Practical Distillation Control

Practical Distillation Control

William L. Luyben, Editor
InmiI VAN NOSTRAND REINHOLD
____ New York
Copyright 1992 by Van Nostrand Reinhold
Softcover reprint of the hardcover 1st edition 1992
Library of Congress Catalog Card Number: 92-10642
ISBN 978-1-4757-0279-8 ISBN 978-1-4757-0277-4 (eBook)
DOl 10.1007/978-1-4757-0277-4
All rights reserved. No part of this work covered by the
copyright hereon may be reproduced or used in any form or
by any means-graphic, electronic, or mechanical, including
photocopying, recording, taping, or information storage and
retrieval systems-without written permission of the
publisher.
Manufactured in the United States of America.
Published by Van Nostrand Reinhold
115 Fifth Avenue
New York, New York 10003
Chapman and Hall
2-6 Boundary Row
London, SE 1 8HN, England
Thomas Nelson Australia
102 Dodds Street
South Melbourne 3205
Victoria, Australia
Nelson Canada
1120 Birchmount Road
Scarborough, Ontario MIK 5G4, Canada
16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1
Ubrary 01 Congress Catalotling-in-Publication Data
Practical distillation control/edited by William L. Luyben
p. Ctn.
Includes bibliographical references and index.
ISBN 978-1-4757-0279-8
1. Distillation apparatus.
I. Luyben, William L.
TP159.D5P73 1992
66O'.28425-dc20
2. Chemical process control.
92-10642
CIP
John E. Anderson
Hoechst Celanese
Corpus Christi, TX 78469
(Chapter 19)
Page S. Buckley
Consultant
Newark Delaware 19713
(Chapter 2)
James J. Downs
Advanced Controls Technology
Eastman Chemical Company
Kingsport, TN 37662
(Chapter 20)
James G. Gerstle
Amoco Corporation
Naperville, Illinois 60566
(Chapter 12)
Vincent G. Grassi II
Air Products and Chemicals, Inc.
Allentown, PA 18195-1501
(Chapters 3 and 18)
Kurt E. Haggblom
Process Control Laboratory
Abo Akademi
20500 Abo, Finland
(Chapter 10)
David A. Hokanson
Exxon Chemicals
H. R. Rotterdam No. 231768
The Netherlands
(Chapter 12)
Henk Leegwater
DSM
6190 AA Beek
The Netherlands
(Chapter 16)
Contributors
William L. Luyben
Department of Chemical Engineering
Lehigh University
Bethlehem, PA 18015
(Chapters 1, 11, 22, 24, and 25)
Randy C. McFarlane
Amoco Corporation
Amoco Research Center
Naperville, IL 60566
(Chapter 7)
Charles Moore
Department of Chemical Engineering
University of Tennessee
Knoxville, TN 37996
(Chapter 8)
Cristian A. Muhrer
Air Products and Chemicals, Inc.
Allentown, PA 18195-1501
(Chapters 23 and 25)
Antonis Papadourakis
Rohm and Haas Co.
Bristol, P A 19007
(Chapter 4)
Ferdinand F. Rhiel
Corporate Division of Research & Development
Bayer AG
D-5090 Leverkusen, Germany
(Chapter 21)
John E. Rijnsdorp
University of Twente
7500AE Enschede
Netherlands
(Chapter 4)
Daniel E. Rivera
Department of Chemical Engineering
Arizona State University
Tempe, Arizona 85287
(Chapter 7)
v
vi Contributors
F. Greg Shinskey
The Foxboro Co.
Foxboro, MA 02035
(Chapter 13)
Sigurd Skogestad
Chemical Engineering
University of Trondheim, NTH
N-7034 Trondheim, Norway
(Chapter 14)
Terry L. Tolliver
Monsanto Co.
St. Louis, MO 63167
(Chapter 17)
Bjorn D. Tyreus
Engineering Department
E. I. DuPont de Nemours & Co.
Wilmington, DE 19898
(Chapters 5 and 9)
Ernest F. Vogel
Advanced Control Technology
Tennessee Eastman Co.
Kingsport, TN 37662
(Chapter 6)
Kurt Waller
Process Control Laboratory
Abo Akademi
20500 Abo, Finland
(Chapters 10 and 15)
This book is dedicated to Bea, Gus, and Joanne Luyben and Bill Nichol, four of
the most avid bridge players I have ever known. Run 'em out!
Preface
Distillation column control has been the
subject of many, many papers over the last
half century. Several books have been de-
voted to various aspects of the subject. The
technology is quite extensive and diffuse.
There are also many conflicting opinions
about some of the important questions.
We hope that the collection under one
cover of contributions from many of the
leading authorities in the field of distillation
control will help to consolidate, unify, and
clarify some of this vast technology. The
contributing authors of this book represent
both industrial and academic perspectives,
and their cumulative experience in the area
of distillation control adds up to over 400
years! The collection of this wealth of expe-
rience under one cover must be unique in
the field. We hope the readers find it effec-
tive and useful.
Most of the authors have participated at
one time or another in the Distillation Con-
trol Short Course that has been given every
two years at Lehigh since 1968. Much of the
material in the book has been subjected to
the "Lehigh inquisition" and survived! So it
has been tested by the fire of both actual
plant experience and review by a hard-nosed
group of practically oriented skeptics.
In selecting the authors and the topics,
the emphasis has been on keeping the ma-
terial practical and useful, so some subjects
that are currently of mathematical and the-
oretical interest, but have not been demon-
strated to have practical importance, have
not been included.
The book is divided about half and half
between methodology and specific applica-
tion examples. Chapters 3 through 14 dis-
cuss techniques and methods that have
proven themselves to be useful tools in at-
tacking distillation control problems. These
methods include dynamic modelling, simu-
lation, experimental identification, singular
value decomposition, analysis of robustness,
and the application of multivariable meth-
ods. Chapters 15 through 25 illustrate how
these and how other methods can be ap-
plied to specific columns or important
classes of columns.
ix
The use in this book of trademarks, trade names, general descriptive names, and
so forth, even if they are not specifically identified, should not be taken as
indication that such names, as understood by the Trade Marks and Merchandise
Act, may be freely used by anyone.
Contents
Preface
ix
Part 1 Techniques and Methods 1
1 Introduction, William L. Luyben 3
1-1 Importance of Distillation in Industry
3
1-2 Basic Control
3
1-2-1 Degrees of Freedom 3
1-2-2 Fundamental Variables for Composition Control 5
1-2-3 Pressure Control 6
1-2-4 Level Control 8
1-3 Uniqueness of Distillation Columns 9
1-4 Interaction between Design and Control 10
1-4-1 Increasing Column Size 10
1-4-2 Holdups in Column Base and Reflux Drum 11
1-4-3 Effects of Contacting Devices 11
1-4-4 Sensors 11
1-5 Special Problems 11
1-5-1 High-Purity Products 11
1-5-2 Small Temperature Differences 12
1-5-3 Large Temperature Differences 12
1-5-4 Gravity-Flow Reflux 12
1-5-5 Dephlegmators 13
1-6 Conclusion 13
References 13
2 Historical Perspective, Page S. Buckley 14
2-1 Introduction 14
2-2 What is Control? 14
2-3 Column Design Methods 15
2-4 Column Tray and Auxiliary Design 15
2-4-1 Tray Design 15
2-4-2 Reboiler Design 16
2-4-3 Flooded Reboilers 16
2-4-4 Column Base Designs 16
2-4-5 Overhead Design 16
2-5 Instrumentation 17
2-6 Control System Design Methods 19
2-7 Process Control Techniques 21
2-8 Influence on Distillation Control 22
2-9 Conclusion 23
xiii
XlV Contents
3
4
2-10 Comments on Reference Texts
References
Part 2 Methods
Rigorous Modelling and Conventional Simulation, Vincent G. Grassi II
3-1 Overview
3-1-1 Conventional Simulation
3-2 Distillation Process Fundamentals
3-2-1 Continuity Equations
3-2-2 Vapor-Liquid Equilibrium
3-2-3 Murphree Vapor Phase Stage Efficiency
3-2-4 Enthalpy
3-2-5 Liquid and Froth Density
3-3 Computer Simulation
3-3-1 Algebraic Convergence Methods
3-3-2 Equilibrium Bubble Point Calculation
3-3-3 Equilibrium Dew Point Calculation
3-3-4 Distillation Stage Dynamic Model
3-3-5 Bottom Sump
3-3-6 Condenser
3-3-7 Reflux Accumulator
3-3-8 Feedback Controllers
3-4 Writing a Dynamic Distillation Simulator
3-5 Plant-Model Verification
3-6 Computational Performance
3-7 Conclusions
3-8 Nomenclature
References
Approximate and Simplified Models, Antonis Papadourakis and
John E. Rijnsdorp
4-1 Introduction
4-2 Classification of Simple Models
4-3 Simple Steady-State Models
4-4 Partitioning of the Overall Dynamic Model
4-4-1 Introduction
4-4-2 Assumptions
4-4-3 Propagation of Vapor Flow and Pressure Responses
4-4-4 Propagation of Liquid Flow and Liquid Holdup Variations
4-4-5 Propagation of Vapor and Liquid Concentrations
4-5 Linear Models
4-5-1 Linear Models in the Time Domain
4-5-2 Linear Models in the Laplace Domain
4-5-3 Linear Models in the Frequency Domain
4-6 Nonlinear Models
4-6-1 Simplifying Assumptions
4-6-2 Number of Components
4-6-3 Number of Stages-Orthogonal Collocation
References
24
24
27
29
29
30
32
32
34
35
36
36
37
37
38
39
39
41
41
42
43
43
44
45
45
46
46
48
48
48
49
50
50
52
53
53
54
55
55
55
59
61
61
62
62
69
Contents xv
5 Object..()riented Simulation, B. D. Tyreus 72
5-1 Introduction 72
5-2 Ideal Dynamic Simulator 73
5-3 Object-Oriented Programming 74
5-3-1 Classes and Objects 75
5-3-2 Modelling a Column Tray 75
5-3-3 Inheritance and Polymorphism 79
5-4 Distillation Column Simulation with Object-Oriented Programming 81
5-4-1 Structured Models 81
5-4-2 Structured Models and Object-Oriented Programming 82
5-5 Experience in Using Object-Oriented Simulation for Distillation 83
5-6 Conclusion 84
References 84
6 Plantwide Process Control Simulation, Ernest F. Vogel 86
6-1 Introduction 86
6-2 Applications of a Plantwide Process Control Simulator 86
6-2-1 Process Control 88
6-2-2 Process Design 89
6-2-3 Process Safety 89
6-2-4 Example 90
6-3 Benefit from Plantwide Process Control Simulation 91
6-4 Defining the Scope of a Plantwide Process Simulation 91
6-5 Building a Plantwide Process Simulation 92
6-5-1 Programming Environment 92
6-5-2 Equation Solving Environment 92
6-5-3 Steady-State-Dynamic Flowsheet Simulation Environment 93
6-5-4 Practical Considerations 93
6-6 Features of a Plantwide Process Simulator for Control Strategy Design 94
References 95
7 Identification of DistiUation Systems, R. C. McFarlane and D. E. Rivera 96
7-1 Introduction 96
7-1-1 Discrete Transfer Function Models for Distillation Systems 97
7-1-2 Iterative Methodology of System Identification 98
7-2 Perturbation Signal Design 99
7-2-1 Discussion 99
7-2-2 Pseudo-random Binary Sequence Signals 100
7-3 Model Structure Selection and Parameter Estimation 103
7-3-1 Bias-Variance Trade-Offs in System Identification 103
7-3-2 Nonparametric Methods 105
7-3-3 Parametric Models 106
7-3-4 Identification for Control System Design 109
7-3-5 Identifiability Conditions for Closed-Loop Systems 115
7-3-6 Treatment of Nonlinearity 117
7-4 Model Validation 119
7-4-1 Classical Techniques 119
7-4-2 Control-Relevant Techniques 120
7-5 Practical Considerations 122
7-6 Example 123
7-7 Nomenclature 136
References 138
xvi Contents
8 Selection of Controlled and Manipulated Variables, Charles F. Moore 140
8-1 Introduction 140
8-2 Sensor and Valve Issues 140
8-2-1 Inventory Control Concerns 140
8-2-2 Separation Control Concerns 142
8-2-3 Loop Sensitivity Issues 142
8-3 Location of Temperature Sensors 145
8-3-1 Determining Temperature Sensitivities 145
8-3-2 Selecting a Temperature Sensor for Single-Ended Control 147
8-3-3 Selecting Temperature Location for Dual-Ended Control 148
8-4 Selecting Sensor Type: Temperature versus Composition 157
8-4-1 Limitation of Temperature Sensors 157
8-4-2 Operational Concerns with Using Process Analyzer 159
8-4-3 Schemes for Using Analyzers in Distillation Control 160
8-4-4 Analyzer Resolution Requirements versus Location 161
8-4-5 Determining Composition Sensitivities 162
8-4-6 Selecting an Analyzer Location for Single-Ended Control 162
8-4-7 Selecting Analyzer Locations and Focus for Dual-Ended Control 165
8-5 Other Roles for Column Analyzers 169
8-5-1 Feedforward Control 169
8-5-2 Recycle Inventory Control 170
8-5-3 Measuring and Documenting Variation 170
8-6 Selecting Manipulated Variables 170
8-6-1 Steady-State Considerations 171
8-6-2 Dynamic Considerations 173
8-6-3 Plantwide Considerations 174
8-7 Summary and Conclusions 176
References 177
9 Selection of Controller Structure, B. D. Tyreus 178
9-1 Introduction 178
9-1-1 Control Design Principles 179
9-2 Manipulative Variables 179
9-2-1 Manipulative Variables and Degrees of Freedom 181
9-3 A Methodology for Selection of Controller Structure 183
9-3-1 Level and Pressure Controls 184
9-3-2 Composition Controls 184
9-3-3 Optimizing Controls 185
9-4 Examples 185
9-4-1 A Column with a Stripping Section Sidestream 185
9-4-2 A Column with a Rectifying Section Sidestream 188
9-5 Conclusion 191
References 191
10 Control Structures, Consistency, and Transformations, Kurt E. Hiiggblom
and Kurt V. Waller 192
10-1 Introduction 192
10-2 Some Basic Properties of Distillation Control Structures 194
10-2-1 Energy Balance Structure (L, V) 194
10-2-2 Material Balance Structures (D, V) and (L, B) 196
Contents xvii
10-3 Consistency Relations 198
10-3-1 (L, V) Structure 198
10-3-2 (D, V) Structure 199
10-3-3 (L, B) Structure 200
10-4 Transformations between Control Structures 201
10-4-1 Transformation from (L, V) to (D, V) 201
10-4-2 Transformation from (L, V) to (L, B) 203
10-5 Control Structure Modelling-the General Case 204
10-5-1 Compact Description of Control Structures 204
10-5-2 Consistency Relations 205
10-5-3 Transformations between Arbitrary Structures 206
10-5-4 Complex Distillation Columns 207
10-6 Application 1: Numerical Examples of Control Structure Transformations 208
10-6-1 (D, V) Structure 209
10-6-2 (LID, V) Structure 209
10-6-3 (LID, VIF) Structure 210
10-6-4 (LID, VIB) Structure 210
10-7 Application 2: Use of Consistency Relations in Transformations 211
10-8 Application 3: Process Dynamics 212
10-9 Application 4: Identification of Consistent Models 214
10-9-1 Reconciliation of Control Structure Models 215
10-9-2 Numerical Example 216
10-10 Application 5: Relative Gain Analysis 219
10-10-1 Some Analytical Relations between Relative Gains 220
10-10-2 Numerical Example 221
10-11 Application 6: Synthesis of Decoupled Control Structures by Transformations
of Output Variables 221
10-11-1 Derivative of Output Transformations 222
10-11-2 Numerical Examples 223
10-11-3 Discussion of Output Decoupling Structures Suggested in the Literature 224
10-12 Application 7: A Control Structure for Disturbance Rejection and Decoupling 225
Acknowledgment 226
References 227
11 Diagonal Controller Tuning, William L. Luyben 229
11-11 Introduction 229
11-1-1 The Problem 230
11-1-2 Alternatives 230
11-1-3 LACEY Procedure 233
11-1-4 Nomenclature 233
11-2 Selection of Controlled Variables 235
11-3 Selection of Manipulated Variables 235
11-3-1 Morari Resiliency Index 235
11-3-2 Condition Number 237
11-4 Tuning Diagonal Controllers in a Multivariable Environment 237
11-4-1 Review of Nyquist Stability Criterion for SISO Systems 237
11-4-2 Extension to MIMO Systems 238
11-4-3 BLT Tuning Procedure 239
11-4-4 Examples 239
11-5 Pairing 243
11-5-1 Elimination of Unworkable Pairings 244
11-5-2 Tyreus Load Rejection Criterion 246
11-6 Conclusion 247
xviii Contents
12 Dynamic Matrix Control Multivariable Controllers, David A. Hokanson and
James G. Gerstle
12-1 Introduction
12-2 Basics of DMC Mathematics
12-2-1 Convolution Models
12-2-2 Prediction Errors
12-2-3 Control Solution
12-2-4 Move Suppression
12-3 Review of Model Identification
12-3-1 DMC Model Identification Background
12-3-2 Integrating Process Model Identification
12-3-3 Multivariable Model Identification
12-3-4 Nonlinear Transformations
12-4 Design Aspects of a Multivariable DMC Controller
12-4-1 Weights
12-4-2 Constraints
12-5 Implementation Steps for a DMC Controller
12-5-1 Initial Design
12-5-2 Pretest
12-5-3 Plant Test
12-5-4 Model Identification
12-5-5 Controller Building and Simulation
12-5-6 Controller and Operator Interface Installation
12-5-7 Controller Commissioning
12-5-8 Measuring Results
12-6 DMC Applications on Industrial Towers
12-6-1 Hydrocracker C
3
- C
4
Splitter
12-6-2 Hydrocracker Preftash Column
12-6-3 Benzene and Toluene Towers
12-6-4 Olefins Plant Demethanizer
12-6-5 Olefins Plant C
2
Splitter
12-7 Summary
References
13 DistiUation Expert System, F. G. Shinskey
13-1 Introduction
13-1-1 On-line versus Off-line Systems
13-1-2 Expertise in a Knowledge Domain
13-1-3 Logical Rule Base
13-1-4 First Principles and Mathematical Modelling
13-2 Configuring Distillation Control Systems
13-2-1 Nonlinear Multivariable System
13-2-2 Relative Gain Analysis
13-2-3 Establishing a Performance Index
13-2-4 Applications and Objectives
13-3 Rule Base for Simple Columns
13-3-1 Economic Objective
13-3-2 Maximizing Recovery
13-3-3 Controlling Both Product Compositions
13-3-4 Floating Pressure Control
248
248
249
249
253
253
254
255
255
256
257
258
260
260
261
261
262
263
263
265
266
267
267
267
268
268
268
268
269
270
270
271
272
272
272
273
273
274
275
275
277
279
283
284
284
285
286
287
Contents xix
13-4 Rule Base for Sidestream Columns 288
13-4-1 Classifications 288
13-4-2 Configuration Rules 288
13-5 Nomenclature 290
References 290
14 Robust Control, Sigurd Skogestad 291
