range.
These load changes were centred within the operating range. Load increases
were of a magnitude which could be supplied by the on-blast stove in the
short term. That is, no load increases were used which necessitated increased
heat inputs over several cycles, i.e. a long term control problem.

Sequence 1, Figure 8.1S, involved alternating step increases and decreases in

blast temperature set-point of a magnitude of 6°C. This was continued
throughout the course of an on-blast period. The blast flow set-point
remained constant at 3000 I h-l. Sequence 2, Figure 8.16, involved
alternating step increases and decreases in blast flow-rate set-point of a
magnitude of 600 I h-l. In this instance, the blast temperature set-point
remained constant at 30°C.

Each step test had a duration of 180 s, which represented approximately 12%
of the total on-blast period. Due to the interaction between the two loops, the
ITAE values were also calculated for the first 180 s after stove changeover.
The latter tests were done to examine control over the changeover
disturbance. The system was always at set-point before commencement of the
step test sequences and again following the test sequence, before stove
changeover. ITAE values were obtained for both stove 1 and stove 2 in on-
blast phase. Consistent results were obtained for both stoves.

The typical test procedure involved operating the two-stove system through
several cycles of on-blast and on-gas phases. Blast temperature and flow-rate

216
 Blast Fumace Stove Control Chapter 8

were controlled by either the GMC or base-case PID control scheme. Test
conditions were maintained as consistent as possible between the different
experimental runs to eliminate spurious effects on controller performance.
The air supply was regulated at 280 kPa gauge. The on-gas manual valve
was fully opened, giving constant on-gas flow of approximately 5000 I hoi.
The heater sei-point was set at 90°C for all tests. Where possible,
experiments were performed with consistent ambient conditions. At the end

BLAST TEMPERATURE (deg C) BLAST FLOW-RATE (LIb)

8800

40 -8400

- 8200
88
- 8000

2800
82 I-

= = = = = = = = " - .2800

28 1 12400
o 180 880 640 720 800 1080 1280 1440 1820
ON-BLAST TIME (Seconds)
- BLAST TEMP SETPOINT -+- BLAST FLOW 8ETPOINT

Figure 8.16 Step test sequence 2, load changes in blast flow-rate

217
BlaYt Fumace Stove Control Chapter 8

of the prescribed step test sequence, the stoves exchanged duties. The new
on-blast stove was then placed through the same sequence of load changes.
This procedure was repeated for several cycles.

(a) Load changes in blast temperature (sequence 1)

Figure 8.17 presents the ITAE values for load changes in blast temperature
for both GMC and PID control. In all cases, GMC performed at least as well
as, and frequently significantly better than, PID.

Non-linearity and process interactions. The most significant improvement in

control was witnessed for load decreases in blast temperature set-point. For
these situations, PID control experienced considerable difficulty in achieving
set-point compared to its performance on equivalent step increases. GMC, on
the other hand, was consistent in its response to both step increases and
decreases. These observations are evident in Figures 8.18 and 8.19 for a step
increase and decrease in temperature set-point, respectively. In Figure 8.18
the manipulated variable responses have also been shown.

The GMC responses to both blast temperature set-point increases and

decreases were smooth, first-order responses to the new set-points. Only
limited, short-term effects on the blast flow loop were observed. The PID
control response to the set-point increase was oscillatOIY, and had increased
interactions with the flow loop. Set-point was, however, still achieved within
the 3 minute test period. The PID control response to the set-point decrease,

218
Blast Furnace Stove Control Chapter 8

-
f-
30
25

III III
~
--
-
....
~
12
10
8
6
4
2
o
UP DOWN UP DOWN UP DOWN
STEP CHANGES OF' 6 DEG C
_ GWC c:::::J PID

Figure 8.17 Controller ITAE values for blast temperature set-point sequence
(sequence 1)

in marked contrast, showed considerably increased oscillation in blast

temperature and severe interactions with the blast flow. In fact, the PID
controller did not achieve the set-point decrease within the test period.

The difference in process behaviour for step increases and decreases in blast
temperature represents a serious system non-linearity which the PID
controllers were ill equipped to handle. Although coping satisfactorily with a

219
Blast Furnace Stove Control Chapter 8

40

....
U
~ 36
38

,/'
.... ....