14-1 Robustness and Uncertainty 291
14-2 Traditional Methods for Dealing with Model Uncertainty 292
14-2-1 Single-Input-Single-Output Systems 292
14-2-2 Multi-Input-Multi-Output Systems 293
14-3 A Multivariable Simulation Example 294
14-3-1 Analysis of the Model 294
14-3-2 Use of Decoupler 296
14-3-3 Use of Decoupler When There is Model Uncertainty 297
14-3-4 Alternative Controllers: Single-Loop PIO 298
14-3-5 Alternative Configurations: DV Control 299
14-3-6 Limitations with the Example: Real Columns 301
14-4 RGA as a Simple Tool to Detect Robustness Problems
301
14-4-1 RGA and Input Uncertainty 301
14-4-2 RGA and Element Uncertainty/Identification 302
14-5 Advanced Tools for Robust Control: IL Analysis 303
14-5-1 Uncertainty Descriptions 304
14-5-2 Conditions for Robust Stability 305
14-5-3 Definition of Performance 307
14-5-4 Conditions for Robust Performance 307
14-6 Nomenclature 308
References 309
Part 3 Case Studies 311
15 Experimental Comparison of Control Structures, Kurt V. Waller 313
15-1 Introduction 313
15-2 Manipulator Choice for Decentralized Control 314
15-3 Experimental Apparatus 315
15-4 Mixture Distilled 316
15-5 Control Structures Studied 316
15-6 General Comments on the Experiments 318
15-7 (D, V), (V, D), and (L, B) Structures 320
15-8 (D/(L + D), V) Structure 321
15-9 (D /(L + D), V /B) Structure 322
15-10 Comparison of Four Conventional Control Structures 324
15-10-1 One-Point Control 324
15-10-2 Two-Point Control 325
15-11 Structure for Disturbance Rejection and Decoupling 326
15-12 Controller Tuning for Robustness against Nonlinearities 328
15-13 Summary and Conclusions 328
Acknowledgment 329
References 329
xx Contents
16 Industrial Experience with Double Quality Control, Henk Leegwater 331
16-1 Introduction 331
16-2 Quality Control 331
16-3 Single Quality Control 331
16-4 Why Double Quality Control? 332
16-4-1 Energy Consumption Versus Degradation of Valuable Product 332
16-4-2 Optimizing Throughput 333
16-5 Quality Measurements 334
16-6 Why Multivariable Control for Double Quality Control? 336
16-7 Heat and Material Balance in Relation to Separation 337
16-8 Net Heat Input 338
16-9 Separation Indicators 339
16-9-1 Separation Performance Indicator 339
16-9-2 Separation Accent Indicator 339
16-9-3 Interpretation 339
16-10 Development of the Control Scheme for the C2 Splitters 340
16-10-1 Characterization of the Column Operation 340
16-10-2 Column Simulation 340
16-11 Introduction of the New Control Schemes into Industrial Practice 341
16-12 C2 Splitter with Heat Integration 341
16-12-1 Basic Controls 342
16-12-2 Quality Measurements 342
16-12-3 Previous Control Scheme 342
16-12-4 Improved Control Scheme 342
16-12-5 Condenser-Reboiler Level Control 343
16-12-6 Experiences 346
16-13 C2 Splitter without Heat Integration 347
16-13-1 Basic Controls 347
16-13-2 Quality Measurements 347
16-13-3 Previous Control Scheme 347
16-13-4 Interaction 347
16-13-5 Improved Control Scheme Using Qnet/F and the Separation Factors 348
16-13-6 Experiences 349
16-14 Conclusion 350
17 Control of Distillation Columns via Distributed Control Systems,
T. L. Tolliver 351
17-1 Introduction
17-1-1 Historical Background on Distributed Control Systems
17-1-2 Advantages over Analog Instrumentation
17-1-3 Future Trends
17-2 Case I
17-2-1 Process Background-Debottlenecking
17-2-2 Control Scheme Design
17-2-3 Implementation Details
17-2-4 Results
17-3 Case II
17-3-1 Process Background-Reduced Capital
17-3-2 Control Scheme Design
17-3-3 Implementation Details
17-3-4 Results
17-4 Case III
17-4-1 Process Background-Energy Conservation
17-4-2 Control Scheme Design
351
351
352
353
353
353
355
357
359
359
359
359
361
363
364
364
364
Contents xxi
17-4-3 Implementation Details 366
17-4-4 Results 367
17-5 Summary 368
References 368
18 Process Design and Control of Extractive Distillation, Vmcent G. Grassi II 370
18-1 Overview 370
18-1-1 Extractive and Azeotropic Distillation 371
18-1-2 History 373
18-1-3 Process Description 374
18-2 Phase Equilibria 375
18-2-1 Relative Volatility 375
18-2-2 Residue Curves 377
18-3 Process Design 380
18-3-1 Degrees of Freedom 380
18-3-2 Process Design Procedure 382
18-3-3 Total Cost Relations 386
18-4 Process Control 387
18-4-1 Control System Economics 388
18-4-2 Measured Variables 392
18-4-3 Extraction Tower Nonlinearities 394
18-4-4 Extraction Tower Open Loop Dynamics 398
18-4-5 Control Schemes 398
18-5 Conclusions 403
References 403
19 Control by Tray Temperature of Extractive Disdlladon,
John E. Anderson 405
19-1 Situation
19-2 Analysis
19-3 Solution
19-4 Conclusions
405
405
408
412
20 Disdlladon Control in a Plantwlde Control Environment, James J. Downs 413
20-1 Introduction
20-2 Plantwide Component Inventory Control
20-2-1 Component Inventory Control for a Tank
20-2-2 Component Inventory Control for a Process
20-3 Acetaldehyde Oxidation Process Case Study
20-3-1 Description of the Problem
20-3-2 Low Boiler Column Analysis
20-3-3 Component Inventory Control Analysis
20-3-4 Control of the Acetaldehyde Oxidation Process
20-4 Conclusions
21 Model-Based Control, F. F. Rhiel
21-1 Introduction
21-2 Design of the Control Concept
21-2-1 Process Description
413
414
414
417
423
423
427
432
438
439
440
440
440
440
xxii Contents
21-2-2 Design of the Observer Model 442
21-2-3 Control Concept 444
21-3 Results of Model-Based Control 444
21-3-1 Observer and Controller Behaviour 444
21-3-2 Introducing the New Control Concept in the Production Plant 445
21-4 Comparison between Conventional and New Control Concept 446
21-5 Conclusions 449
21-6 Appendix 449
References 450
22 Superfractionator Control, William L. Luyben 451
22-1 Occurrence and Importance 451
22-2 Features 451
22-2-1 Many Trays 452
22-2-2 High Reflux Ratios 452
22-2-3 Flat Temperature Profile 452
22-2-4 Slow Dynamics 452
22-3 Alternative Control Structures 453
22-4 Industrial Example 455
22-4-1 Process 455
22-4-2 Plant Dynamic Tests 455
22-4-3 Simulation 455
22-4-4 Results 455
22-5 Tuning the D-B Structure 457
22-5-1 Transformations 457
22-5-2 Example 458
22-5-3 Fragility of D-B Structure 459
22-6 Pitfalls with Ratio Schemes 459
22-7 Superfractionator with Sidestream Example 461
22-7-1 Process 461
22-7-2 Plant Dynamic Tests 462
22-7-3 Steady-State Analysis 462
22-7-4 Simulation Results 466
22-8 Conclusion 467
References 467
23 Control of Vapor Recompression Distillation Columns, Cristian A. Muhrer 468
23-1 Introduction 468
23-2 Design 469
23-3 Dynamics and Control 470
23-4 Alternative Compressor Control Systems 471
23-4-1 Compressor Performance Curves 472
23-4-2 Plant Characteristic Curve 474
23-4-3 Description of Alternative Compressor Controls 475
23-4-4 Dynamic Performance 477
23-5 Case Studies 478
23-5-1 Steady-State Design 478
23-5-2 Dynamic Models 481
23-5-3 Control System Design 482
23-5-4 Results for Specific Systems 488
23-6 Conclusion
23-7 Nomenclature
References
24 Heat-Integrated Columns, Wdliam L. Luyben
24-1 Introduction
24-2 Types of Systems
24-2-1 Energy Integration Only
24-2-2 Energy and Process Integration
24-3 Economic Incentives
24-4 Limitations
24-5 Control Problem
24-6 Total Heat-Input Control
24-7 Incentives for Composition Control of All Products
24-8 Conclusion
References
25 Batch Distillation, Cristian A. Muhrer and Wdliam L. Luyben
25-1 Introduction
25-2 Basic Operations
25-2-1 Process
25-2-2 Composition Profiles
25-2-3 Slop Cuts
25-3 Assessment of Performance: Capacity Factor
25-4 Models
25-4-1 Differential Distillation
25-4-2 Pseudo-Steady-State Models
25-4-3 Rigorous Dynamic Models
25-4-4 Fitting Models to Experimental Batch Distillation Data
25-5 Comparison with Continuous Distillation
25-6 Reflux Ratio Trajectories
25-7 Pressure Trajectories
25-7-1 Constant Pressure
25-7-2 Constant Reflux-Drum Temperature
25-8 Column Design
25-9 Slop Cut Processing
25-9-1 Alternatives
25-9-2 Results
25-10 Inferential Control of Batch Distillation
25-10-1 Problem
25-10-2 Basic Insight
25-10-3 Quasidynamic Model
25-10-4 Extended Luenberger Observer
25-11 Conclusion
References
Index
Contents xxiii
489
490
490
492
492
492
492
498
499
501
502
502
504
507
507
50S
508
508
508
509
510
510
511
511
512
512
513
514
514
515
515
515
515
517
517
518
519
519
519
521
523
525
528
529
1
Techniques and Methods
1
I ntrod uction
William L. Luyben
Lehigh University
1-1 IMPORTANCE OF
DISTILLATION IN INDUSTRY
Despite many predictions over the years to
the contrary, distillation remains the most
important separation method in the chemi-
cal and petroleum industries. Distillation
columns constitute a significant fraction of
the capital investment in chemical plants
and refineries around the world, and the
operating costs of distillation columns are
often a major part of the total operating
costs of many processes. Therefore, the
availability of practical techniques for devel-
oping effective and reliable control systems
for efficient and safe operation of distilla-
tion systems is very important.
These considerations are the basis for
the development of this book. We hope that
the generic techniques and the industrial
case studies presented in these chapters will
help working engineers in their important
task of controlling distillation columns.
Distillation columns present challenging
control problems. They are highly multivari-
able and usually quite nonlinear. They have
many constraints and are subject to many
disturbances. Therefore, their control is not
a trivial task.
1-2 BASIC CONTROL
This section presents some of the funda-
mentals of distillation column control that
are applicable to almost all distillation
columns. Some of the ideas may seem obvi-
ous and trivial, but unless they are kept in
mind, it is often easy to make some funda-
mental mistakes that will result in very poor
control.
We will concentrate on the basic,
"plain-vanilla" simple distillation column
shown in Figure 1-1. The nomenclature is
shown on the figure. There is a single feed,
and two products are produced. Heat is
added in the partial reboiler and removed
in the total condenser. Reflux is added on
the top tray. Trays are numbered from the
bottom.
1-2-1 Degrees of Freedom
In the context of process control, the de-
grees of freedom of a process is the number
of variables that can or must be controlled.
It is always useful to be clear about what
this number is for any process so that you
do not attempt to over- or undercontrol any
process.
3
4 Practical Distillation Control
R
, N,
Z
3
2
B
FIGURE 11. Basic column and nomenclature.
The mathematical approach to finding
the degrees of freedom of any process is to
total all the variables and subtract the num
ber of independent equations. This is an
interesting exercise, but there is a much
easier approach. Simply add the total num
ber of rationally placed control valves. The
"rationally placed" qualification is to em
phasize that we have avoided poorly con
ceived designs such as the placement of two
control valves in series in a liquid filled
system.
In Figure 11 we see that there are five
control valves, one on each of the following
streams: distillate, reflux, coolant, bottoms,
and heating medium. We are assuming for
the moment that the feed stream is set by
the upstream unit. So this simple column
has 5 degrees of freedom. But inventories in
any process always must be controlled. In
ventory loops involve liquid levels and pres
sures. In our simple distillation column ex-
ample, this means that the liquid level in
the reflux drum, the liquid level in the base
of the column, and the column pressure
must be controlled. When we say that the
pressure is controlled, we do not necessarily
mean that it is held constant. If we mini-
mize it, we are "controlling" it.
If we subtract the three variables that
must be controlled from S, we end up with 2
degrees of freedom. Thus, there are two
and only two additional variables that can
(and must) be controlled in this distillation
column. Notice that we have made no as-
sumptions about the number or type of
chemical components we are separating. So
a simple, ideal, binary system has 2 degrees
of freedom; a complex, multicomponent,
nonideal distillation system also has 2 de-
grees of freedom.
The two variables that are chosen to be
controlled depend on many factors. Some
common situations are:
1. Control the composition of the light-key
impurity in the bottoms and the compo-
sition of the heavy-key impurity in the
distillate.
2. Control a temperature in the rectifying
section of the column and a temperature
in the stripping section of the column.
3. Control the flow rate of reflux and a
temperature somewhere in the column.
4. Control the flow rate of steam to the
reboiler and a temperature near the top
of the column.
S. Control the reflux ratio (ratio of reflux
flow to distillate flow) and a temperature
in the column.
These examples illustrate that (a) only
two things can be controlled and (b) nor-
mally at least one composition (or tempera-
ture) somewhere in the column must be
controlled.
Once the five variables to be controlled
have been specified (e.g., two temperatures,
two levels, and pressure), we still have
the problem of deciding what manipulated
variable to use to control what controlled
variable. This "pairing" problem is called
determining the structure of the control sys-
Introduction 5
tem. It will be discussed in many chapters in
this book.
1-2-2 Fundamental Variables for
Composition Control
The compositions of the products from a
distillation column are affected by two fun-
damental manipulated variables: feed split
and fractionation. The feed split (or cut-
point) variable refers to the fraction of the
feed that is taken overhead or out the bot-
tom. The "fractionation" variable refers to
the energy that is put into the column to
accomplish the separation. Both of these
fundamental variables affect both product
compositions but in different ways and with
different sensitivities.
Feed split: Taking more of the feed out of
the top of the column as distillate tends
to decrease the purity of the distillate
and increase the purity of the bottoms.
Taking more of the feed out of the bot-
tom tends to increase distillate purity and
decrease bottoms purity. On a McCabe-
Thiele diagram, assuming a constant re-
flux ratio, we are shifting the operating
lines to the left as we increase distillate
flow and to the right as we decrease
distillate flow.
Fractionation: Increasing the reflux ratio (or
steam-to-feed ratio) produces more of a
difference between the compositions of
the products from the column. An in-
crease in reflux ratio will reduce the im-
purities in both distillate and bottoms.
Feed split usually has a much stronger
effect on product compositions than does
fractionation. This is true for most distilla-
tion columns except those that have very
low product purities (less than 90%).
One of the important consequences of
the overwhelming effect of feed split is that
it is usually impossible to control any com-
position (or temperature) in a column if the
feed split is fixed, that is, if the distillate or
bottoms flows are held constant. Any small
6 Practical Distillation Control
changes in feed rate or feed composition
will drastically affect the compositions of
both products, and it will not be possible to
change fractionation enough to counter this
effect. A simple example illustrates the
point. Suppose we are feeding 50 mol of
component A and 50 mol of component B.
Distillate is 49 mol of A and 1 mol of B;
bottoms is 1 mol of A and 49 mol of B.
Thus product purities are 98%. Now sup-
pose the feed changes to 40 mol of A and
60 mol of B, but the distillate flow is fixed at
a total of 50 mol. No matter how the reflux
ratio is changed, the distillate will contain
almost 40 mol of A and 10 mol of B, so its
purity cannot be changed from 80%.
The fundamental manipulated variables
(fractionation and feed split) can be changed
in a variety of ways by adjustment of the
control valves that set distillate, reflux, bot-
toms, steam, and cooling water flow rates.
Fractionation can be set by adjusting reflux
ratio, steam-to-feed ratio, reflux-to-feed ra-
tio, and so forth. Feed split can be set
directly by adjusting distillate or bottoms
flows or indirectly by adjusting reflux or
steam and letting level controllers change
the product streams.
12-3 Pressure Control
Pressure in distillation columns is usually
held fairly constant. In some columns where
the difficulty of separation is reduced (rela-
tive volatility increased) by decreasing pres-
sure, pressure is allowed to float so that it is
as low as possible to minimize energy con-
sumption. In any case it is important to
prevent pressure from changing rapidly, ei-
ther up or down. Sudden decreases in pres-
sure can cause flashing of the liquid on the
trays, and the excessive vapor rates can flood
the column. Sudden increases in pressure
can cause condensation of vapor, and the
low vapor rates can cause weeping and
dumping of trays.
There are a host of pressure control
techniques. Figure 1-2 shows some common
examples.
(a) Manipulate coolant: A control valve
changes the flow rate of cooling water
or refrigerant. If an air cooled con-
denser is used, fan speed or pitch is
changed. Note that the liquid in the
reflux drum is at its bubble point.
Changes in coolant temperature are
compensated for by the pressure con-
troller.
(b) Vent-bleed: Inert gas is added or bled
from the system using a dual split-ranged
valve system so that under normal con-
ditions both valves are closed. The re-
flux must be significantly subcooled in
order to keep the concentration of
product in the vent gas stream low.
Changes in coolant temperature cause
changes in reflux temperature.
(c) Direct: A control valve in the vapor line
from the column controls column pres-
sure. This system is only useful for fairly
small columns.
(d) Flooded: Liquid is backed up into the
condenser to vary the heat transfer area.
Reflux is subcooled. If inerts are pre-
sent, the condenser must be mounted
horizontally to permit venting.
(e) Hot-vapor bypass: A blanket of hot va-
por exists above a pool of cold liquid in
the reflux drum. Heat transfer and con-
densation occur at this interface, and it
is important to avoid disturbances that
would change the interfacial area. The
formation of ripples on the surface can
result in the cold liquid "swallowing"
the hot vapor and produce rapid drops
in pressure. If the reflux drum is located
below the condenser, a large vapor line
connects the drum to the condenser in-
let and a control valve in the liquid line
below the condenser floods the con-
denser. If the reflux drum is located
above the condenser, a control valve in
the vapor line controls the pressure in
the reflux drum (and indirectly the pres-
sure in the column). Reflux is sub-
cooled.
(f) Floating pressure: A valve-position con-
troller is used to keep the cooling water
--f.#tJVv+----t:IiIEt-.. Coolant
(0)
.....::.;.;.-...... Vent
(b)
o}'
"." " ,,' .t/o', '
(c)
(d)
FIGURE 12. Pressure control schemes. (a) Coolant manipulation; (b) vent' bleed;
(c) direct; (d) flooded condenser; (e) hot vapor bypass; (f) floating pressure.
7
8 Practical Distillation Control
FIGURE 12. Continued
valve nearly wide open by slowly chang-
ing the setpoint of a fast pressure con
troller. Dual composition control (or
some scheme to reduce reflux ratio as
pressure is reduced) must be used to
realize the energy savings.
1-2-4 Level Control
The two liquid levels that must be con
trolled are in the reflux drum and column
(I)
Hot all/cold liquid
iDlerf_
=lK
base (or in the reboiler if a kettle reboiler is
used). These levels are controlled in very
different ways, depending on a number of
factors.
If the column is part of a series of units
in a plant, it is usually important from a
plantwide control standpoint to use the liq
uid levels as surge capacities to dampen out
disturbances. In such an environment, it is
usually preferable to control base level with
bottoms flow and reflux drum level with
distillate flow, using a proportional-only
controller. Proportional controllers mini-
mize the flow disturbances that propagate
to downstream units.
However, this is not always possible. One
important example is a high reflux ratio
column (R/D > 5). Using distillate to con-
trol level would require large changes in D
for fairly small changes in R or V. Thus the
disturbances would be amplified in the vari-
ations in the distillate flow rate. In columns
with high reflux ratios, the reflux drum level
should be controlled by reflux. This is an
example of a very simple and useful rule
that was proposed by Richardson (1990) of
Union Carbide:
Richardson's Rule: Always control level
with the largest stream.
This heuristic works well in a remarkably
large percentage of processes.
Application of the same logic would sug-
gest that base level should be controlled by
heat input in high reflux ratio columns. This
is done in some columns, but you should be
careful about potential problems with "in-
verse response" that can occur when this
loop is closed. Increasing heat input will
decrease base liquid level in the long term,
but there may be a short transient period
when the level momentarily increases in-
stead of decreases with an increase in heat
input. This inverse response can result from
the "swell" effect in the reboiler and/or on
the trays of the column itself.
In the reboiler, an increase in heat input
can quickly increase the fraction of vapor.
In a thermosiphon reboiler this can push
liquid back into the base of the column,
resulting in a momentary increase in the
liquid level in the column base. In a kettle
reboiler, the increase in vapor fraction
causes the material in the reboiler to swell
and more liquid flows over the outlet weir
into the surge volume in the end of the
reboiler. Therefore, the liquid level in this
section momentarily increases.
Introduction 9
On the trays in the column, an increase
in vapor rate can cause the liquid on all the
trays to swell. This will increase liquid rates
flowing over the weir. Of course the in-
crease in vapor rate will also increase the
pressure drop through the trays, and this
will require a higher height of liquid in the
downcomer, which tends to decrease the
liquid flow onto the tray. Thus, there are
two competing effects: swell tends to in-
crease internal liquid flows in the column
and pressure drop tends to decrease them.
If the former effect is larger (e.g., in valve
tray columns where pressure drop does not
change much with vapor flow), an increase
in heat input can result in a momentary
increase in liquid flow rate into the reboiler,
which can make the base level increase for
a short period of time.
This inverse response phenomena corre-
sponds to a positive zero in the transfer
function (i.e., a zero that is located in the
right half of the s plane). Because the root-
locus plot always goes to the zeros of the
open-loop system transfer function, the
presence of a positive zero pulls the root-
locus plot toward the unstable region of the
s plane. This clearly shows why inverse re-
sponse results in poor control performance.
1-3 UNIQUENESS OF
DISTILLATION COLUMNS
One of the often-quoted expression in the
distillation area is: "There are no two distil-
lation columns that are alike." This is true
in the vast majority of cases. We will discuss
briefly in succeeding text some of the rea-
sons for this and its impact on both design
and control.
It is true that there are many columns
that fit into general classes:
Stabilizers: The distillate stream is only a
small fraction of the feed and the relative
volatility between the key components is
large.
Superfractionators: The separation is diffi-
cult (relative volatilities < 1.2) and the
10 Practical Distillation Control
columns have many trays and high reflux
ratios (see Chapter 22).
High "K" columns: The separation is ex-
tremely easy, resulting in a very sharp
temperature profile. This can lead to
difficult control problems because of non-
linearity.
Side stream columns: These complex config-
uration produce more than two products
by removing sidestreams from the col-
umn. Sometimes sidestream strippers or
sidestream rectifiers are also used.
The preceding partial list illustrates some
of the generic columns that occur in distilla-
tion. However, despite the similarities, the
individual columns within each of these
classifications can be very different in both
design and control for a number of reasons.
Some conditions that vary from column to
column are summarized:
Feed conditions: The number of components
and the types of components in the feed
can have a drastic effect on the type of
column, condenser, and reboiler used and
on the control system. The feed thermal
condition can also strongly affect the col-
umn design and the control system. Sub-
cooled liquid feed requires a different
column and a different control structure
that does superheated vapor feed.
Product specifications: High purity columns
are more nonlinear and sensitive to dis-
turbances. Larger feed tanks and temp-
erature-composition cascade control
systems may be required. If "on-aim"
control is required (purity must be held
within a narrow band), a blending system
may be required.
Energy costs: Probably the major source of
variability among distillation columns is
the cost of energy in a particular plant
environment. If excess low-pressure
steam is available in the plant, the energy
cost for any distillation column that can
use it is very small.
For example, if you need a propylene-
propane distillation column in a plant with
excess low-pressure steam, you would prob-
ably build a conventional column with many
trays (200) and high reflux ratio (14) and
operate at 17 atm so that cooling water
could be used in the condenser. The control
system would maximize recovery of propy-
lene by using as much heat input as possible
(operating against a flooding or high-pres-
sure constraint) and controlling only the
distillate composition.
If, on the other hand, in your plant you
have to produce incremental steam by burn-
ing more fuel, you would probably build a
vapor recompression system with fewer trays
(150), lower reflux ratio (11) and lower pres-
sure (11 atm). The control system would try
to keep the column at the optimum distil-
late and bottoms compositions (dual compo-
sition control) that represent the best trade-
off between compressor horsepower costs
and propylene recovery.
1-4 INTERACTION BETWEEN
DESIGN AND CONTROL
As with any process, in distillation systems
there are interesting and challenging inter-
actions between the basic design of the pro-
cess, which historically has been based on
only steady-state economics, and its control-
lability. Discussion of some examples fol-
lows.
1-4-1 Increasing Column Size
(Trays and Diameter)
Traditional designs yield columns that oper-
ate at reflux ratios of 1.1 to 1.2 times the
minimum reflux ratio. Usually this corre-
sponds to a column with about twice the
minimum number of trays (which can be
calculated from the Fenske equation for
constant relative volatility systems). Column
diameters and heat exchangers are designed
to handle flow rates that are 10 to 20%
above design.
These designs leave little excess capacity
that can be used to handle the inevitable
disturbances that occur. It is usually good
engineering practice to increase the number
Introduction 11
of trays by 10% and increase the capacity of will not vary excessively during distur-
the column, reboiler, and condenser by 20% bances.
to make it easier to control the column.
1-4-2 Holdups in Column Base and
Reflux Drum
From a steady-state economic standpoint,
these holdups should be as small as possible
so as to minimize capital investment. In-
creasing base holdup means that the col-
umn shell must be longer and requires that
the column be positioned higher above
grade. Both of these effects increase capital
costs. Bigger reflux drums mean high costs
for both the drum and its supporting super-
structure. Large holdups of liquids may also
be undesirable because of safety considera-
tions (if the material is toxic, explosive,
thermally sensitive, etc,),
However, from the standpoint of dynam-
ics and control, the ability of the column to
ride through disturbances is usually im-
proved by having more liquid holdup avail-
able. Thus there is a conflict between
steady-state economics and controllability,
and an engineering trade-off must be made.
Liquid holdup times of about 5 min are
fairly typical. This holdup time is based on
the total material entering and leaving the
column base or reflux drum, not just the
stream that is being used to control the
level. If the column is very large, subject to
many large disturbances, or forms part of a
series of operation units, somewhat larger
holdup times may be appropriate. If small
liquid holdups are required because of safety
or thermal degradation considerations,
somewhat smaller holdup time may be ap-
propriate.
But remember, one of the most common
reasons why columns give control difficulties
is insufficient liquid holdups, so make sure
the process designers do not squeeze too
hard for the sake of a few dollars. Safety
and thermal degradation problems can of-
ten be handled by using external tankage
for surge capacity so that liquid levels will
not be lost or flow rate to downstream units
1-4-3 Effects of Contacting Devices
The type of contact devices used in the
column can significantly affect the dynamics
of the column. Most columns with trays
have more liquid holdup and therefore
slower dynamic responses than columns with
packing. The structured packing that has
become more widely used in recent years in
processes where pressure drop is important
has faster dynamics than tray columns.
These columns respond more quickly to
changes in manipulated variables, but they
also respond more quickly to disturbances.
These considerations should be kept in mind
when designing control systems.
1-4-4 Sensors
It is important for the operation of the
column to provide adequate sensors: enough
thermocouples located so that the tempera-
ture profile can be determined, differential
pressure measurements to sense flooding,
flow measurements of all streams, and so
forth. It is important to include these mea-
surements in the original equipment design
when it is quite inexpensive and easy. Hav-
ing to make field modification to an operat-
ing column can be very expensive and may
take a long time to accomplish.
1-5 SPECIAL PROBLEMS
Throughout the chapters in this book there
will be much discussion and many examples
of various types of columns and systems.
Several commonly encountered situations
that present special control problems are
discussed briefly in the following text.
1-5-1 High-Purity Products
As noted earlier, distillation columns that
produce high-purity products are very non-
linear and sensitive to disturbances. Very
small changes in the feed split can produce
12 Practical Distillation Control
drastic changes in product concentrations at
steady state. The theoretical linear time
constants of these high-purity column are
very large.
These columns can be effectively con-
trolled if disturbances can be detected
quickly and dynamic corrective action be
taken in time to keep the column near the
desired operating point. The system is
somewhat analogous to the "inverted pen-
dulum" process that is studied by mechani-
cal engineers (i.e., balancing a stick on the
palm of your hand): If the position of the
stick can be detected quickly, a vertical po-
sition can be maintained.
Thus it is vital that measurement and
sampling times be minimized. It is also im-
portant to slow down the effects of distur-
bances to these columns. Large upstream
feed tanks may be required to filter distur-
bances.
1-5-2 Small Temperature Differences
Difficult separations occur when boiling
point differences between key components
are small. This results in a temperature
profile in the column that is quite flat and
can make the use of temperature to infer
composition quite ineffective. Sometimes
more sensitivity can be obtained by using
differential or double-differential tempera-
tures, but nonmonotonic behavior can
sometimes occur, which can crash the sys-
tem. If the separation is quite difficult (rela-
tive volatilities less than 1.2), temperature
becomes essentially useless and some type
of composition analyzer is required to con-
trol composition.
1-5-3 Large Temperature Differences
Extremely easy separation (e.g., peanut but-
ter and hydrogen) yields temperature pro-
files that are very sharp and lead to control
problems because of high process gains and
nonlinearity (saturation of the measurement
signal). This problem can be handled effec-
tively by using a "profile position" control
system: several (four to five) temperatures
are measured at tray locations in the col-
umn below and above the tray where the
temperature break occurs under steady-state
design conditions. These temperatures are
averaged, and the average temperature is
controlled, typically by manipulation of heat
input. This technique reduces the process
gain and avoids the measurement saturation
problem that would be experienced if only a
single tray temperature were measured.
1-5-4 Gravity-Flow Reflux
Instead of pumping reflux back to the col-
umn from a reflux drum located somewhat
above grade (the most common configura-
tion), it is sometimes desirable to locate the
condenser and reflux drum above the col-
umn in the superstructure. Then gravity can
be used to overcome the pressure differen-
tial between the reflux drum and the top of
the column. This avoids a pump, which can
be a real advantage in highly corrosive, toxic,
or dangerous chemical systems.
However, the design of the gravity-flow
reflux system requires some explicit consid-
eration of dynamics. The condenser must be
located high enough above the column so
that there is enough head to get liquid back
into the column in the worst-case situation.
The head must overcome pressure drop
through the vapor line, the condenser, the
liquid line, the reflux flow measurement de-
vice, and the reflux control valve. Because
these pressure drops vary as the square of
the flow rates, a 30% increase in flow in-
creases the pressure drops through the fixed
resistances by a factor of (1.3)2 = 1.69. This
69% increase in pressure drop can only
come from some combination of an increase
in liquid height and a decrease in control
valve pressure drop.
Gravity-flow reflux systems are often not
designed with control considerations in
mind, and insufficient height is provided
between the condenser and the top of the
column. This can lead to very poor control.
1-5-5 Dephlegmators
An extension of the gravity-flow-reflux
scheme is to eliminate not only the pump
but also the external condenser. The con-
denser is mounted directly on the top of the
column. These systems are particularly pop-
ular in low-temperature systems where heat
losses must be minimized.
The usual installation has the vapor
flowing upward through the vertical tubes in
the condenser. The condensed liquid flows
countercurrent down the tubes and refluxes
the column. The uncondensed vapor leaving
the top of the condenser is the vapor distil-
late product.
From a control engineer's point of view,
these dephlegmator systems can be night-
mares. The lack of any liquid holdup makes
the system very sensitive to disturbances.
For example, a slug of noncondensibles can
drop the rate of condensation drastically
and the reflux flow will drop immediately.
Fractionation on the trays in the column is
immediately adversely affected. Columns do
not recover quickly from changes in liquid
flow rates because they propagate down the
column fairly slowly (3 to 6 s per tray).
Introduction 13
I recommend that a total trap-out tray be
installed under the condenser to provide
reflux surge capacity. This type of tray has a
chimney to let the vapor pass through to the
condenser, and should be designed with
sufficient depth so that about 5 min of liq-
uid holdup is provided. Reflux from this
reflux-surge tray to the top of the column
should be controlled by an external control
valve and the flow rate measured. Clearly
the reflux-surge tray has to be located high
enough to provide the required pressure
drops.
1-6 CONCLUSION
This chapter has attempted to present some
of the basic principles of distillation column
control. Some of the problems and con-
straints have been briefly discussed. Many
of these topics will be treated in more detail
in later chapters.
Reference
Richardson, R. (1990). 1990 Lehigh Distillation
Control Short Course. Lehigh, P A: Lehigh
University.
2
Historical Perspective
Page S. Buckley
Process Control Consultant
2-1 INTRODUCTION
History is normally approached in a chrono-
logical fashion, starting at an earlier date
and proceeding stepwise and logically to a
later date. To discuss distillation control in
this manner is difficult. Most important de-
velopments have been evolutionary, occur-
ring gradually over a period of time. Conse-
quently, we will be discussing this subject
primarily in terms of eras, rather than spe-
cific dates. Furthermore, the time period of
most interest for distillation control history
is very short, about 40 years. As every histo-
rian knows, the closer one comes to the
present, the more difficult it is to achieve a
valid perspective. Bias worsens when the
historian has been a participant in the events
under discussion.
2-2 WHAT IS CONTROL?
Before talking about distillation control, let
us define the meaning of control, at least as
it is used in this chapter. A chemical plant
or refinery must produce a product or prod-
ucts that meet certain quality specifications.
Preferably, there should be enough inven-
tory to ship upon receipt of orders, although
this is not always feasible, and some orders
call for shipping a certain amount continu-
14
ously or intermittently over a period of time.
Plant operation must meet production re-
quirements while observing certain con-
straints of safety, environmental protection,
and limits of efficient operation. We do not
sell flow rates (usually), liquid levels, tem-
peratures, pressures, and so forth. Contrary
to popular belief, these variables need not
be held constant; primarily they should be
manipulated to achieve the following opera-
tional objectives:
Material balance control: An overall plant
material balance must be maintained.