.......
CL
:l 34 r
r
....
III
j
32
,,I

30 ~,

CD (0)
28
.: 3600
d; ;\
~ 3200 k-f-i-,l\~~"'r" ;;"'"';''
'''';' ' -.....".,,,.,...,.,..-='-'''''------1
~ 2800 '-l: r" ---" ~
~ ~
j....L'~_...I...-___''--__'__
J '"
CD 2400 _'_(:...b):'''''''O'''_ _ ' __ .L_____'_ ___1
70
~
0
60
SO
~
... 40
~
30
z>";;: 20

'" 10 (e)

III
~ 20
CL
~ 10U-_...I...----''--~_ _'_~(d~)~~_ _L_~~
o W ~ ~ ~ 100 lW 1~ 1~ 1~

TIME (S.cond.)
SETPOINT _~~ _ CMC ...... PIO

Figure 8.18 Blast temperature set-point step increase from 30°C to 36°C
(step 1, sequence 1)

step increase, the same PIO tuning constants caused excessive control action
for the equivalent step decrease. Detuning of the PIO controllers would
lessen the control action. It would, however, simultaneously deteriorate
performance for other load changes. Furthermore, oscillation between the
two PIO control loops for blast temperature and flow emphasized the strong
interactions existing between these control loops. The PIO controllers could

220
Blast Fumace Stove Control Chapter 8

not compensate for these interactions.

J8
U' ~6
go
~4
~
n.
32

..~
:::i
~
JO .......: .....
28
CD
(0)
26
4400
Z
~ 4000
. .......
...~
"-
3600
::""
-
jI

Iii ~200
, I
:' '- ... .... .~'"
....
~ ......... "
............ .
OJ 2800
ro~__~~~··~
·· ~~~(~b)~__~~~~~
-cv 60
& 50
~ 40
~ 30 l\
~ 20 \, / ... --7'...... _... _-:.. . . ------.::.:~---~::OO:::----~:-'oi~
~ 10 \ " __ ,,,/ .

o ro w w ~ 100 lro lW lW 1~

TillE (Second,)
- SETPOINT • __ . CIIC ..... PID

Figure 8.19 Blast temperature set-point step decrease from 36 0 to 30 °C

(step 2, sequence 1)

However, the GMC controller was able to give consistently good responses
for both temperature load increases and decreases, while significantly
reducing interactions. The good response to temperature load change may be
attributed to the steady state model approximation to the required flow ratio

221
Blast Furnace Stove Control Chapter 8

for the current set-point and stove condition. The traditional PID controller.

on the other hand. relied entirely on its error signal T3* - T3 to detennine
•
its control action.

The reduced temperature and flow interactions with GMC control resulted
from the flow ratio tenn incorporated in the blast flow control law. The
Rnlt> provided a feedforward estimate of what the flow
'filtered' flow ratio,
ratio was expected to be in the next sampling instant. In effect Rnll was a
measured disturbance as far as the flow control law was concerned. To this
end. controller operation was analogous to feedforward or decoupling control.

Time-varying process behaviour. As the on-blast stove cooled and the stove
exit temperature dropped. blast air delivery was made under different process
conditions from those which existed earlier in the on-blast phase. With
reference again to Figure 8.17. the IT AE values for PID control responses to
repeated step increases and decreases in blast temperature showed a trend
with respect to time. The flow and temperature IT AE for equal step
decreases actually decreased as the on-blast time increased. The temperature
IT AE values for step increases showed an increase as time increased (flow
ITAEs for the latter showed no significant change). The GMC controller.
however, was essentially consistent in its IT AE values throughout the course
of the on-blast period and for both step increases and decreases.

Figure 8.20 shows the GMC and PID control responses for a step decrease in
blast temperature near the end of the on-blast period (step 6 in sequence I).

222
Blast Furnace Stove Control Chapter 8

The GMC response was similar to its response for the earlier step decrease
(see Figure 8.19). The PID control response, however, was improved
compared to its response to the earlier step decrease. For the later step
decrease, set-point was actually approached within the 3 min test period,
unlike the earlier step decrease.

The GMC controller, in marked contrast to the PID controllers, showed fairly
consistent control performance over time. This was again directly attributable
to the incorporation of the process model in the control law.

(b) Load changes in blast flow-rate (sequence 2)

Figure 8.21 presents the blast flow and temperature ITAE values for GMC
and PID control of the rig under step test sequence 2 (see Figure 8.16). As in
the case· of the blast temperature sequence, the GMC control again performed
at least as well as, and generally significantly better than, PID control.

Time varying behaviour and process interactions. Figure 8.21 shows an

obvious trend in PID control performance over the course of the on-blast
phase. At the outset, the PID controllers demonstrated extremely poor
control, as exhibited by the very high IT AE values for both blast temperature
and flow-rate. Performance, however, improved steadily, essentially
becoming equivalent to that of GMC at the very end of the period.

The GMC controller responses, again in marked contrast, were consistently

good throughout. No significant variation over the on-blast period was
exhibited.