Production rate must, on the average,
equal rate of sales. Flow rate changes
should be gradual to avoid upsetting pro-
cess equipment.
Product quality control: Final product or
products must meet sales specifications.
Product quality needs to be constant at
only one point in a process-final inven-
tory.
Constraints: If there is a serious malfunction
of equipment, interlocks may shut the
plant down. Otherwise, override controls
may nudge a process away from excessive
pressure or temperature, excessively high
or low liquid levels, excessive pollutants
in waste streams, and so on.
An important implication of the preced-
ing is that for optimum operation, one must
let all variables vary somewhat except final
product quality, and, sometimes, even that
may vary within prescribed limits.
Classical control theory emphasizes (a)
rapid response to setpoint changes and (b)
rapid return to setpoint in the face of dis-
turbances. These performance requirements
really apply only to product quality controls
and to some constraint controls. Some of
the successes of computer control have re-
sulted from restoring control flexibility that
was originally lost by excessive reliance on
fixed setpoint control of individual vari-
ables.
2-3 COLUMN DESIGN METHODS
The reader is presumably familiar with
steady-state column design, so little more
will be said about it.
In the 1930s and 194Os, what chemical
engineers call "unit operations theory"
made tremendous strides. The objectives
were twofold:
1. To ensure that design of piping, heat
exchangers, distillation columns, and so
forth, would have at least flow sheet ca-
pacity.
2. To ensure that this equipment would be
as little overdesigned as possible to mini-
mize capital investment.
For distillation columns, the McCabe-
Thiele procedure for estimating the re-
quired number of theoretical trays was pub-
lished in 1925. It did not, however (at least
in my experience), achieve wide usage until
the early 1940s. During World War II a
group of us trying to increase production
rates in a heavy chemicals plant found that
existing columns designed before 1930 occa-
sionally had tremendous overdesign factors:
they were 10 to 15 times larger than neces-
sary.
By 1948, we were using a safety factor of
2 for designing distillation columns and heat
exchangers, and 25 to 50% for piping.
Historical Perspective 15
Today, overdesign safety factors are very
small. Columns are designed to run much
closer to flooding and thermosyphon reboil-
ers often operate close to choked-flow in-
stability. Older columns with more safety
factor in design, and particularly with
bubble-cap trays, were easier to control.
2-4 COLUMN TRAY AND
AUXILIARY DESIGN
2-4-1 Tray Design
Circa 1950 we began to design columns with
sieve trays instead of bubble-cap trays. For
a given load, a sieve tray column is smaller
and cheaper. However, sieve tray columns
have more limited turndown: about 2: 1 in-
stead of about 8: 1. In addition they are
subject to weeping and dumping. These
characteristics, together with tighter design
practices, increase control problems.
More recently, valve trays have become
popular. They have more turndown capabil-
ities than sieve trays, say 4 : 1, but have the
same potential for weeping and dumping. In
addition, they commonly have another prob-
lem: inverse response.
As noted by Rijnsdorp (1961), with an
increase in boilup some columns show a
momentary increase in internal reflux, fol-
lowed eventually by a decrease in internal
reflux. This is termed inverse response be-
cause it causes a temporary increase in low
boiler composition at the bottom of the
column, followed by an eventual decrease.
The mechanism was elucidated by Buckley,
Cox, and Rollins (1975), who found that
most sieve tray columns demonstrated this
effect at low boilup rates, whereas most
valve tray columns demonstrated it over the
entire turndown. Thistlethwaite (1980) also
studied inverse response and worked out
more details. Composition control via boilup
at the base of a column is vastly more
difficult when a column is afflicted with in-
verse response. Momentarily, the controller
seems to be hooked up backwards. For the
control engineer this can be even worse
than deadtime.
16 Practical Distillation Control
Since the mid-1980s packed columns,
particularly those with structured packing,
have been used more extensively. They have
about the same turndown as valve trays, say
4 : 1, but do not suffer from inverse re-
sponse. On the other hand, changes in re-
flux flow can change column pressure drop
significantly.
2-4-2 Reboller Design
Thermosyphon reboilers, particularly of the
vertical type, are popular because of low
cost. Several kinds of dynamic problems,
however, can result from their use (Buckley,
1974). One is choked flow instability
(Shellene et aI., 1967), which occurs when
the reboiler is sized too snugly with exces-
sive heat flux.
Another problem is "swell." At low heat
loads there is little vapor volume in the
tubes; at high heat loads there is much
more vapor in the tubes. Hence, an increase
in steam flow will cause liquid to be dis-
placed into the column base, causing a tem-
porary increase in base level. If base level is
controlled by steam flow, the controller mo-
mentarily acts like it is hooked up back-
wards. When "swell" is accompanied by in-
verse response of reflux, controlling the base
level via heating medium may be impossi-
ble.
Recent trends in thermosyphon design
seem to have had at least one beneficial
effect: critical aT is rarely a problem. Prior
to 1950, it could be observed fairly often.
This improvement apparently is due to much
higher circulating rates.
2-4-3 Flooded Rebollers
Particularly with larger columns, it is com-
mon to use steam condensate pots with
level controllers instead of traps. In recent
years some reboilers have been designed to
run partially flooded on the steam side.
The large steam control valve may be
replaced by a much smaller valve for steam
condensate. Also, it is usually easier to mea-
sure condensate flow rather than steam flow.
Flooded reboilers are, however, more slug-
gish than the nonflooded ones, typically by a
factor of 10 or more. Flooded reboilers
sometimes permit use of low pressure, waste
steam. There may be little pressure drop
available for a steam supply valve.
2-4-4 Column Base Designs
The size of the column base, the design of
the internals (if any), and the topological
relationship to the reboiler are all impor-
tant factors in column control. Column base
holdups may range from a few seconds to
many hours (in terms of bottom product
flow rate). If there are no intermediate
holdups between columns or process steps,
the column base must serve as surge capac-
ity for the next step.
Originally, column bases were very sim-
ple in design, but in recent years elaborate
internal baffie schemes have often been
used. The objective is to achieve one more
stage of separation. Many of these schemes
make column operation more difficult and
limit the use of the holdup volume as surge
capacity. Overall, their economics is ques-
tionable; increasing the number of trays in a
column by one is relatively cheap.
Elimination of intermediate tanks can
save a lot of money, both in fixed invest-
ment (tanks) and working capital (inven-
tory).
2-4-5 Overhead DeSign
In the chemical industry, particularly for
smaller columns, it is common to use gravity
flow reflux and vertical, coolant-in-shell
condensers. Instrumentation and controls
can be much simpler than with pumped-back
reflux. Regardless, however, of condenser
design, gravity flow reflux systems have a
dynamics problem if distillate is the con-
trolled flow and reflux is the difference flow.
This is the well known problem of "reflux
cycle." Its mechanism and several correc-
tive measures are discussed in a paper by
Buckley (1966).
When horizontal, coolant-in-tube con-
densers are used, control is sometimes sim-
plified by running the condensers partially
flooded. This technique is especially useful
when the top product is a vapor. Flooded
condensers seem to be becoming more pop-
ular, but like flooded reboilers, are more
sluggish. For both, however, we have worked
out mathematical models that enable us to
calculate the necessary derivative compen-
sation (Buckley, Luyben, and Shunta, 1985).
25 INSTRUMENTATION
Before automatic controls and automatic
valves became prevalent, control was
achieved by having operators observe local
gages and adjust hand valves to control pro-
cess variables to desired values. According
to stories told to me by old timers, some
applications required such careful attention
and frequent adjustment that an operator
would be kept at a single gage and hand
valve. To help keep the operator alert, a
one-legged stool was provided.
Gradually, this practice diminished as
manufacturers developed automatic con-
trols in which measurement, controller, and
valve were combined in one package. Some-
times the valve itself, usually powered by air
and a spring and a diaphragm actuator, was
a separate entity. Slowly, the practice of
separating measurement, controller, and
valve evolved.
The development of measurement de-
vices that could transmit pneumatic or elec-
tric signals to a remote location was well
underway by 1940. This permitted the use
of central control rooms and, together with
more automatic controls, a reduction in the
required number of operators.
Historical Perspective 17
Prior to 1940 most controllers were ei-
ther proportional-only or had manual rather
than automatic reset. A few vendors, how-
ever, had begun to use primitive versions of
automatic reset.
By 1945, the chief instruments available
for distillation control were:
Flow: Orifice with mercury manometer
plus pneumatic transmission, or rotame-
ters.
Temperature: Thermocouple or mercury-
in-bulb thermometers.
Pressure: Bourdon tube plus pneumatic
transmission.
Control valve: Air-operated with spring-
and-diaphragm or electric motor opera-
tor. Most valves, except for small ones,
were of the double-seated variety; small
ones were single-seated. Valve positioners
occasionally were used.
By 1950 the picture had begun to change
dramatically. The so-called force balance
principle began to be used by instrument
designers for pneumatic process variable
transmitters and controllers. This feature,
together with higher capacity pneumatic pi-
lots, provides more sensitivity, freedom from
drift, and speed of response. Mercury-type
flow transmitters gave way to "dry" trans-
mitters and mercury-in-bulb temperature
transmitters were replaced by gas-filled bulb
transmitters.
Piston-operated valves with integral valve
positioners appeared on the market. The
practicality of larger sized single-seated
valves permitted less expensive valves for a
given flow requirement and also provided
increased sensitivity and speed of response.
With improved actuators and positioners,
reliability improved, and many users aban-
doned control valve bypasses except for a
few cases. This saves money (investment)
and in many cases reduces maintenance.
In 1974, the Instrument Society of Amer-
ica (ISA) published new standards for calcu-
lating flow through control valves. This was
one of the most valuable developments of
18 Practical Distillation Control
the past 25 years. It permitted much more
accurate valve sizing, even for flashing, cavi-
tation, or incipient cavitation. In a parallel
development, valve vendors began to make
available more accurate plots or tables re-
lating C
u
' the valve flow coefficient, to valve
stem position. Together with nonlinear
functions available in new microprocessor
controls, these permitted the design of valve
flow compensators (Buckley, 1982), which in
tum, permits much more constant stability
over a wide range of flows.
New pneumatic controllers with external
reset feedback appeared, which permitted
the use of anti-reset windup schemes. Circa
1952, one vendor marketted an extensive
line of pneumatic computing relays, which
permitted addition, subtraction, and multi-
plication by a constant, high- and low-signal
selection, and so-called impulse feed-
forward. Decent pneumatic devices for
multiplying or dividing two signals did not
appear until the mid-1960s. Taken together,
these various devices permitted on-line cal-
culation of heat and material balances. Cal-
culation of internal reflux from external re-
flux and two temperatures also became
practical, and simple pressure-compensated
temperature schemes became feasible.
The improved AP transmitters, origi-
nally developed for orifice flow measure-
ments, permitted sensitive specific gravity
and density measurements. They could also
be used to measure column A P, increas-
ingly important for columns designed to run
closer to flooding.
Parallel to developments in conventional
instruments, there was a much slower devel-
opment in on-line analyzers. In the late
1940s, Monsanto developed an automatic
refractometer, which was used in a styrene
monomer distillation train. In the early
1950s, several prototypes of on-line infrared
and optical analyzers were available. Prob-
lems with reliability and with sample system
design plus high cost severely limited their
applications. pH measurements, on the
other hand, expanded more rapidly, as the
result of improved cell design, better cable
insulation, and high input impedance elec-
tronic circuitry.
The first analog electronic controls ap-
peared in the mid-1950s. They had good
sensitivity and speed of response, but due to
the use of vacuum tubes, had somewhat
high maintenance. In addition, they were
far less flexible than pneumatics. Subse-
quent switching to solid state circuitry im-
proved reliability, but flexibility still suf-
fered. It was not until the early 1970s that
one manufacturer produced an electronic
line with all of the computation and logic
functions available in existing pneumatics.
In the meantime, the cost competitiveness
of pneumatics was improved by a switch
from metallic transmission tubing to black
polyethylene tubing (outdoors) and soft vinyl
tubing (in the control room).
In the late 1970s a potentially momen-
tous change got underway-the introduc-
tion of so-called distributed digital control
systems. By using a number of independent
microprocessors, manufacturers were able
to "distribute" the control room hardware
and achieve better reliability than would be
obtained with a single, large time-shared
computer (avoid "all eggs in one basket").
Early versions were quite inflexible com-
pared with the best pneumatic and elec-
tronic hardware, but this situation has now
changed. Some of the newest systems have
more flexibility than any analog system and
lend themselves well to the design of ad-
vanced control systems. Transmitters and
control valves in the field usually may be
either pneumatic or electronic.
No discussion of instrumentation is com-
plete without some mention of computers.
The first process control computer was put
on the market in the early 1950s. A well-
publicized application was a Texaco refin-
ery, where some optimization was per-
formed. Another optimization application
of that era was a Monsanto ammonia plant
at Luling, LA. Other machines came on the
market in the next few years, but all of the
early machines suffered from lack of relia-
bility, slow operation, and insufficient mem-
ory. In the 1960s and 1970s applications
were chiefly as data loggers and supervisory
controls. Production personnel found that
CRT displays, computer memory, and com-
puter printouts were a big improvement over
conventional gages and recorders. As the
memory and speed of computers have in-
creased and as the expertise of users has
improved, we have seen a resurgence of
interest in optimization.
The latest development in instrumenta-
tion began in mid-1983 when one vendor
put a line of "smart" transmitters on the
market. These featured built-in micropro-
cessors to permit remote automatic calibra-
tion, zeroing, and interrogation. These fea-
tures, together with reduced drift and
sensitivity to ambient conditions, and with
more accuracy and linearity, improve plant
operation. They also facilitate and reduce
maintenance. However, test data I have seen
on such instruments indicate that they are
slower (poorer frequency response) than
their pneumatic counterparts. Eventual ex-
tension to instruments with multiple func-
tions such as temperature and pressure is
foreseen. Present indications are that within
the next several years a number of manufac-
turers will put "smart" instruments on the
market.
2-6 CONTROL SYSTEM DESIGN
METHODS
Prior to 1948, distillation control systems as
well as other process control systems were
designed by what might be called the instru-
mentation method. This is a qualitative ap-
proach based principally on past practice
and intuition. Many "how-to" papers have
appeared in the literature. A good summary
of the state-of-the-art in 1948 is given by
Boyd (1948).
In the period 1948 to 1952, there ap-
peared a large number of books based on
control technology which had been devel-
oped for military purposes during World
War II. Commonly the title mentioned
"servomechanisms theory." Two of the best
Historical Perspective 19
known books are by Brown and Campbell
(1948) and Chestnut and Mayer (1951). A
wartime text by Oldenbourg and Sartorius
(1948) was translated from German into En-
glish and became well known. Although the
theory required significant modifications to
adapt it to process control, for the first time
it permitted us to analyze many process
control situations quantitatively.
The technique involves converting pro-
cess equations, including differential equa-
tions, into linear perturbation equations. By
perturbations we mean small changes
around an average value. Linear differential
equations with constant coefficients are eas-
ily solved by use of the Laplace transforma-
tion. It is usually easy to combine a number
of system equations into one. This equation,
which relates output changes to input
changes, is called the transfer function. Time
domain solutions are readily calculated, and
if the uncontrolled system is stable, the
transfer function may be expressed in terms
of frequency response.
The transfer function approach permits
convenient study of system stability, permits
calculation of controller gain and reset time,
and permits calculation of system speed of
response to various forcing functions. In the
precomputer era, these calculations were
made with a slide rule and several graphical
aids. Today, the transfer function tech-
niques are still useful but the calculations
usually may be made more rapidly with a
programmable calculator or a computer.
Early applications were in flow, pressure,
and liquid level systems. Although fre-
quency response methods of calculation
were most commonly used, some users pre-
fer root locus methods.
Throughout the 19508 and 1960s the
transfer function approach was used exten-
sively for piping, mixing vessels, heat ex-
changers, extruders, weigh belts, and, occa-
sionally, distillation columns. Two books
that appeared in the 1950s illustrated the
application of transfer function techniques
to process control (Campbell, 1958; Young,
1957).
20 Practical Distillation Control
The relatively simple modelling tech-
niques developed earlier were of limited use
for systems with significant nonlinearities or
those describable by either partial differen-
tial equations or by a large number of ordi-
nary differential equations. Gould (1969)
explored some more advanced modelling
techniques for such systems.
The limitations of the simple transfer
function approach for distillation columns
led some workers to try another approach
-computer simulation. The work of
Williams (1971) and Williams, Harnett, and
Rose (1956) in applying this approach is
well known. Originally, simulations were
mostly run on analog computers. Then,
when improved digital computers and FOR-
TRAN programming became available,
there was a shift to digital techniques. Basi-
cally, simulation consists of writing all of the
differential equations for the uncontrolled
system and putting them on the computer.
Various control loops can then be added,
and after empirically tuning the controllers,
the entire system response can be obtained.
In some cases real controllers, not simu-
lated ones, have been connected to the
computer.
In 1960, Rippin and Lamb (1960) pre-
sented a combined transfer function-simu-
lation approach to distillation control. The
computer simulation was used to develop
open loop transfer functions. Then standard
frequency response methods were used to
calculate controller gain and reset as well as
desired feedforward compensation. In our
own work, we have found this combination
approach to be quite fruitful.
Since about 1980, there has been enor-
mous progress in small or personal comput-
ers (PCs) accompanied by a more gradual
evolution in software. The machines have
more memory and speed, and both ma-
chines and software are more user friendly.
Simple distillation columns are readily mod-
eled and programmed; see, for example, the
program illustrated in Luyben (1990). Mod-
elling, including newer approaches, is dis-
cussed by other authors in this book.
As the speed and memory of technical
computers increased, distillation models
have become more detailed and elaborate.
Tolliver and Waggoner (1980) have pre-
sented an excellent discussion of this sub-
ject.
Many newer control techniques have
been devised. Among these are "internal
model control" and "dynamic matrix con-
trol." These and other techniques are sum-
marized and reviewed by Luyben (1990).
There has been some application to distilla-
tion control, as discussed in Chapter 12.
Although most process measurements are
continuous, some are not, as, for example,
once per day sampling and laboratory analy-
sis of column product streams. This led to
an interest in the techniques of sampled-
data control. This permits the quantitative
design of discontinuous control systems.
There have been few applications to distilla-
tion control in spite of the fact that many
composition measurements are discontinu-
ous. One application has been the use of
Shannon's data reconstruction theorem to
establish the relationship between a speci-
fied sampling frequency and required pro-
cess holdup, or between a specified holdup
and required sampling frequency.
These newer techniques require a lot of
training beyond that required for transfer
function techniques, and industrial applica-
tions have so far been limited. There is no
question, however, that there exist ample
incentives and opportunities for this newer
control technology.
An interesting trend in recent years has
been toward the development of standard-
ized programs for control system design.
This is commonly called computer aided
design (CAD). Although originally aimed at
larger machines, many useful programs exist
for small computers and even pro-
grammable calculators. The portability and
permanent memory features of many of the
latter render them very useful in a plant or
maintenance shop.
Since about 1970, there has been increas-
ing interest in newer approaches to control.
The incentives have been twofold:
1. Because flow sheet data are often incom-
plete or not exact, it is desirable to mea-
sure process characteristics of the real
plant after startup. Furthermore, process
dynamics change as operating conditions
change. The increasing availability of
on-line computers permits the use of
"identification" techniques. Information
thus obtained may be used to retune
controllers either manually or automati-
cally. If a column is not equipped with
analyzers, identification techniques may
be used to deduce compositions from a
combination of temperature, pressure,
and flow measurements. A particular ap-
proach that was developed by Brosilow
and Tong (1978) is termed inferential
control. For binary or almost binary dis-
tillation, pressure-compensated tempera-
ture usually suffices.
2. Many processes are plagued by interac-
tions. There are a variety of techniques
for dealing with this that may be lumped
under the heading "modern control the-
ory" (Ray, 1981). That aspect relating to
time-optimal control will likely be help-
ful in future design of batch distillation
columns. Implementation of this technol-
ogy usually requires an on-line computer.
Some other methods of dealing with in-
teractions include relative gain array (RGA),
inverse Nyquist array (INA), and decou-
pIers. See Luyben (1990) for a good short
summary.
2-7 PROCESS CONTROL
TECHNIQUES
In 1945 there existed no overall process
control strategy. As one process design
manager put it, "instruments are sprinkled
on the flow sheet like ornaments on a
Christmas tree." Not until 1964 was there
published an overall strategy for laying out
operational controls from one end of a pro-
cess to the other (Buckley, 1964). This strat-
Historical Perspective 21
egy, as mentioned earlier, has three main
facets: material balance control, product
quality control, and constraints. What kind
of controls do we use to implement them?
Material balance controls are tradition-
ally averaging level and averaging pressure
controls. Although the basic concept dates
back to 1937, averaging controls are still
widely misunderstood. The original theory
was expanded in 1964 (Buckley, 1964) and
again quite recently (Buckley, 1983). To
minimize required inventory and achieve
maximum flow smoothing we use quasilin-
ear or nonlinear proportional integral (PI)
controllers cascaded to flow controllers. If
we can use level control of a column base to
manipulate bottom product flow and level
control of the overhead receiver to adjust
top product flow, then we can usually avoid
the use of surge tanks between columns or
other process steps. This saves both fixed
investment and working capital.
Product quality controllers are usually
also of the PI type. Overhead composition is
usually controlled by manipulating reflux
and base composition by manipulating heat
flow to the reboiler. It is also increasingly
common to provide feedforward compensa-
tion for feed rate changes. This minimizes
swings in boilup and reflux, and in top and
bottom compositions.
One of the most important developments
in control techniques is that of variable con-
figuration and variable structure controls.
Although it is much more common to use
fixed configuration and fixed structures, in
reality either or both should change as pro-
cess conditions change. For example, the
steam valve for a distillation column re-
boiler may, depending on circumstances, re-
spond to controllers for:
Steam flow rate
Column b.p
Column pressure
Base temperature
Column feed rate
Column base level
Column bottom product rate
22 Practical Distillation Control
The seven variables listed may also exert
control on five or six other valves. We call
this variable configuration.
For level control, the quasilinear struc-
ture referred to earlier consists normally of
PI control, but proportional-only control is
used when the level is too high or too low.
We call this variable structure. To accom-
plish this automatically, it was proposed in
1965 (Buckley, 1968) to use a type of control
called overrides. In ensuing years a number
of other publications, (Buckley and Cox,
1971; Buckley, 1969b; Cox, 1973) described
the applications to distillation. The math
and theory are presented in a two-part arti-
cle (Buckley, 1971).
Fairly recently, there has been a growing
interest in "self-tuning regulators." One of
the most prolific contributors has been
Astrom (1979) in Sweden. So far, applica-
tions to industrial distillation columns are
not known, but they have been tried on
pilot columns.
Although little has been published, there
has been a certain amount of work with
nonlinear controllers. Unlike linear theory,
which has a cohesive, well-developed
methodology, nonlinear theory is essentially
a bag of tricks. The best approach I have
found is to run a large number of simula-
tions with different algorithms and for vari-
ous types and sizes of forcing functions.
For averaging level control we found
(Buckley, 1986) that with linear reset and
gain proportional to the error squared, we
got flow smoothing far superior to that
achievable with any linear controller,
whether P or PI, and experiences where
P-only controllers were replaced with this
type of controller confirmed simulation re-
sults.
For composition control, we tried nonlin-
ear gain to filter out noise. Sometimes it
worked well, sometimes not.
Our experience with trial-and-error tun-
ing of nonlinear controllers in a plant has
been negative. These applications should
be simulated. For those interested in non-
linear control, we recommend a book by
Oldenburger (1966).
28 INFLUENCE ON DISTILLATION
CONTROL
Some years ago I was preseni: when one of
our most experienced distillation experts was
asked "What is the best way to control a
column?" His answer: "Gently." When ap-
proaching column control one should keep
in mind that conditions in a column cannot
be changed rapidly. Rapid changes in boilup
or rate of vapor removal, for example, may
cause momentary flooding or dumping. The
developments since 1945 have therefore not
been aimed at rapid control but rather at
smoother operation with better separation
at lower cost.
If we now look back at 1945 and move
forward we can see how the various devel-
opments previously discussed have inter-
acted to influence distillation control.
1. Switch to sieve and valve trays and tighter
column design. These have encouraged
the use of minimum and maximum boilup
overrides. A high I1p override on steam
is now universally employed by some en-
gineering organizations. We also fre-
quently provide maximum and minimum
overrides for reflux.
2. Increased use of packed columns. As
mentioned earlier, packed columns, par-
ticularly those with structured packing,
have become more popular.
3. Tighter design of heat exchangers, espe-
cially reboilers. To prevent choked flow
instability we sometimes provide maxi-
mum steam flow limiters.
4. Increased use of side-draw columns. One
side-draw column is cheaper than two
conventional columns with two takeoffs.
Turndown is less, however, and the col-
umn is harder to control (Buckley, 1969a;
Doukas and Luyben, 1978; Luyben, 1966).
Such columns usually require more over-
rides, particularly to guarantee adequate
reflux below the side draw or draws.
5. Energy crunch of 1973. This has had two
major effects:
a. Increased energy conservation of con-
ventional columns. After improved in-
sulation, one of the big items has be-
come double-ended composition con-
trol. A column runs more efficiently
and has more capacity if composition
is controlled at both ends. Another
consequence has been more interest
in composition measurements instead
of just temperature. Temperature is a
nonspecific composition measurement
except for a binary system at constant
pressure. This, together with the com-
mon practice of having temperature
control of only one end of a column,
usually leads production personnel to
run a column with excess boilup and
reflux to ensure meeting or exceeding
product purity specifications. This
wastes steam and limits column ca-
pacity. Experience indicates that 10 to
30% savings are often achievable. To
get an idea of the numbers involved,
consider a column that uses 1000 pph
of steam. If we can save 100 pph, if
average operating time is 8000 h/yr,
and if steam costs $5 per 1000 lb, then
annual savings are $4000. This means
we can invest $8000 to $12,000 in im-
proved controls to achieve the desired
steam saving.
b. Heat recovery schemes. Much larger
energy savings are often possible with
energy recovery schemes, usually in-
volving the condensation of vapor
from one column in the reboiler of
another column (Buckley, 1981). Our
experience has been that designing
controls for energy recovery schemes
is much more difficult and time-con-
suming than for conventional columns.
This extra cost, however, is usually a
small fraction of savings. An interest-
ing feature of some heat recovery
Historical Perspective 23
schemes is that the pressure in the
column supplying heat is allowed to
float. Without going into details I will
simply say that we have found that
this simplifies process design (usually
no auxiliary condensers are needed)
and simplifies instrumentation and
control (no pressure controller is
needed). Because maximum operating
pressure occurs when boilup is a max-
imum, there is no problem with
flooding. There are other problems,
however. For example, the feed valve
to a column with floating pressure has
a much larger aP variation with feed
rate than is commonly encountered.
Additionally, the column bottom
product valve has a low ap at low
rates and a high aP at high rates. To
obtain reasonably constant stability,
one may need to provide "flow char-
acteristic compensation" (Buckley,
1982). This need is increased by a
trend toward lower ap for valves in
pumped systems. There are some dif-
ferences in control requirements be-
tween the chemical and petroleum in-
dustries. Tolliver and Waggoner
(1980) have presented an excellent
discussion of this subject.
Other approaches to heat recovery are
discussed in Chapter 24.
2-9 CONCLUSION
From the preceding discussions we can see
that since 1945 we have moved from gener-
ously designed columns with bubble-cap
trays and simple controls to tightly designed
columns with sieve or valve trays or packing
and with more sophisticated controls. For
new design projects with difficult applica-
tions it may be appropriate in some cases to
consider going back to bubble-cap trays and
larger design safety factors to avoid the
necessity of very complex and perhaps
touchy controls.
24 Practical Distillation Control
The literature on distillation control, par-
ticularly that from academic contributors,
emphasizes product composition control.
Our experience has been that many operat-
ing difficulties, in addition to those due to
tight design, are due to improperly designed
column bases and auxiliaries, including pip-
ing. In my own experience, the single factor
that contributes most to low cost operation
and good composition control has been
properly designed averaging liquid level
controls.
As of mid-1990, lack of adequate product
quality measurements is probably the
Achilles' heel of distillation control. With
adequate quality measurements and compo-
sition control of each product stream, the
next problem area will be that of interac-
tions.
Although a great deal of progress has
been made in quantitative design of control
systems, I would say that today we are about
where distillation design was in 1948.
2-10 COMMENTS ON REFERENCE
TEXTS
There are now several texts on distillation
control. One of them, which has the most
complete technical treatment, is that by
Rademaker, Rijnsdorp, and Maarleveld
(1975). Another, by Shinskey (1977) is no-
table for its treatment of energy conserva-
tion. A third, by Nisenfeld and Seemann
(1981), has a well-organized treatment of
distillation fundamentals and types of
columns. Neither of these last two books
deals with control in a technical sense: there
is no stability theory, Laplace transforms, or
frequency response. But each contains many
gems of practical details about column con-
trol. There are two more recent books. One,
by Buckley, Luyben and Shunta (1985), has
an initial section devoted to practical details
and a second half that discusses distillation
dynamics and control with modest theory. A
second book, by Deshpande (1985), has a
more advanced treatment.
References
...\strom, K J. (1979). Self-tuning regulators. Re-
port LUTFD2/(TRFT-7177)/1-068, Lund
Institute of Technology, Sweden.
Boyd, D. M. (1948). Fractionation instrumenta-
tion and control. Petro Ref., October and
November.
Brosilow, C. B. and Tong, M. (1978). Inferential
control of processes: Part II. AIChE 1. 24, 485.
Brown, G. S. and Campbell, D. P. (1948). Princi-
ples of Servomechanisms. New York: John Wi-
ley.
Buckley, P. S. (1964). Techniques of Process Con-
trol. New York: John Wiley.
Buckley, P. S. (1966). Reflux cycle in distillation
columns. Presented at IFAC Conference,
London, 1966.
Buckley, P. S. (1968). Override controls for distil-
lation columns. InTech. 15, August, 51-58.
Buckley, P. S. (1969a). Control for sidestream
drawoff columns. Chern. Eng. Prog. 65, 45-51.
Buckley, P. S. (1969b). Protective controls for
sidestream drawoff columns. Presented at
AIChE Meeting, New Orleans, March 1969.
Buckley, P. S. (1971). Designing override and
feedforward controls. Part 1. Control Eng.
August, 48-51; Part 2, October, 82-85.
Buckley, P. S. (1974). Material balance control in
distillation columns. Presented at AIChE
Workshop, Tampa, FL, November 1974.
Buckley, P. S. (1981). Control of heat-in-
tegrated distillation columns. Presented at
Engineering Foundation Conference, Sea Is-
land, GA, January 1981.
Buckley, P. S. (1982). Optimum control valves for
pumped systems. Presented at Texas A&M
Symposium, January 1982.
Buckley, P. S. (1983). Recent advances in averag-
ing level control. Presented at Instrument So-
ciety of America meeting, Houston, April
1983.
Buckley, P. S. (1986). Nonlinear averaging level
control with digital controllers. Presented at
Texas A&M Instrumentation Symposium,
January 1986.
Buckley, P. S. and Cox, R. K (1971). New devel-
opments in overrides for distillation columns.
1&4 Trans. 10, 386-394.
Buckley, P. S., Cox, R. K, and Rollins, D. L.
(1975). Inverse response in a distillation col-
umn. Chern. Eng. Prog. 71, 83-84.
Buckley, P. S., Luyben, W. L., and Shunta, J. P.
(1985). Design of Distillation Column Control
Systems. Research Triangle Park, NC.: Instru-
ment Society of America.
Campbell, D. P. (1958). Process Dynamics. New
York: John Wiley.
Chestnut, H. and Mayer, R. W. (1951). Ser-
vomechanisms and Regulating System Design,
Vols. I and II. New York: John Wiley.
Cox, R. K. (1973). Some practical considerations
in the application of overrides. Presented at
ISA Symposium, St. Louis, MO, April 1973.
Deshpande, P. B. (1985). Distillation Dynamics
and Control. Research Triangle Park, NC:
Instrument Society of America, 1985.
Doukas, N. and Luyben, W. L. (1978). Control of
sidestream columns separating ternary mix-
tures. In Tech. 25, 43-48.
Gould, L. A. (969). Chemical Process Control.
Reading, MA: Addison-Wesley.
Luyben, W. L. (1966). 10 schemes to control
distillation columns with side stream drawoff.
ISA J. July, 37-42.
Luyben, W. L. (1990). Process Modeling, Simula-
tion and Control for Chemical Engineers, 2nd
ed. New York: McGraw-Hill, 1990.
Nisenfeld, A. E. and Seemann, R. C. (1981).
Distillation Columns. Research Triangle Park,
NC: Instrument Society of America.
Oldenburger, R. (1966). Optimal Control. New
York: Holt, Reinhart, and Winston.
Oldenbourg, R. C. and Sartorius, H. (1948). The
Dynamics of Automatic Controls. New York:
ASME.
Rademaker, 0., Rijnsdorp, J. E., and Maar-
leveld, A. (1975). Dynamics and Control of
Historical Perspective 25
Continuous Distillation Columns. New York:
Elsevier.
Ray, W. H. (1981). Advanced Process Control.
New York: McGraw-Hill.
Rijnsdorp, J. E. (1961). Birmingham University
Chemical Engineer 12,5-14, 1961.
Rippin, D. W. T. and Lamb, D. E. (1960). A
Theoretical Study of the Dynamics and Control
of Binary Distillation. Newark, DE: University
of Delaware.
Shellene, K. R., Stempling, C. V., Snyder, N. H.,
and Church, D. M. (1967). Experimental study
of a vertical thermosyphon reboiler. Pre-
sented at Ninth National Heat Transfer Con-
ference AIChE-ASME, Seattle, WA, August
1967.
Shinskey, F. G. (1977). Distillation Control. New
York, McGraw-Hill.
Thistlethwaite, E. A. (1980). Analysis of inverse
response behavior in distillation columns. M.S.
Thesis, Department of Chemical Engineering,
Louisiana State University.
Tolliver, T. L. and Waggoner, R. C. (1980). Dis-
tillation column control: A review and per-
spective from the CPI. ISA Adv. Instrum. 35,
83-106.
Williams, T. 1. (1971). Distillation column con-
trol systems. Proceedings of the 12th Annual
chemical and Petroleum Industries Sympo-
sium, Houston, April 1971.
Williams, T. J., Harnett, B. T., and Rose, A.
(1956). Ind. Eng. Chem. 48,1008-1019.
Young, A. 1., Ed. (1957). Plant and Process Dy-
namic Characteristics. New York: Academic.
2
Methods
3
Rigorous Modelling and
Conventional Simulation
Vincent G. Grassi II
Air Products and Chemicals, Inc.
31 OVERVIEW
The aim of this chapter is to show the
reader how to develop a dynamic simulation
for a distillation tower and its control from
first principles. The different classes of dy-
namic distillation models and various ap-
proaches to solving these models will be
presented. The author hopes to dispel the
myth that modelling and simulation of dis-
tillation dynamics must be difficult and com-
plex.
Dynamic modelling and simulation has
proven to be an insightful and productive
process engineering tool. It can be used to
design a distillation process that will pro-
duce quality products in the most economic
fashion possible, even under undesirable
process disturbances. Working dynamic
models provide a process engineering tool
that has a long and useful life.
Dynamic simulation can be used early in
a project to aid in the process and control
system design. It ensures that the process is
operable and can meet product specifica-
tions when the process varies from steady-
state design. Later in the project the simula-
tion can be used to complete the detailed
control system design and solve plantwide
operability problems. After the project, the
same simulation is useful for training. Years
later, as product and economic conditions
change, the simulation can be used for plant
improvement programs.