223
Blast Furnace Stove Control Chapter 8

38
'0

·
36 :":\
co
,.'.'-
\
\
~ 34
IL
::t 32
.•.. '
~ ...... ...........
:lO .. ::::::::,..-
§ 28
CD
' ....... ~

(a)
26
4000
~
d 3600
\,.. .
-

~~ 3200
;1 , ..... : ..
"-
l-
2800
" I I ..... ,-~.- "
~ I I
.... .:
'" 2400
V '

(b)
70

·
'Coo
~ 50
~ 40 ~'i\
~ 3D : ~. .•. .... ..
:;i, I -, ..- " . -----~,.~::- ---.:::'-'~-----
~ 20 \'~'-" ,/';- ~7-"'---""~"'~
~ 10 -, ---'
(0)

·
'Cs8
g. 70
t!.oo
~ 50
.,> 40
~ :lO
~ 20 w-__ ~~ __ ~~~~
(d) __ ~~ __ ~~

o w ~ 60 60 ~ 1W 1~ 160 160
lilAC (Seconds)
- SETP01N1' ---. Gwe ...... P!D

Figure 8.20 Blast temperature set-point step decrease from 36° to 30°C late
in on-blast phase (step 6, sequence 1)

Figure 8.22 presents the GMC and PID control responses to the 600 lb-! step
increase in blast flow, at the beginning of an on-blast period (step 1,
sequence 2). The GMC response was rapid, with little overshoot or
oscillation in the blast flow. Minimal effect on the blast temperature
response was observed. The PID control response, however, showed
considerable oscillation in the blast flow-rate and significant interaction

224
Blast Fumace Stove Control Chapter 8

-- 15.0
a

-....
I I)
(
0 12.5

10.0
~
•9... 7.5

5.0

§m 2.5
Il "--.
-..,0.0 --'----I

--....
( 8
0

6
~
Q..
::::E .4-
~

§m 2

0
UP DOWN UP DOWN UP DOWN UP DOWN
STEP CHANGES OF 600 L/h
_ G~C c:::::J PID
Figure 8.21 Controller ITAE values for blast flow-rate set-point sequence
(sequence 2)

between the flow and temperature loops.

Figure 8.23 shows GMC and PID control responses to the last 600 1 h-I step
increase in blast flow set-point (step 7, sequence 2). This step change
occurred over 1000 s after the first step increase discussed above. The GMC
control response was very similar to that of the earlier step increase (compare
with Figure 8.22) and was again very good. The PID controller response was

225
Blast Furnace Stove Control Chapter 8

(b)

~80
~ 70
~ 60
'$I ~
~
III
40
!:2 30
.. (d)
In 20 LL_.l..----I._-'-_.l..----I._ - ' -_ .l..-----'_....!
o W ~ ro W 100 IW I~ lro IW
m.cE (Seconds)
- SETPOINT --- GIIC ...... PID

Figure 8.22 Blast flow-rate set-point step increase from 2800 to 3400 1 h- I
late in on-blast phase (step 1, sequence 2)

dramatically improved over its earlier performance. Set-point was achieved

more quickly, with little effect on the blast temperature loop.

Linear PID control was unable to adapt to the time-varying process

conditions. The model-based GMC controller did accommodate this
behaviour, and therefore supplied improved responses throughout. This

226
Blast Fumace Stove Control Chapter 8

36
U' 34

!'" 32
~ 30 I-...c.:!.~:':':'';';
''' ~
'''"" -.-_-I
'' _"",:,:~---_ _ _

.... 28
t;
~ 26 (I)
4~P-~--~~--~~--L-~--~~

:d 3600 ,"', ., .:. '". ,',

~
...
3200 ~jr
: '--'
~2800 ' :
CD

(d)

Figure 8.23 Blast flow-rate set-point step increase from 2800 to 3400 I hoI
late in on-blast phase (step 7, sequence 2)

improvement resulted largely from reduced interactions with the blast

temperature loop, as discussed earlier.

227
Blast Furnace Stove Control Chapter 8

(c) Responses to stove changeover disturbance.

Stove changeover represents a particularly disruptive event in stove

operation. There is a sudden change from a coolon-blast stove with little
bypass air, to a hot stove with high bypass flow. The controller responses
over the changeover transient are significant with the GMC controller
showing a dramatic reduction in ITAE values compared with PID control.

Figure 8.24 presents typical GMC and PID control changeover behaviour.
The GMC response showed only minor deviation in blast temperature
following the changeover from a cool to a hot on-blast stove. Deviation in
blast flow from its set-point was also minor, and was quickly eliminated. No
significant interaction between the two loops was evident. The PID control
response, however, showed significant oscillation in the responses of both the
blast temperature and blast flow-rate. Set-points were not achieved within
the test period.