The authors of this book offer different
approaches to distillation modelling. Chap-
ter 4 discusses reduced and simplified mod-
els. Chapter 5 presents a novel concept in
object orientated simulation. Chapter 6 pre-
sents concepts necessary to develop a
plantwide simulator. These chapters on
modelling and simulation will provide read-
ers with a solid framework so the methods
and ideas presented in the remaining chap-
ters of this book can be implemented.
This chapter will be devoted to develop-
ing process models that realistically predict
plant dynamics and their formulation in al-
gorithms suitable for digital computer codes.
Computer source codes are readily available
from many sources (Franks, 1972; Luyben,
1990), so they will not be repeated here. I
intend to share my experience in properly
selecting models that accurately predict the
dynamics of real distillation columns one is
likely to find in the plant.
A dynamic model is needed to study and
design composition controls. To do this we
will develop a sufficiently rigorous tray-
by-tray model with nonideal vapor-liquid
29
30 Practical Distillation Control
and stage equilibrium. Proportional-integral
feedback controllers will control product
compositions or tray temperatures. Vapor
flow and pressure dynamics often can be
assumed negligible. I will discuss how sys-
tems with vapor hydraulic pressure dynam-
ics can be modelled and simulated. My ap-
proach is based on fundamental process
engineering principles. Only the relation-
ships that are necessary to solve the prob-
lem should be modelled. Most importantly,
these are models that any chemical engi-
neer can easily simulate. They are suitable
for a small personal computer, in whatever
programming language you prefer.
3-11 Conventional Simulation
Let me begin by defining what I mean by
conventional simulation. I view conventional
simulation as a technique used by many
practicing engineers over the past 30 years
based on a physical approach. I consider
nonconventional techniques to be simula-
tion methods based on mathematical ap-
proaches. The conventional approach usu-
ally follows the process flows, solving each
physical phenomenon in an organized se-
quence of steps. The nonconventional ap-
proach writes the governing equations for
each unit operation and solves the entire
system as a set of simultaneous differential
and algebraic equations. Each method has
advantages and disadvantages. The conven-
tional method usually results in a model
that is numerically easier to solve, but spe-
cific to process. Mathematically, a conven-
tional simulation is easier to initialize and
more robust when dealing with discontinu-
ities. The nonconventional method is nu-
merically more difficult to solve, but is flex-
ible and less specific. Flexibility is important
when developing a general purpose dynamic
simulator.
Chemical engineering models can be
classified by type. Models can be classified .
in four categories: algebraic, ordinary dif-
ferential, differential algebraic, and partial
differential. Table 3-1 contains a general
classification for chemical engineering mod-
els.
Our models for distillation towers use
the equilibrium stage approach. Differen-
tial-algebraic equations (DAB) govern this
tower model. Differential-algebraic models
are characterized by their index (Brenan,
Campbell, and Petzold, 1989). The index is
equal to the number of times we must dif-
ferentiate the set of algebraic equations to
obtain continuous differential equations for
all unknown variables. An index of zero is a
set of differential equations without any al-
gebraic equations. If the algebraic equations
are solved separately from the differential
equations, the system is of index O. Higher
levels of index result when the algebraic and
differential equations are solved simultane-
ously. Figure 3-1a contains three examples
of various index problems. The unknown
variables are Yl' Y2' and z in these exam-
ples.
Index 0 problems are much easier to
solve numerically than those with index 1.
Indices greater than 1 are very difficult to
solve, if solvable at all. There are commer-
cial integration packages available for solv-
ing index 1 problems reliably, such as
DASSL (Brenan, Campbell, and Petzold,
1989). Currently reliable algorithms to solve
problems with index 2 or greater do not
exist. If you try to simulate a differential-
TABLE 3-1 Classification 01 Chemical Engineering
Models
Model Type
Algebraic
Ordinary differential
Differential algebraic
Partial differential
Characteristics
Linear
Nonlinear
Linear
Nonlinear
Nonstiff
Stiff
Index = 0
Index = 1
Index> 1
Parabolic
Hyperbolic
Elliptical
Rigorous Modelling and Conventional Simulation 31
INDEX = 1
YI = allYl + a12YZ + z
Yz = aZIYI + azzYz + z
YI + Yz + Z = 0
Differentiate once to get
YI' Yz, and i
INDEX = 2
YI = allYl + a12Y2 + Z
Yz = aZIYI + a2zYz + Z
Differentiate twice to get
YI' Yz, and i
YI + Yz = 0
INDEX = 3
Yr = Y3Yt
jiz = Y3YZ - g
Differentiate three times to get
YI' Yz, and Y3
Yf + - L
Z
= 0
(a)
"-+--Q
dM
dt = Fin - Fout
dE
dt = HinFin - HoutFout + Q
E = HoutM
H
out
=AT + BT2
Specify Q, calculate T Index = 1
Specify T, calculate Q => Index = 2
where F = molar flow rate
H = specific molar enthalpy
E = total energy
Q = heat addition rate
A, B = enthalpy correlation constants
t = time
T = temperature
in = inlet stream
out = outlet stream
(b)
FIGURE 3-1. (a) Index. The index is the number of
differentiations required to obtain all derivatives ex-
plicitly. (b) Index example.
algebraic distillation model, be very careful
to pose the model so that the index does
not exceed 1.
Figure 3-1b contains an example of the
index problem similar to the type we find in
distillation dynamics. The example is a sim-
ple tank. A known pure component feed
enters a tank. Heat is added through a
heating coil. The outlet flow is set down-
stream from the tank.
We write the total differential material
and heat balances. We can calculate the
outlet enthalpy by a correlation of tempera-
ture. Our model has two differential and
two algebraic equations that can be solved
using the DAB approach.
The index of this problem can change
depending on what variables we set and
what variables we calculate. For example,
we could set the heat addition rate and
solve the problem to calculate the tempera-
ture. We would need to differentiate the
algebraic equation with respect to time once
to get the temperature differential. All the
unknown variables are then given in differ-
ential form, resulting in a problem of
index 1.
Alternatively, we could set the tempera-
ture and solve the problem to calculate the
heat addition rate. This problem is index
greater than 1. We would have to differen-
tiate the energy balance once to get the
heat addition differential. This differentia-
tion creates a second derivative of the total
energy, resulting in an index 2 problem. We
can reduce the problem to index 1 by adding
another algebraic equation for the heat. This
equation could be a heat transfer equation
for the heating coil.
Another numerical difficulty that appears
in the solution of differential equations is
stiffness. When we model the dynamics of a
chemical process, we are seeking the time
response of the process to changes in its
inputs. The process is modelled by many
differential equations. Each differential
equation takes a characteristic amount of
time to reach steady state. Some equations
can take hours whereas others might take
32 Practical Distillation Control
seconds. Stiffness can be measured as the
ratio of the longest to the shortest time for
these differential equations to approach
steady state. We solve the system of differ-
ential equations by stepping forward in time,
but we cannot take steps larger than those
necessary to solve the shortest differential
equation. As stiffness increases we must take
more steps to solve the equations, and the
simulation runs slower.
The conventional simulation methods
that we will develop in this chapter result in
index 0 problems. We avoid the index prob-
lem by solving the algebraic equations as
separate modules before integrating the
differential equations. In this way we do not
have to solve the differential-algebraic sys-
tem simultaneously. These models can be
stiff or not, depending on what problem we
need to solve.
3-2 DISTILLATION PROCESS
FUNDAMENTALS
3-2-1 Continuity Equations
The development of a dynamic distillation
model can be simple if you wisely use your
process knowledge and engineering judg-
ment. It can be difficult if you try to model
details that are not needed to solve the
problem at hand. You might find it interest-
ing that it can be easier to simulate the
dynamics of a distillation column than its
steady state. A steady-state model must
solve all the algebraic equations for mass
and energy simulation. In a dynamic simula-
tion we compute the change in mass and
energy and step forward in time. A dynamic
simulation is more intuitive with the way in
which we think about distillation operation
than a steady-state simulation.
To develop our dynamic model we begin
by writing dynamic continuity equations of
mass and energy for each unit operation
where mass or energy can accumulate. For
a distillation tower we develop the continu-
ity equations for the stages, condenser, re-
flux accumulator, bottom sump, reboiler,
thermowells, integrating controllers, and so
forth. Let us start by discussing the dynamic
material and energy balances for a general
system.
The dynamic continuity equations state
that the rate of accumulation of material
(mass or energy) in a system is equal to the
amount of material entering and generated,
less the amount leaving and consumed
within the system:
[
rate of accumulation]
of mass ( energy)
= (energy) flOW]
mto the system
_ [mass ( energy) flOW]
out of the system
+ ( energy) generated]
wIthm the system
_ [mass (energy) consumed]
within the system
The accumulation term is a first order time
derivative of the total mass or energy. The
flow terms are algebraic. This results in a
first order ordinary differential equation that
is usually nonlinear. We write material bal-
ances for each component (or the total flow
combined with all but one component flow)
and one energy balance. The component
balances are modelled as
dxjM .. ..
-- = - x.F + M.g
en
- M.con
dt I lout I I
(3-1)
where Z j = feed composition (mole frac-
tion)
x j = composition of material in sys-
tem (mole fraction)
M = material in system (Ib-moI)
F = flow rate Ob-moljh}
= rate of generation of compo-
nent i in system Ob-moljh)
Mj
con
= rate of consumption of com-
ponent i in system Ob-moljh}
Rigorous Modelling and Conventional Simulation 33
We will not consider reactive distillation
systems in our model, so there are no chem-
ical components generated or consumed in
any part of the distillation tower. The mate-
rial generation and consumption terms are
zero.
as
A general energy balance is represented
= EinF
in
- EoutFout + Q - W (3-2a)
E = U + PE + KE
U=H- PV
(3-2b)
(3-2c)
where M = material in system Ob-moO
F = material flow in/out of the
system Ob-moljh)
E = specific total energy (Btu /lb-
moO
Q = heat added to system (Btu/h)
W = work produced by the system
(Btu/h)
U = specific internal energy (Btu/
Ib-moO
PE = specific potential energy
(Btu/lb-moO
KE = specific kinetic energy (Btu/
Ib-moO
H = specific enthalpy (Btu/ Ib-
mol)
PV = pressure-volume work (Btu/
Ib-moO
The PE, KE, and PV terms of total energy
are negligible in distillation towers. We will
not consider vapor recompression systems
in this chapter, so there is no work pro-
duced by the system. Vapor recompression
will be considered in Chapter 23.
Heat can be added or removed from
many unit operations in distillation, such as
reboilers, condensers, or interstage heat ex-
changers. This means that the energy terms
considered in our distillation model are
equal to enthalpy. Our overall energy bal-
Rate 01 AccumuIaIIon - In - Out + Generation - Consumption
dM,
Tt -
Inhlal Condllon: M, (1-0) _
I- {1.2.3 ... n EO'
Dift'crenIiaJ-AJaebraic:
'Index-Oor!
Stiff I NOlI-Stiff
FIGURE 3-2. Integration block diagram.
ance can be written as
(3-2d)
We solve the system of differential equa-
tions, for all units in the process, by inte-
grating forward in time. First we specify a
set of initial conditions for the state vari-
ables at the starting time. Next, all differen-
tials are calculated from the material and
energy flows at the current time. Each
equation is integrated to the next time step.
The current time is increased by the time
step, and the integration process continues.
Figure 3-2 contains a representation of steps
in this process.
The integration procedure must be
started with a set of initial conditions for
each state variable. Initial conditions mayor
may not correspond to a steady-state solu-
tion. If a steady-state solution is not avail-
able for the first run, a set of consistent
values may be chosen. The solution will
34 Practical Distillation Control
usually converge, dynamically, to a bona
fide steady state as the integration pro-
ceeds. This method of initialization is useful
to get a dynamic simulation running quickly.
In fact, often it is numerically more stable
to converge to the steady state dynamically
than to solve for the steady state alge-
braically.
The following simple first order Euler
integrator will be used to integrate the
model:
dXI
X,H, = X, + dt at (3-3a)
X( t = 0) = Xo (3-3b)
where X is a function of time (t). The Euler
integrator is very easy to implement and is
very effective at solving a system of nonstiff
differential equations. The Euler integrator
works well for solving the distillation model
that we develop in this chapter. Models that
include vapor hydraulics and pressure dy-
namics are stiff. It is usually more efficient
to use an implicit integrator such as LSODE
(Hindmarsh, 1980) for such stiff systems. It
is important to remember the difficulty with
stiffness is that it slows down the simulation.
Because the computational speed of com-
puters has dramatically increased over a
short period, systems that were considered
a few years ago to be too stiff to solve
explicitly are now quickly solved by present
day computers. Explicit methods are gener-
ally easier to implement than implicit inte-
grators and the solution behavior of an ex-
plicit integrator is easier to follow than an
implicit integrator. This is a big benefit when
you are debugging your simulation.
We have discussed how to develop and
simulate dynamic models from continuity
balances. In the next few sections, we will
develop the principles necessary to compute
the terms of the differential equations
specifically for distillation towers. We will
look at the algorithms that we need to solve
the algebraic models and we will develop
the complete differential models of the vari-
ous subsections within the distillation tower.
Finally, we will discuss the performance of
these models on various computer platforms
and find how well the models match real
plant towers.
3-2-2 Vapor-Uquld Equilibrium
It is essential to have an accurate model or
correlation for the vapor-liquid equilibrium
(VLE) and physical properties of the com-
ponents in the distillation tower. There are
many useful sources for obtaining literature
data and correlating equations to represent
your system (Dauber and Danner, 1988;
Gmehling and Onken, 1977; Reid, Praus-
nitz, and Sherwood, 1977; Walas, 1985;
Yaws, 1977). The phase equilibrium of some
systems is essentially ideal. Raoult's law,
YiP =xiP/
(3-4a)
represents ideal systems. Others may have
constant relative volatility:
Some can be correlated simply as a polyno-
mial in temperature:
y.
K = ....!.. =A + B.T + C-T
2
+ ... (3-4c)
I Xi I I I
where Yi' Xi = vapor, liquid composition
(mole fraction)
P = total pressure (psia)
P/ = vapor pressure (psia)
aij = relative volatility
K
j
= vapor-liquid composition
ratio
A, B, C = correlation constants
Many chemical systems have significant
nonidealities in the liquid phase, and some-
times also the vapor phase. A more rigorous
VLE model is required. The fugacities of
the vapor and liquid phases are equal at
Rigorous Modelling and Conventional Simulation 35
equilibrium:
fr=/l
(3-5a)
Expressions can be written for the fugacity
of each phase:
(3-5b)
...I.ry.p = 'V.x .p.s exp I I (
_ p.S) )
'1'1 I II I I RT
(3-5c)
where fr, fjL = vapor, liquid fugacity
cPr, cPt = vapor, liquid fugacity coef-
ficient
P/ = vapor pressure (psia)
vf = specific critical volume
(ft3 jib-moD
'Yj = liquid activity coefficient
R = gas constant (ft
3
psiajlb-
mol OR)
T = temperature eF)
If an equation of state is available that
accurately represents both the vapor and
liquid phases, use Equation 3-5b. If the
liquid phase cannot be modelled by an
equation of state, which is usually the case,
use Equation 3-5c, which models the non-
idealities of the liquid phase by a correla-
tion of the liquid phase activity coefficient
'Y. The exponential term corrects the liquid
fugacity for high pressure (greater than 150
psia). Most of the nonideal effects for low
pressure (less than 150 psia) chemical sys-
tems can be represented by a simplified
form of Equation 3-5c, given as
The vapor phase is very close to ideal and
the exponential correction is negligible. It
should be noted that Equation 3-5d reduces
to Raoult's law if the liquid phase is ideal,
that is, the liquid phase activity coefficient is
unity.
The liquid activity coefficient can be
modelled for a component system by a cor-
relating equation (Gmehling, 1977; Walas,
1985) such as the Van Laar, Wilson, NRTL,
or UNIQUAC equations. Pure component
vapor pressure is modelled by the Antoine
equation,
(3-5e)
where C
t
, C
2
, C
3
are Antoine coefficients.
3-2-3 Murphree Vapor Phase Stage
Efficiency
Tray vapor composition can be calculated
from the relationships developed in the pre-
vious section. Mass transfer limitations pre-
vent the vapor leaving a tray from being in
precise equilibrium with the liquid on the
tray. This limitation can be modelled as a
deviation from equilibrium. Three types of
stage efficiencies are commonly used:
1. Overall efficiency.
2. Murphree efficiency.
3. Local efficiency.
Overall efficiency pertains to the entire col-
umn, relating the total number of actual to
ideal stages. Murphree efficiency pertains to
the efficiency at a specific stage whereas
local efficiency pertains to a specific loca-
tion on a single stage. We will use Mur-
phree stage efficiency in our model. We
base the Murphree stage efficiency on the
vapor composition's approach to phase
equilibrium. Figure 3-3 contains a geometric
interpretation of the Murphree vapor phase
stage efficiency for a binary system. The
Murphree stage efficiency is the ratio be-
tween the actual change in vapor composi-
tion between two stages and the change that
would occur if the vapor was in equilibrium
with the liquid leaving the stage.
36 Practical Distillation Control
E _ - - En +
n -
Yn - Yn-1
FIGURE 3-3. Murphree vapor phase stage effi-
ciency.