Once again, the use of the process model within GMC was largely
responsible for the improved control. Stove changeover and the subsequent
increase in stove exit temperature, T2' represented disturbances to the system.
Feedforward compensation minimized the effect of changeover on the blast
temperature. The base-case PID control response had to wait until the effect
on the blast temperature, T3' had become evident, and only then did it react.

The process interactions between the blast temperature and flow-rate were
again reduced with GMC by the use of the flow ratio, Rr.lt' in the blast flow

228
Blaft Fumoce Stove Control Chapter 8

" 34
o
~ 32
~ :;".
Q. 30 ,},;.-...'>:........ ~J;--;.L."\
~ •...•....-1.... •.•...,.,.' .: .".

Iii 28
:s
m 26
(0)
3400
"~ 3200
...I
'oJ

~ 3000
g
I&. 2800
I-
III
:s 2600
m
2400 (b)
o w~ ~ ~ l00lWl~lrol~

TIME (Seconds)
- SETPOINT --- GMC ........ PIO
Figure 8.24 Controller responses to stove changeover

control law. As observed in the temperature set-point step changes (sequence

1), the control actions of the blast temperature controller were disturbance
inputs to the blast flow controller. Feedforward compensation was made to
maintain the blast flow set-point and minimize oscillation.

229
Blast Furnace Stove Control Chapter 8

8.S.2 Long-1enn ResuUs

Online optimal control experiments involved running the blast stove rig for
15 periods, with a load increase after the fifth period. The constant load
periods before and after the period 5 increase allowed the rig's operation to
stabilise out The load increase implemented was a step increase in blast
temperature from 36 to 40°C. Blast flow-rate remained constant at 3500 Llh.
The on-gas inlet temperature also remained constant at 90°C.

(a) Preferential Weighting of Different Terms in the Objective Function.

Online optimal control tests were performed on the stove rig with different
preferential weighting of objective function terms. A relatively higher
weighting was always used for the stability weighting, Ws. This was to
ensure operation remained above the practical collapse limit of the stove rig.

Figure 8.25 illustrates contour plots for the objective function and each of its
three contributing terms for a typical experiment The optimum shown in
Figure 8.25 (a) was determined during the online optimisation of rig
operation for the period 5 load increase. The optimisation was performed
based on the stove conditions at the end of period 4 and the desired load for
period 5. The optimum defines the setpoints to be delivered during period 5.
The weighting factors were specified as:

230
Blast Furnace Stove Control Chapter 8

Thermal Efficiency Weighting, Wt = 1.0

Cyclic Stability Weighting, WI = 100
Load Response Weighting, W,= 5.0

In the region of the optimum in Figure 8.25 the contributions by each of the
three terms are of approximately the same order of magnitude. As the on-gas
heat input is decreased and the collapse point is approached the objective
function becomes dominated by the cyclic stability term. This is apparent by
the increasing slope of the functions in Figure 8.25 (a) and (c). In each case
the effect of the quadratic form of the objective function is obvious by the
changing distance between contour lines.

The optimum point in Figure 8.25 (a) specifies a blast temperature below the
desired load of 40°C. This was a consequence of the dominant thermal
efficiency and cyclic stability terms, which prevented the full load from being
delivered.

(b) Changing the Thermal Efficiency Weighting

Figure 8.26 shows the blast temperature and on-gas flow setpoints determined
by the optimal control strategy for different values of the thermal efficiency
weighting factor, Wt. The first experiment employed a value of 1.0, twice
the size of that used in the second run. For both runs the other two
weighting factors were We = 5 for response speed, and W. = 100 for stability.
Both runs showed similar responses in the blast temperature response, with
the new load being fully delivered for the second period after the load

231
Blast Fumace Stove Control Chapter 8

40;:;..0_-.-_ _---,.--_ _-,

a..
a
Q.
::IE
t- ~ .4'4

§
£D
§
£D

2.52
L.;.;1.;;;J3=mULL.J....lo._::;"""_.-£---I

2000 4000 6000 2000 4000 6000

ON-GAS FLOW ON-GAS FLOW

(a) OBJECTIVE FXN J (b) RESPONSE SPEED TERM

~
£D

L.:..;;..;L'--'-"'---'--'--'-'-'--'--L-..J15•11
2000 4000 6000 2000 4000 6000
ON-GAS FLOW ON-GAS FLOW

(c) CYCUC STABIUTY TERM (d) THERMAL EFFICIENCY TERM

Figure 8.25 Contour Plots of the Objective Function and its three component
tenus at the period 5 load increase

change. A marked difference, however, is evident in the heat input

specifications made by the optimiser. The first run, with the higher thermal
efficiency weighting, attains an eventual heat input rate 10% lower than the
other run.

232
Blast Fumoce Stove Control Chapter 8

-
41 ~----~--~~--~----~----~----~----~~

. 2,40
B;
~
...
~
.. ..
~ ~
...
~
..
~
...
~
,..
~
..
~

~ 39

~
a.
38

~
- --
37

~ 36 - . .-..-.. ..-.

35~--~~--~-----+-----+----~----~----~~
2 4 6 8 10 12 14

3000~--~~--~----~----~----~----~----~~
o 2 4 6 8 10 12 14
ON-BLAST PERIOD NUMBER
o IUGH 'IEIGB'1'DfG V LOW WBIGB'1'DfG

Figure 8.26 Effects of Changing the Thennal Efficiency Weighting, WI

This result would be expected for greater relative importance being placed on
thennal efficiency. Even the overdamped response is indicative of a higher

233
Blast Fumace Stove Control Chapter 8

priority on energy saving. The second run, with its lower efficiency
weighting, actually shows an overshoot in heat input In this case, the
relatively higher weighted stability term can prescribe a high heat input to
maintain a greater stability margin.

(c) Changirig the Cyclic Stability Weighting Factor

Figure 8.27 shows the effects of varying the cyclic stability weighting factor,
Ws' The first run employed a value of 100, whereas the second used a value
of 150. For both experiments the other two weighting factors were We =5
(response speed) and Wt = 1.0 (thermal efficiency).

The experimental run with the higher stability weighting displays a much
more conservative delivery of the new setpoint It requires four periods to
fully deliver the new setpoint, whereas the other experiment delivers it in two
periods. The more gradual delivery of the new load would allow the stove
system extra time to increase its overall heat content, and therefore increase
the stability margin. Moreover, the high weighting causes a higher overall
on-gas flow-rate towards the end of the run. This again gives the rig system
a higher relative stability margin than obtained in the other experiment This
result would be intuitively expected for greater relative importance being
placed on cyclic stability.

(d) Changing the Response Speed Weighting Factor

Figure 8.28 shows the effects of varying the cyclic stability weighting factor,

234
Blmt Fumace Stove Contml Chapter 8

W f' The first experimental run employed a value of 10, and the second used
5. The common weighting factors were WI = 150 and WI = 1.0 for cyclic
stability and thennal efficiency, respectively.

41

- - - - - - ...
P
g 40 -
Bi
1&.1 39
ar::

! 38 /
m 37

! 38

35
-- - -- ...
2 4 8 B 10 12 14
8DDD
2
~
Bi 5D00
~
~I 4000
z
0

3D00
0 2 4 8 B 10 12 14
ON-BLAST PERIOD NUMBER
0 LOW WBIOBTIHO v moa "DIOIft'IlfO

Figure 8.27 Effects of Changing the Cyclic Stability Weighting, WI

235
Blast Fumace Stove Control Chapter 8

The experimental run with the higher response speed weighting displays a
much more rapid delivery of the new setpoint than the second run. The first

- - - - - - -- ...
41

-e
CL.
en
40

39
~

m 38
CL.

m37
§ 36
ID ..
35
2 4 8 B 10 12 14

-~ BOOO

CL.
en 5000

~
~I 4000
z
0

3000
0 2 4 6 B 10 12 14
ON-BLAST PERIOD NUMBER
0 HIGH WBIGB'l'IRG V L01f 'lBIGllTING

Figure 8.28 Effects of Changing the Control Response Weighting, We

236
Blast Fumace Stove Control Chapter 8

run requires only two periods to fully deliver the new setpoint, whereas the
second delivers it in three periods.

The more rapid control response due to the high weighting occurs with a
much more h1>eral addition of on-gas heat to the rig system. In fact, the on-
gas flow setpoint overshoots its eventual equilibrium value. Thermal
efficiency was clearly compromised with the higher emphasis on rapid
control response. The second experiment, in contrast, displays a much more
conservative response both in blast temperature delivered and on-gas flow-
rate supplied. Once again, this result would be intuitively expected when
higher importance is placed on rapid control responses.

An interesting result was obtained when the response speed weighting was
specified at a much lower value. Optimal control with the response weight
Wr = 1 gave the results shown in Figure 8.29. At period 5, when the
optimisation program was executed for the load increase, the low response
weighting actually allowed the blast temperature to drop below its previous
value. Clearly, the relatively higher stability and thermal efficiency
weightings were dominant and drove the system in the direction opposite to
the intended load change.

(e) Comparison to Conventional Stove Control

The conventional operating method for blast stoves in industIy has long been
the fixed-period method. The advantage of this fixed-period method solely
lied in its simplicity. Strict control over the average heat input was not

237
Blmt Fumoce Stove Control Chapter 8

38~----r-----~----~----~----~----~----~--'

g 37
B;
iii! 38

I:
! 33

5oaa~--~~--~-----+----~----~----+-----r-~
2 4 8 8 10 12 14

~
Bs
4000
-
...........
13000
I 2000

o 2 4 8 8 10 12 14
ON-BLAST PERIOD NUMBER

Figure 8.29 Effects of VeIY Low Control Response Weighting

necessary as a significant level of residual heat was maintained in the stove.