3-2-4 Enthalpy
Total energy reduces to enthalpy in our
distillation model. We must therefore calcu-
late the enthalpies of the vapor and liquid
streams as functions of temperature, pres-
sure, and composition. Liquid enthalpy is
not a function of pressure because liquids
are incompressible. Pressure effects on va-
por enthalpy are negligible for low to mod-
erate pressure systems. Enthalpy can be
easily correlated with composition and tem-
perature, based on a linear fit of heat capac-
ity with temperature. Selecting OF as the
reference temperature results in the corre-
lations
M = + B!-T2 (3-6a)
, , ,
hr = ArT + BrT
2
+ llHt (OF) (3-6b)
where hr, hf = vapor, liquid pure com-
ponent specific enthalpy
A, B = correlation constants
llHt (OF) = pure component heat
of vaporization at OF
We can obtain the correlation constants for
Equations 3-6a and 3-6b from heat capacity
data. Because we used OF for our enthalpy
reference temperature, the heat of vapor-
ization in Equation 3-6b must be that at
OF. Mixing rules are used to find multicom-
ponent properties. The vapor phase en-
thalpy is mixed ideally as the molar average
of the pure component enthalpies. The liq-
uid phase can contain significant nonideali-
ties (heat of mixing) unless the liquid activ-
ity coefficient is unity. Heat of mixing can
be modelled with
L 2 a In 1i
hmix = - RT l;:Xi---aT (3-6c)
I
The liquid activity coefficient in Equation
3-6c is modelled with the same correlation
we used for vapor-liquid equilibrium in
Section 3-1-2. Good results can be obtained
often, even with significant liquid nonideali-
ties, neglecting the heat of mixing or substi-
tuting a constant value.
Vapor and liquid enthalpy for mixtures
can be represented as
3-2-5 Liquid and Froth Density
Liquid density can be calculated from a
correlation or an equation of state. A useful
correlation for the pure component liquid
density (Yaws, 1977) is
Pi = AB-(1-T,)2/1 (3-8a)
Yaws gives constants for many common
components, or you can regress data to ob-
Rigorous Modelling and Conventional Simulation 37
tain the parameters. To compute the liquid
density of a mixture use
(3-Sb)
where Pi = pure component liquid density
Objft
3
)
PL = mixture liquid density Objft3)
M L = liquid mixture molecular
weight Ob lIb-mol)
Mi = pure component molecular
weight (Ib lIb-mol)
Tr = reduced temperature
A, B = pure component density corre-
lation constants
Vapor passing through a distillation tower
tray aerates the liquid, creating a froth. The
clear liquid density, Equation 3-8b, is cor-
rected for frothing by correlations, usually
obtained from the tray vendor design guides
or generalized correlations (Van Winkel,
1967). Computing the froth density as a
function of vapor rate is necessary if you
want to study flow hydraulics. Changes in
froth density with vapor rate can lead to
inverse response behavior in bottom sump
composition and level dynamics. There is no
need to compute the froth density continu-
ously if you are not interested in froth dy-
namics. Omitting this calculation at each
time step will improve the speed of the
simulation.
3-3 COMPUTER SIMULATION
3-3-1 Algebraic Convergence Methods
Our conventional simulation of a distillation
tower requires solving algebraic equations
before we integrate the system of ordinary
differential equations. We will use two sim-
ple and intuitive methods to solve implicit
algebraic relations. The first of these is the
Wegstein, or secant, convergence method.
Figure 3-4 presents this method graphically.
The Wegstein method requires the objective
function to be written in the form x = I(x).
The method requires two initial estimates of
the solution (Xl and x
2
), and the function is
evaluated at these two points (/1 and 12)' A
secant is then drawn between the two val-
ues. A third point ([3) is calculated where
the secant intersects the line where I(x) =
x, and the function is evaluated at this third
point (/3)' If 13 is sufficiently close to x
3
'
the solution has been found; if not, the last
two points are used as the next set of initial
guesses and the procedure is repeated. The
Wegstein method is very easy to implement
in a computer program.
The second convergence method that
is very useful for distillation problem is the
Newton-Raphson method. The Newton-
Raphson method requires the objective
function to be written in the form I(x) = O.
This method requires the first derivative of
the function to be evaluated, preferably an-
alytically. This requirement adds some com-
plication that the Wegstein method does
not have, but improves the rate of conver-
gence. Newton-Raphson is a very powerful
numerical method for thermodynamic cal-
culations, such as the bubble point. The
analytical derivative is readily available for
x - I(x)
,\;f'
, .

'"
,
12
2
I,
,
'"
I(x)
'"
'"
'"
,
'"
'"
'"
'"
,
'"
,
X,
X
2
X3
F1GURE 3-4. Wegstein convergence.
38 Practical Distillation Control
11
I(x) - 0
I(x)

1
-0
-1--1'
- I
XI - X2
FIGURE 3-5. Newton-Raphson convergence.
-
these calculations. Figure 3-5 presents the
Newton-Raphson method graphically.
The Newton-Raphson method requires
an initial estimate of the solution (Xl). An
analytical expression gives the derivative of
the objective function at the estimate. The
derivative is the tangent to the curve at this
point. The next estimate (x
2
) is computed
where the tangent intersects the X axis at
f = O. This solution procedure is very simi-
lar to the Wegstein method but a tangent is
used to compute the next estimate instead
of a secant.
These convergence methods require ini-
tial estimates for the solution. During the
dynamic simulation, the solution at the pre-
vious time step is used for these estimates.
This is very close to the solution at the
current time step. The algebraic equations
are solved very quickly, in only a few itera-
tions.
3-3-2 Equilibrium Bubble Point
Calculation
The development of an equilibrium stage
model requires the calculation of the vapor
T
LIquid
x
l=l-LYi=O
i
y
ps = exp(c
i
.
1 ,I T+C.
3,1
Converge T: Tk+1 = Tk + 1 /(
FIGURE 3-6. Equilibrium bubble point calculation
(binary system at constant pressure).
composition and temperature in equilibrium
with a known liquid composition at a known
pressure. This is a bubble point calculation.
Equation 3-5d will be used for our thermo-
dynamic model. A starting temperature is
guessed. The Antoine equation, Equation
3-5e, computes the vapor pressures. A sin-
gle dimensional Newton-Raphson method
is used to converge temperature and com-
pute the equilibrium vapor composition. The
relations and a graphical representation of
the solution are presented in Figure 3-6.
The function that we choose to solve is
unity minus the sum of the calculated vapor
compositions. The derivative of this func-
tion with respect to temperature is given in
Figure 3-6. The liquid activity coefficient is
a weak function in temperature and can be
excluded from the analytic derivative. This
method to compute the tray bubble point
temperature and equilibrium vapor compo-
sition is very fast and robust.
Rigorous Modelling and Conventional Simulation 39
3-3-3 Equilibrium Dew Point
Calculation
The reverse of the bubble point is the dew
point. Here we would like to calculate the
temperature and liquid composition in equi-
librium with a vapor. This calculation is'
necessary to model a partial condenser. The
entire model for a partial condenser is not
developed in this chapter, but the dew point
calculation is presented in Figure 3-7.
3-3-4 Distillation Stage Dynamic Model
We will now develop a general model for a
single distillation stage. We assume single-
flow pass tray hydraulics. Liquid enters the
tray through the downcomer of the tray
above. Vapor enters the tray from the tray
below. The vapor and liquid completely mix
on the tray. Vapor leaves the tray, in equi-
librium with the tray liquid composition,
and passes to the tray above. Liquid flows
over the outlet weir into the downcomer
and the tray below. The tray may contain a
feed or a sidestream. A dynamic model for
the tray will contain Nc differential material
T
LIquid
x
y
YiP
x=-
, 'Yi P/
p' = exp(c
1
.
, "T+ C .
3"
f= 1- LXi = 0
FIGURE 3-7. Equilibrium dew point calculation (bi-
nary system at constant pressure).
balances, where Nc is the number of com-
ponents in the system, and one overall en-
ergy balance. I like to model the material
balances as Nc - 1 component balances and
one overall material balance. You can model
the system as Nc component balances if you
prefer; both are equivalent. Do not forget
to include the liquid in the downcomer as
material in the stage.
If the tower pressure is high (greater
than 150 psi), the vapor can represent a
significant fraction of the total stage mass.
The liquid mass should be increased to ac-
count for the vapor if it is up to 30% of the
total material. If the vapor mass is more
than 30 or 40% of the total stage material
and variable, the vapor phase must be mod-
elled independently.
The change in specific liquid enthalpy is
usually very small compared to the total tray
enthalpy. Therefore, the energy balance may
often be reduced to an algebraic relation
that can be used to calculate the vapor rate
leaving the stage. A schematic of this model
is given in Figure 3-8.
Given this structure, the following proce-
dure can be used to solve the stage model.
The procedure is started from the bottom
and proceeds up through the column.
1. Calculate the equilibrium vapor compo-
sition and temperature from pressure and
liquid composition by a bubble point cal-
culation.
2. Calculate the actual vapor composition
from the Murphree vapor phase stage
efficiency.
3. Calculate the vapor and liquid enthalpy
from their composition and stage tem-
perature.
4. Calculate the clear liquid density from
composition and temperature.
5. Calculate the liquid froth density at the
stage vapor rate.
6. Calculate the liquid rate leaving the stage
from the Francis weir formula.
7. Calculate the vapor flow leaving the stage
from the energy balance.
8. Calculate the component and total mass
derivatives.
40 Practical Distillation Control
Material Balance
dM"
dt OK F + L,,+l + v,,-l - L" - v"
Component Balance
Energy Balance
d(h;M,,) L v L V
dt = hFF + h,,+lL,,+l + h,,-lv,,-l - h"L" - h"v"
dh
L
-" =0
dt
Liquid Hydraulics
(Francis weir) Ln = (C is a constant for units conversion)
Vapor-liquid Equilibrium
OGURE 3-8. Distillation tray model (2a).
The stage model is the keystone of the
dynamic distillation tower model. For each
stage in the column, we repeat it in a loop.
The stage model presented here can be
modified easily to meet your specific mod-
elling needs. For example, if precise froth
density is not important for your needs, the
correlation can be replaced easily with a
constant frothing factor. Sidestream prod-
ucts are added easily by an additional term
for the sidestream flow leaving the stage.
We have treated stage pressure as a con-
stant in this model. Variable pressure drop
requires the addition of a pressure drop
relation expressed in terms of vapor flow.
This requires the energy balance to be
solved as a differential equation. The change
in specific enthalpy is not neglected.
1. Calculate the specific enthalpy and liquid
composition by integration.
p'
eo K. = 11,,, I,"
',n P"
2. Calculate the tray temperature from en-
thalpy and tray composition.
3. Calculate the tray pressure from temper-
ature and liquid composition.
4. Calculate the vapor rate leaving the tray
from the tray pressure and tray vapor
flow hydraulics.
Vapor hydraulics are orders of magni-
tude faster than liquid composition dynam-
ics. This results in a very stiff set of differ-
ential equations. It will require a small time
step, a fraction of a second, to integrate the
equations. It probably will be more efficient
to use a stiff integrator, such as LSODE, to
solve this model. Fortunately, vapor and
pressure dynamics are not usually important
for product composition control problems
and we do not have to resort to this model
too often.
We can use the stage model we have
developed here to help classify different
Rigorous Modelling and Conventional Simulation 41
types of distillation stage models. Five mod-
els are given in Figure 3-9. The simplest
model (1) is appropriate when the molar
latent heats of the system components are
very close, producing equimolar overflow.
The vapor rate through each section of the
column is constant, eliminating the energy
balance. We integrate the differential mass
continuity equations, and the Francis weir
formula gives the liquid rate.
Models 2a, 2b, and 3 are the most widely
applicable dynamic distillation tower mod-
els used by the chemical industry. We have
developed model 2a here. I have found this
model sufficient to solve 95% of the indus-
trial distillation problems that I have worked
on. The remaining 5% were solved using
models 2b or 3. Use model 2b for high
pressure (greater than 150 psia) systems
where the vapor holdup is constant. Use
Modell
Equimolal Overflow
Negligible Vapor Holdup
Constant Pressure
Single Liquid Phase
Model2a
Negligible Vapor Holdup
Negligible Specific Enthalpy Change
Constant Pressure or Tray Pressure Drop
Model2b
Significant, but Constant, Vapor Holdup
Negligible Specific Enthalpy Change
Constant Pressure or Tray Pressure Drop
Model 3
Negligible Vapor Holdup
Significant Specific Enthalpy Change
Variable Tray Pressure Drop
Model 4
Significant Vapor Holdup
Variable Tray Pressure Drop
FIGURE 39. Distillation model classifications.
model 3 if there is significant variation in
tray pressure drop, or if pressure dynamics
are important. Distillation columns operat-
ing under vacuum might require model 3.
The differential equations become stiff when
solving model 3, so model 3 is best solved
using an implicit integrator, such as LSODE.
Model 4 is rarely applicable to industrial
columns. You need to use this model when
there is significant vapor holdup, that is,
pressure greater than 150 psia, and variable
tray pressure drop is important. Variable
tray pressure drop is usually only important
in vacuum columns. Yet vacuum columns
have insignificant vapor holdup, so model 3
is usually used. Applications for model 4 are
very rare, but should the need for it arise,
you will be ready. This model is probably
solved best using a nonconventional simula-
tion, that is, solving the system of differen-
tial and algebraic equations simultaneously.
The differential-algebraic integration pro-
gram DASSL can be used to solve the
model.
3-35 Bottom Sump
The model of the distillation tower bottom
sump is similar to that of the tower stage.
The bottom sump is modelled as an equilib-
rium stage. There is heat input from the
reboiler. Material holdup is large and vari-
able. Changes in specific enthalpy will not
be neglected. The dynamic model is pre-
sented in Figure 3-10.
The time derivative of the specific en-
thalpy can be approximated numerically by
backward difference. The specific enthalpy
is saved at each time step. The current
value is subtracted from the previous value
and the result is divided by the time step. It
is necessary to compute the specific en-
thalpy derivative so that the vapor rate leav-
ing the sump can be calculated from the
energy balance.
3-36 Condenser
The energy dynamics of the condenser are
small relative to column composition dy-
42 Practical Distillation Control
Material Balance
""_--s_
B, XI>
dM
- =L
l
- Vo-B
dt
d(Xb,jM)
= Xl ,.L
l
- YO Fo - Xb jB
dt ' , ,
Component Balance
Energy Balance
d( = hfLJ _ - + Q
R
dt L) dhLb
(
dM dhb V
Vo = + + Mdt - hfLI + QR /ho where dt .. Tt
"/jP/
Vapor-Liquid Equilibrium YO,j = --p;;Xb, j
FIGURE 310. Tower bottom sump model.
namics. We will specify the condensate tem-
perature of the liquid leaving the condenser
to allow for subcooling. The condenser duty
is equal to the latent heat required to con-
dense the overhead vapor to its bubble point
plus the sensible heat for any subcooling.
Our simulation reduces the enthalpy of the
overhead stream by the condenser duty be-
fore the stream is passed to the reflux accu-
mulator:
where H = stream specific molar enthalpy
F = stream molar flow rate
in = vapor inlet stream into the
condenser
out = liquid outlet stream from the
condenser
3-3-7 Reflux Accumulator
The reflux accumulator contains a signifi-
cant and variable quantity of material. There
is no heat into or leaving the accumulator
and there is no vapor leaving the accumula-
tor. The total enthalpy is computed by inte-
grating the energy balance directly. Dividing
the total enthalpy by the total mass gives us
the specific enthalpy of the liquid contents.
Integrating the material and energy bal-
ances produces the liquid composition and
specific enthalpy. We must compute the
temperature of the accumulator contents
from composition and enthalpy. Equation
3-10 can be solved, using Wegstein conver-
gence, to compute the temperature of the
contents given specific enthalpy and compo-
sition:
hL
T = L L (3-10)
EiXiAi + TEixiBi
where A, B are correlation constants.
Rigorous Modelling and Conventional Simulation 43
3-3-8 Feedback Controllers
Column inventories and product composi-
tions are controlled with proportional (P)
and proportional-integral (PI) controllers. I
do not like to use proportional-integral-
derivative (PID) controllers in dynamic dis-
tillation simulations. Derivative action per-
forms better in noise-free simulations than
in the real plant. If you use PI controllers in
the simulation, your expectations for the
plant will be conservative. Derivative action
can always be added in the field to improve
the response further.
Proportional controllers are used for sim-
ple level loops that do not require tight
setpoints. Proportional control will allow the
steady-state level to move as disturbances
pass through the process and smooth the
exit flow. Downstream disturbances are
minimized. This is very important if this
stream is fed to another unit operation. The
control law for a proportional controller is
where Co = controller output (%)
Kc = controller gain (%/%)
P
v
= process variable (%)
Sp = setpoint (%)
Bo = valve bias (%)
(3-11a)
All the controller variables are scaled to
instrument ranges. The process variable and
setpoint are divided by the transmitter span.
The controller output is a percentage of the
total range of the control valve. The actual
flow through the control valve is a function
of the controller output and valve flow char-
acteristics. The bias is equal to the desired
controller output when the difference be-
tween the process and setpoint is zero. Be-
cause the bias term is constant, there will be
a steady-state offset when the controller
moves the output to a different steady state.
A PI controller eliminates steady-state
offset. This is desirable for product compo-
sition control. The control law for a PI
controller is
Co = Kc[<Pv - Sp) + - Sp)dt]
(3-11b)
where Co = controller output (%)
Kc = controller gain (%/%)
Tj = controller reset time (h)
P
v
= process variable (%)
Sp = setpoint (%)
t = time (h)
Again, the controller variables are appropri-
ately scaled to instrument ranges. The inte-
gral term is evaluated by integrating the
operand with the differential continuity
equations. The bias term is the error time
integral. This term drives the error to zero
at all steady states because it varies.
3-4 WRITING A DYNAMIC
DISTILLATION SIMULATOR
With the ideas and references I have
presented in this chapter you now have
enough information to write a dynamic dis-
tillation simulator. You can write your simu-
lator using the conventional or nonconven-
tional approach; it is entirely up to you.
Both methods work. Each has benefits over
the other depending on your background
and the nature of your model. I have used
the following procedure to write a conven-
tional simulator using model 2a.
1. Create a data file with the tower me-
chanical, instrument, and physical prop-
erty data.