Similarly, no condition was placed on the final fraction of bypass mixing.
The operator could accommodate small load increases utilising this residual
heat. For larger load increases, a small increase in load would initially be
delivered, with subsequent heat increases in the next on-gas phases to
accommodate the remaining load increase.

238
Blast Fumace Stove Control Chapter 8

41

.
g40 : = -- - = --
;~-
II
at
38

FI
"
J
-- = -
•• ,•
38

38 • • • • •
,,+
i .l. 1b ,~
eooo

..
~
I
eooo
- -
1
: : :

z
'.
/

, ,
4000

- • ~
.
aoo
•
1---
~
• 1t ,t ,,+

g -
800 -~ ~ ~ ';" ';' ~ ~ ~ ';" ';" ~

I
!
400

200

•
• •
0
0 2 4 10 12 14
ON-aASf PERIOD MU.....
0 OPI'IIUL
" COMI..-rIOIIAL

Figure 8.30 Optimal and Conventional (Fixed-Period) Control Performance -

Period Number

239
Blast Fumace Stove Control Chapter 8

The simplicity of this method, however, comes at a cost. Jeffreson (1979)

showed that thermal efficiency decreases for a given load as the level of
residual heat increases, and that any form of fixed operating period regime
would show lower thermal efficiency.

The conventional control method was compared to the optimal control

scheme using simulations. Figure 8.30 shows the results of a fifteen period
run with a step increase in blast temperature at period five.

The fixed-period method displayed a poor load response. It clearly required

more time to fully deliver the new load than did the optimal control strategy.
The maintenance of a fixed period required five periods whereas the optimal
control achieved setpoint in approximately two periods. When "real time" is
considered, rather than "periods", the difference is even more dramatic. The
optimal controller meets setpoint within 310 seconds, while the fixed-period
control requires 3000 seconds. This is shown in Figure 8.31.

Another significant difference exists. The eventual on-gas flow-rate for the
conventional control was 5450 gIh. The optimal controller, in contrast,
specified a frugal 4770 gIh on-gas flow. This represents an energy saving of
14% by application of advanced control.

8.6 Economic Benefits

The economic benefits of implementing the long term controller were

suggested in the previous section. The long term controller made a 14%

240
Blart Furnace Stove Control Chapter 8

-.......
(.)

Cl..
II)
41
40 - ---
~---- - - ~ ""
~~

.
I.iJ
~ 39
::::>
~
I.iJ
38
Cl..
::::E 37
~
36
~
-
.s=
~
.......
CXl

6000
35

£--
Cl..

-- -
II)

~ 5000 ~ :::: :::: "" ::::

~
~

~
~

r§ 4000
Z
a
I ... ==
3000
0 1000 2000 3000 4000 5000 6000 7000 8000
TIME (SECONDS)
0 OPTIMAL v CONVENTIONAL

Figure 8.31 Optimal and Conventional (Fixed-Period) Control Performance -

Real Time
reduction in energy usage and an order of magnitude increase in the speed of
response. Of course, these benefits have been demonstrated on the pilot scale
apparatus and require future verification on full scale stoves. The economic
value associated with these improvements requires knowledge of the cost of
energy, and on the effect of quicker delivery of blast air at the specified
temperature and flowrate has on the iron quality and quality produced in the

241
Blast Fumace Stove Control Chapter 8

blast furnace. These factors are likely to be site specific.

The benefits of the improved short term control are harder to quantify.
Again, the effects of "smoother" delivery of the blast air requirements to be
blast furnace will have an effect on the iron produced in the blast furnace.
This requires quantification. Secondly, the very success of the long-term
controller is dependent on the ability of the short-term controller to follow
setpoint changes. Degradation in the quality of the short term control will
erode the benefits to be obtained from the optimization procedure. In this
light, the short term controller may be considered as a necessary pre-cursor to
achieving the long term control benefits.

8.7 llefeleDCeS

Bateman I. (1992) Optimal Control of a Copper Anode Smelter. Ph.D

Thesis, Department of Chemical Engineering, The University of
Queensland.
Beets J., Elshout J. and de Jong G. (1977). Computer Control of a Hot Blast
Stove System. Journal A (Belgium) 18,1:31-37
Box M.J. (1965) A New Method of Constrained Optimization and a
Comparison with Other Methods. Computer Journal 8:42-52
Jeffreson C.P. (1979) Feedforward Control of Blast Furnace Stoves.
Automatica Vol 15:149-159
Mitter G., Delavos G. and Conert W. (1981) Computer Control and
Optimization of a Blast Furnace Stove Operation. Iron and Steel
Engineer 58,10:43-50

242
Blast Fumace Stove Control Chapter 8

Nose K., Takemura M. and Morita T. (1984) Systems Engineering Approach

to the Optimisation of Hot Blast Stoves. IFAC Proc. Series 1984 5:47-54
Smith C.A. and Corripio A.B. (1985) Principles and Practice of Automatic
Process Control. Wiley
Turnbull Control Systems (1988) Reference Manual. Sussex, England.

243
 CHAPTER 9

CONCLUSIONS

One of the aims of this monograph was to illustrate, through a number of

industrial problems, the benefits of applying modem model-based control.
The examples presented in this work have clearly demonstrated the
applicability of one such modem control strategy, namely Generic Model
Control (GMC). While it can always be argued that a particular problem
brings a unique set of circumstances, this work illustrates that a wide variety
of different control problems can be "solved" using a flexible control
algorithm. It hopefully does meet industrial practitioners' needs of providing
reference examples to give others confidence that modem "theory" can indeed
be successfully applied.

The monograph has also illustrated implementation issues that have to be

considered to arrive at a successful conclusion. The narrative style of chapter
seven clearly illustrates that at times perseverance and sometimes faith is
required. Rarely does an application work the first time. Chapter three
demonstrates that the 90: 10 principle can also apply - 90% of the return for
10% of the effort. Further refinement of the results obtained ~e probably
possible, but are they justifiable?
Conclusions Chapter 9

The degree of sophistication of implementation hardware and software

required to implement GMC has been shown to be modest. The pH
controller of chapter three was implemented in an on-line computer, albeit a
small machine, using standard commercial software. The blast furnace stove
controller of chapter eight, and the distillation column controller of chapter
four were implemented using PC computers, while the extruder controller of
chapter six used a commercial PLC.

Operator acceptance of advanced control schemes is very important. Chapter

four discussed this point and illustrated many practical considerations
required to obtain this vital component. The ability of the DRYER MASTER
controller, described in chapter five, to achieve operator acceptance can be
inferred from the commercial success of the company in marketing this
product.

Each individual case study has highlighted the superior performance of the
model based controller. Substantial benefits have been illustrated, with a
payback time of six months for the dIyer controller of chapter five,
SUS2S0,OOO to SUS2,OOO,OOO for examples in distillation control in chapter
four, and SCDN2,OOO,OOO for the reactor control described in chapter seven.
These are substantial savings and clearly the companies involved have reaped
large rewards.

One of the advantages of GMC illustrated through this monograph is its

ability to use a variety of model types. Chapter seven used an empirical
model, chapter four illustrated semi-empirical models based on approximate

246
Conclusions Chapter 9

dynamic representations and semi-rigorous steady-state descriptions, while

chapter 8 took a more fundamental approach and derived a model from first
principles. The ability to use a variety of model types, which all must
contain various degrees of inaccuracies, improves the flexibility of the
method and enhances its usability.

The monograph has also illustrated three other issues. Firstly, the ability to
control processes in the presence of constraints was illustrated in chapters
four and eight. In particular, chapter eight showed a clear method of not
only incorporating traditional bound constraints on both the manipulated and
the controlled variables, but also functional constraints representing higher
level operating objectives. Secondly, controlling processes that incorporate
dead-times was illustrated in chapter seven. Finally, the flexible specification
of the desired closed loop performance was demonstrated. Chapter 2 and
other chapters illustrated a time-domain approach while chapter three
illustrated a closed-loop pole placement strategy.

This work has highlighted the need for process models to obtain good
process control. Despite the inherent properties of GMC, ultimately good
performance is dependent on a good model. A modelling methodology that
assists practitioners in obtaining reliable process models would greatly
advance the practice of modem control theory.

Finally this monograph was not meant to be a comparison of different

modem control methods, concluding that technique A was better than
technique B. It was meant, as has been stated earlier, to provide some

247
Conclusions Chapter 9

further reference examples for industrial practitioners to "take heart and sally
forth". Others may care to provide detailed comparisons.

248
INDEX

Adapmtion 157, 158, 160, 163, 167-9 Feedback 11, 31, 37, 56, 64, 78, 88,
Adaptive 40, 44, 45, 65, 151, 157, 109, 110, 112, 113, 175, 183, 189,
170, 172-5 . 193
Algebraic Models 79,80,81 Feedforward 11, 17, 19,24, 37, 56,
Anaerobic Digestor 14, 15, 23 60, 64, 73, 78, 79, 101, 112, 113,
140, 141, 144, 191, 192, 222, 228,
Constraints 4,21,25-7,29-31,36,72, 229,242
88, 89, 92, 110, 152, 247
Control Valve 46, 56, 181, 183, 205 Gain 71, 75, 77-81, 84, 93, 94, 136,
138, 141, 186, 187, 214
Deadtime 21,51, 77, 93, 131, 137, Gains 45, 56, 63, 69-71, 74, 80, 93,
153-5,213 101, 140-2, 144, 186, 187
Decoupling 2, 4, 7, 37, 72, 73, 78, 79,
101, 140, 141, 144, 222 lAB 144
Differential Geometry 3, 7 Insmbility 34
Disturbance Variables 10,24, 153, 165, Integral Action 53, 141
173 Interactions 2, 184, 186,218-23,227,
Disturbances 8, 11, 12, 31, 32, 43, 56, 228
61, 111, 133, 143, 157, 163, 172, ITAB 215, 216, 218, 219, 222, 223,
173, 191, 192, 228 224,228
Dynamic Model 8, 21, 44, 48, 76, 85,
86, 120, 197 Minimum Variance 154, 155, 159, 168,
173
Economic Benefits 67, 99, 106, 110, Model Error 34
126, 127, 151, 168,240 Moisture Control 105, 106, 111-4, 116,
Economics 172, 173 124
Efficiency 27, 30, 69, 71, 73, 75, 79, Multivariable 2,28,37,41, 131, 134,
83, 85, 107, 123, 128, 180, 183-7, 139, 141, 149, 152, 186
196, 198, 199, 200-3, 211, 231-5,
237,240 Noise 32, 136, 138, 143, 145, 154, 163
Energy I, 67, 84, 88, 89, 99, 106, Noise-free 136
107, 110,111, 123-6, 128, 185, 188, Noisy 151
194,195,234,240,241
 Index

Nonlinear 2, 4, 7-9, 11, 16, 22-5, 28, Robustness 31, 144

30-2,37,39-41,43-6,48,57,59,60, Rules 173
61-5,69,71-6,79-81,85,89-91,93,
95-7, 99, 101-3, 151-3, 156, 162, Sampling Interval 31, 142, 192, 194
167, 173, 174 Sensor 105, 108, 110, 112, 131, 138,
Nonlinearity 69, 70, 73, 101, 167 143, 181, 188, 194, 195, 204, 205
Setpoint 3, 8, 9, 11, 18, 20, 24, 25,
Offset 9, 20,27, 137, 139 27,48, 57, 59, 61, 78, 118, 120,
Operators 56,57, 61, 80, 90-2, 106, 124, 125, 140, 141, 147, 160, 163,
108-11, 116 186, 189, 192, 198, 200, 201, 205,
Optimization 2, 25, 28, 30, 39, 101, 207, 214, 216, 217, 218-26, 228,
152, 195, 196, 210, 242 229, 234, 236, 237, 240, 242
Overshoot 17,57, 140, 141, 160, 162, Simulation 2, 16, 18, 30, 36, 37, 41,
163, 224, 234 65, 86, 102, 115, 136, 140, 142-7,
211
pH 4, 38,43-6,49-51,53,54,56,57, Software 101, 105, 121, 122, 204-6,
59, 61, 63, 65, 242, 246 210,246
PI 69, 73, 74, 77, 93, 96, 98, 140-6 SPC 94,95, 154
PID 45, 69, 71, 72, 114, 142, 189, SPQ 72,95
191, 192, 204, 205, 207, 213, 214, SRK 86
217-20,222,223-6,228 Stability 31,32, 35, 42, 133, 184-7,
Pole Placement 53, 247 196, 200-3,211,230, 231, 234, 235,
PRBS 165, 166, 168, 175 237
Product Quality 129, 153, 155, 156, Steady-state Model 23, 24, 76-8, 221
159-61, 164, 169-71, 173
Throughput 1, 106, 108, 113, 125,
Quality 1,61,72, 123, 127-9, 151-6, 128, 154, 162, 170
155, 157-61, 160, 162-4, 163, 166, Time Constant 16, 22, 23, 72, 78, 79,
168, 169, 170, 171, 173-5,241,242 88,93,120,136,153,162,163,174,
188
Reference Trajectory 16, 18,26, 31, Transfer Function 13
33,34, 121 Tuning 34, 44, 45, 52, 53, 55-7, 63,
Relative Gain Array 138, 186, 214 65, 73, 79, 80, 86, 88, 93, 96, 101,
Reset Windup 57 108, 157, 158, 159, 160, 162, 163,
Rise Time 141, 142 168, 173-5, 213, 214, 220
Robust 12, 20, 31, 34, 35, 42, 174,
193, 212