2. Specify the initial conditions of the state
variables. If you have made previous runs,
a snapshot of the process variables is
saved in a data file that can be used as
the initial conditions. If this is the first
run, specify a set of consistent initial
conditions.
3. Calculate the derivatives of each stage
starting with the bottoms sump and
working up the column.
44 Practical Distillation Control
4. Calculate the derivatives for the con-
denser and reflux accumulator.
5. Calculate the derivatives for the con-
trollers containing integral action.
6. Integrate all the derivatives.
7. Increment time by the integration time
step.
8. If necessary, print the display variables
and save a snapshot of the current pro-
cess variables.
9. Go back to step 3 and continue.
In step 2 I suggest that if this is a new
simulation consistent initial conditions be
chosen. To do this I ask the user to input
the feed stream, reflux, heat input, and
product compositions. Tray liquid composi-
tions are then computed as a linear profile
from top to bottom. A constant vapor rate is
used based on the heat input. The liquid
rate is set using equimolar overflow and
assuming the feed is a liquid. Bubble point,
enthalpy, liquid density, and froth density
are calculated for each tray. The dynamic
simulation is then run from these conditions
until it converges to a bona fide steady
state.
A method that can be used to simplify
programming of the simulator is to inte-
grate the differential equations at each stage
instead of integrating the entire tower. The
way I have presented these models and sim-
ulation approach is well suited to this tech-
nique. This avoids creating arrays for all the
differential equations. If you use this tech-
nique, be very careful to make sure that
your variables do not get out of time step.
Remember, once you integrate a differen-
tial equation, the state variable is at the
next time step.
3-5 PLANT -MODEL VERIFICATION
Other chapters in this book discuss method-
ologies on how to design, control, and oper-
ate distillation towers. The dynamic models
developed in this chapter may be used as a
tool to apply or test these methods. Before
you begin to apply your simulation to a real
Tray Number
Top 38
30
20
10
Bottom
150 170 190 210 230
Temperature (De". F.I
- Simulation 0 Plant
(a)
Sta"e 32 Temperature (De" FI
157.2
158.8
0
158.4
0
158.0 0
0
0
155.8 1..-.-1...---'_-'---'-_-'----'-__ -'-_'-----'
o 20 40 80 80 100 120 140 180 180 200
Time IMin.)
- Simulation 0 Plant Data
(b)
FIGURE 311. (a) Alcohol tower steadystate tem-
perature profile verification and (b) dynamic response
verification.
Rigorous Modelling and Conventional Simulation 45
tower, you need to verify that the model
does mimic plant operation.
A good model will match the plant at
steady state and accurately track the pro-
cess during dynamic upsets. A verification
of the model includes a check of the overall
material and energy balances, steady-state
temperature profile, and open loop dynamic
disturbance tests. A final test of closed loop
performance will ensure that your model is
accurate. Once you have verified many dif-
ferent columns, you will have confidence in
modelling a new design that cannot be veri-
fied until startup.
Figures 3-11a and 3-11b are two verifi-
cation plots completed for a real 38-tray
alcohol-water column. Figure 3-11a is the
steady-state temperature profile computed
using the simulation with plant data over-
laid. Figure 3-11b is a dynamic open loop
test of a tray temperature for a step change
in reflux flow. These curves are illustrative
of my experience in successfully verifying
dynamic distillation models with plant data.
The simulation can now be used to develop
new and improved process control strate-
gies.
3-6 COMPUTATIONAL
PERFORMANCE
Computational performance and productiv-
ity have progressed well beyond our expec-
tations over the last 30 years. It is likely to
exceed our current expectations in the next
few decades.
A decade ago, complex engineering prob-
lems were only discussed qualitatively. To-
day, these problems are quickly quantified
and solutions put into production.
We now have the power of supercomput-
ing at an affordable price, either on the
desk or tied to it through a network of
computers. System software and services
have reached the point that engineers with-
out extensive training in computer technol-
ogy can program steady-state and dynamic
models. The modelling activity itself gives
TABLE 32 Computer System Execution Speed
RealTime
Computer MFLOPS
a
Factor
b
Cray Y-MP (1 processor) 161 860
HP9000/720 10 580
IBM RS/6000-530 14 390
DECstation 5000/200 4 250
VAX 8700 1 90
Compaq 386/25 MHz 0.2 27
aDongara (1991).
bSimulation based on a 7.S ft, 38 tray alcohol-water col-
umn with a setpoint change:
process real time
real time factor (RTF) = I . . .
e apsed sUDulatlon time
the engineer improved process insight and
solves very significant problems.
Table 3-2 contains the run time perfor-
mance, from various computing platforms,
of a dynamic distillation simulation. These
data correspond to the 38-tray binary alco-
hol-water distillation tower that we verified
in the previous section.
We measure execution performance in
terms of a real time factor. This is the
number of times faster than real time the
simulation runs. Computing performance is
growing geometrically each year. The desk-
top, laptop, or fingertop computer of tomor-
row will be fast enough that entire plants of
five or more towers will be added to the top
of Table 3-2.
3-7 CONCLUSIONS
I have presented an approach to modelling
distillation dynamics that builds on engi-
neering experience and judgment. The ap-
proach is based on modelling the physical
elements as modules and connecting the
components together in the body of a simu-
lation. The modules chosen are those that
are important to solve our engineering ob-
jectives for the problem at hand.
We developed a rigorous model for a
general distillation tower. The model is
based on first principles. We implemented
the model in a computer simulation using
46 Practical Distillation Control
an explicit integrator. The simulation accu- w Weir length ft
rately predicted operating data from a real x Liquid composition mole fraction
plant. The simulation runs very fast on a y Vapor composition mole fraction
personal computer. Written in FORTRAN z Feed composition mole fraction
77, it is portable to many other computer
platforms. We will use this simulator later Greek
in Chapter 18. We will use model 2a to e Error
understand the process dynamics and de-
4J
Vapor fugacity coeffi-
sign a control system of an extractive distil- cient
lation system.
'Y
Liquid activity coeffi-
cient
3-8 NOMENCLATURE
11 Liquid specific volume ft3 lIb-mol
p Liquid density lb/lb-mol
B Controller bias %
'T'i
Controller reset time h
C Correlation coefficient
E Murphree stage effi- Subscripts
ciency con Material consumed
E Specific total energy Btu/lb-mol within system
F Feed flow rate Ib-moIjh gen Material generated
f
Fugacity psia within system
H Specific enthalpy Btu/lb-mol in Flow in system
h Specific enthalpy Btu/lb-mol L Liquid
Kc
Controller gain %j% len Length
K Vapor-liquid compos- mix Heat of mixing
ition ratio out Flow from system
KE Specific kinetic
energy Btu/lb-mol Superscripts
L Liquid flow rate lb-moljh
*
Equilibrium
M Total material lb-mol L Liquid
Nc
Number of compo- s Saturated
nents (vapor pressure)
P
v
Controller process V Vapor
variable % Time rate of change
P Pressure psia
PE Specific potential
References
energy Btu/lb-mol Brenan, K. E., Campbell, S. L., and Petzold,
PV Pressure-volume work Btu/lb-mol L. R. (1989). Numerical Solution of Initial-
Q
Heat input Btu/h
Value Problems in Differential-Algebraic Equa-
R Gas constant
tions. New York: North-Holland.
Sp Controller setpoint %
Daubert, T. E. and Danner, R. P. (1988). Data
T Temperature
OF
compilation, tables of properties of pure com-
Tr
Reduced temperature
pounds. DIPPR, AIChE, New York.
Dongarra, J. T. (1991). Performance of various
(actuaIj critical)
computers using standard linear equations
t Time h
software. University of Tennessee, Knoxville,
U Specific internal ener-
TN.
gy Btu/lb-mol
Franks, R. G. E. (1972). Modeling and Simu-
V Vapor flow rate Ib-moIjh !ation in Chemical Engineering. New York:
W Work produced Btu/h Wiley-Interscience.
Rigorous Modelling and Conventional Simulation 47
Gmehling, 1. and Onken, U. (1977). Vapor-
Liquid Equilibrium Data Collection. Frankfurt,
Germany: DECHEMA.
Hindmarsh, A. C. (1980). LSODE: Livermore
solver for ordinary differential equations.
Lawrence Livermore National Laboratory.
Luyben, W. L. (1990). Process Modeling Simula-
tion and Control for Chemical Engineers,
2nd ed. New York: McGraw-Hill.
Reid, R. C., Prausnitz, 1. M., and Sherwood,
T. K. (1977). The Properties of Gases and
Liquids. New York: McGraw-Hill.
Van Winkel, M. (1967). Distillation. New York:
McGraw-Hill.
Walas, S. M. (1985). Phase Equilibrium in Chemi-
cal Engineering. Boston: Butterworth.
Yaws, C. L. (1977). Physical Properties. New
York: McGraw-Hill.
4
Approximate and
Simplified Models
Antonis Papadourakis
Rohm and Haas Company
John E. Rijnsdorp
University of Twente
4-1 INTRODUCTION
Although the basic principle of distillation
is simple, modelling columns with many trays
leads to large models with complex overall
behavior. In the past, this has encouraged
the development of many shortcut models
for process design purposes. However, with
the present availability of cheap and power-
ful computer hardware and software it has
become possible to utilize rigorous static
and dynamic models (see preceding chap-
ter) for off-line use.
Why do we then still need approximate
and simplified models? First, we can better
understand the behavior of a distillation
process. Without understanding it is almost
impossible to design good control structures
or to improve process controllability. Sec-
ond, simple models are needed in process
computer systems for optimizing and ad-
vanced regulatory control, where strict
real-time and reliability conditions prohibit
complex ones. Third, advanced plant design
brings about strongly integrated process flow
schemes, which cannot be segregated into
individual process units. Then the overall
complexity of integrated rigorous dynamic
models is even beyond the possibilities of
present-day work stations.
48
In what follows, an attempt is made to
look at simple static and dynamic models in
a systematic way, classify them according to
their major characteristics, and demonstrate
the application of the best among them.
4-2 CLASSIFICATION OF SIMPLE
MODELS
In general, the model structure has to be in
line with the purpose for which it will be
used. Table 4-1 shows some common uses
of models.
For continuous distillation processes,
highly simplified static models can be used
to track relatively small changes in the opti-
mum operation point, provided the model is
validated by means of a rigorous model or
by means of test run data (Hawkins, Tolfo,
and Chauvin, 1987). Combining optimizing
control with model adaptation requires non-
linear ARMA-type models (Bamberger and
Isermann, 1978, have used Hammerstein
models for this purpose). In the case of
batch distillation the major dynamics are in
the kettle, but the rectifier dynamics may
not be completely ignored (Betlem, 1991).
Here dynamic optimization techniques are
only feasible in practice if the overall model
is of low order, say of order 10 or less.
TABLE 4-1 Purposes for Simple Distillation Models
Understanding process behavior
Optimization of process operation
Startup, switchover, or shutdown
Assessing process controllability
Designing regulatory control structures
Composition estimation
Model-based control
Simulation, particularly of integrated plants
(for operability
and controllability analysis and for
operator training)
Startup, shutdown, and major switchovers
can only be analyzed by nonlinear models.
For complete plants, consisting of many
process units, order reduction is mandatory,
but not very easy. For an approximate anal-
ysis of minor switchovers, linear models are
sufficient and high order does not impose
any great difficulties.
There is a need to assess process con-
trollability in early design phases, when
the flowsheet has not yet been frozen.
Here singular value decomposition (see
Skogestad and Morari, 1987a, for a brief
description and a worthwhile extension) has
proven to be a simple and valid method
(Perkins, 1990). This method requires a (lin-
ear) frequency response matrix between
process inputs and the variables to be con-
trolled.
The same model can be used for de-
signing the basic control structures, which
includes a separation between single-
input-single-output (SISO) and multiple-
input-multiple output (MIMO) subsystems.
For the latter, the block relative gain array
method (Arkun, 1987) has proven to be
quite useful.
Real-time composition estimation and
model-based control are only practical when
the model order is low. In addition, in case
of high-purity separations, nonlinearities
play an important role, which brings us again
to the problem of order reduction in nonlin-
ear models (see Section 4-6 for more detail).
Finally, in later process design phases,
simulation of complete plants is desirable
Approximate and Simplified Models 49
for a more thorough analysis of plant oper-
ability, controllability, and switchability, and
for operator training. In the latter case,
speed of computation is more critical than
in the former case, which warrants model
simplification.
4-3 SIMPLE STEADY-STATE
MODELS
As has already been indicated in the pre-
ceding section, simple models can be used
for tracing optimum operation points, pro-
vided they are validated for the actualoper-
ating region. A very simple model has been
proposed by Shinskey (1967):
V/F = fJc In Sc (4-1)
where V = vapor flow rate (mol/s)
F = feed flow rate (mol/s)
fJ c = column characterization factor
S c = separation factor, equal to
where LK = mole fraction of the light key
component
HK = mole fraction of the heavy
key component
D = top product
B = bottom product
The column characterization factor is a
function of the number of trays, the tray
efficiency, and the relative volatility of the
two key components. It is the parameter to
be fitted to the operating region.
Equation 4-1 does not contain the influ-
ences of feed enthalpy and feed tray loca-
tion. The first type influence can be ac-
counted for by defining separation factors
for each column section (Rijnsdorp, 1991):
LR/D = fJR In SR (4-3)
Vs/B = fJ
s
In Ss (4-4)
50 Practical Distillation Control
where L R = liquid flow in the rectifying
section (moljs)
Vs = vapor flow in the stripping
section (moljs)
S R' S S = characterization factors for
the rectifying and the strip-
ping section, respectively, de-
fined by
SR = (LKD/LKF)(HKF/HK
D
) (4-5)
Ss = (LKF/LKB)(HKB/HK
F
) (4-6)
where LK F = mole fraction of the light
key component at the feed
tray
HK F = mole fraction of the heavy
key component at the feed
tray
Equations 4-3 to 4-6 can be combined, with
substitution of 4-2, into
Most column design methods implicitly as-
sume that the feed enters at the optimum
location. However, in actual operation the
feed tray location is rarely adapted to the
prevailing conditions (although this could
be quite worthwhile for large deviations
from the design conditions). The effect on
the column performance can be accounted
for by comparing the ratio of the key com-
ponent mole fractions at the feed tray ac-
cording to Equations 4-3 and 4-4 to the
optimal mole fractions, which are given by
the intersection of the operating lines for
the two column sections:
YF = (LR/VR)X
F
+ (D/VR)XD (4-8)
YF = (LS/VS)XF - (B/VS)XB (4-9)
where x F = component mole fraction in
the liquid on the feed tray
YF = component mole fraction in
the vapor above the feed tray
Solution yields
The difference between the actual (accord-
ing to Equations 4-3 and 4-4) and the opti-
mal (according to Equation 4-10) ratio of
the two key component fractions is now
used to increase the values of the character-
ization factors in a suitable way, for in-
stance,
f3R = f3R.OPt[ 1 + cR(pos{( LKF/HKF ) opt
- (LKF/HKF ) act}) 2] (4-11)
f3s = f3s.opt[l + cS(pos{(LKF/HKF)act
- (LKF/HKF )opt})2] (4-12)
where C R' C S = parameters to be deter-
mined by comparison with
tray-to-tray calculations
post } = its argument if the latter is
positive, and equals zero if
the latter is nonpositive
4-4 PARTITIONING OF THE
OVERALL DYNAMIC MODEL
4-4-1 Introduction
At first sight, a complete dynamic model of
distillation column behavior offers a bewil-
dering interaction between many variables.
For every tray there are pressures, vapor
and liquid flows, and vapor and liquid con-
centrations. Similar variables also appear
for the reboiler, the condenser, and the top
and bottom accumulators. It would be very
worthwhile to partition these variables into
a number of clusters, so that the overall
dynamics can be represented by less compli-
cated submodels, each associated with a
particular control issue.
As will be shown in the next subsection,
this can be realized under relatively mild
assumptions, with submodels for the follow-
ing clusters of process variables (see Figure
4-1):
1. Pressure and vapor flow rates, associated
with pressure and pressure drop control.
2. Liquid holdups and flow rates, associated
with liquid level control in the top and
bottom accumulators.
C
Tc
L.AM
Approximate and Simplified Models 51
3. Vapor and liquid concentrations, associ-
ated with temperature and composition
control.
4. Condenser temperatures and heat flow
rates.
5. Reboiler temperatures and heat flow
rates.
The first three submodels have no inter-
actions, but there is an important action
from the first to the second and third, and
V.AP
rectifyiq
IeCtial
PAY. rec:tifyq L. P.t v.
H
T"
V.AP
.trippiDa
HClim
C Cooliaa medium flow
R Refiu
B Bauoma flow
aectim
LIAM
PAY. Itrippq L. P .t V.
aectim
H HNIiDa mecli1llll flow
D DiIli1late flow
F Feeclflow
FIGURE 4-1. Partitioning of the dynamic model.
XIAJI.
Itrippq
IOClim
52 Practical Distillation Control
from the second to the third. In fact, the
influences from the latter on the former
mainly go through the reboiler and the con-
denser submodels, which cause inherent
regulation and deregulation, respectively.
4-4-2 Assumptions
The assumptions are:
Equal (pseudo) molal overflow.
Small influence from liquid tray holdup
on tray pressure drop.
Tray pressure drop small compared to the
absolute value of the pressure.
Negligible heat loss.
In distillation column design, a popular
assumption is that of equal molal overflow.
In terms of the relations between specific
enthalpies and compositions this assump-
tion is equivalent to (where the summation
is over all except the least volatile compo-
nent)
HL(s,t) =Ho(s,t) - Eajxj(x,t) (4-13)
and
Hv(s, t) = Ho(s, t) + Hc(s, t)
- EajYj(s, t) (4-14)
where xis, t) = liquid mole fraction for
component j on tray s
yis, t) = vapor mole fraction for
component j on tray s
a j = sensitivity coefficient for
component j, assumed to
be equal for vapor and
liquid
Ho(s, t) = specific liquid enthalpy of
the least volatile compo-
nent (J ImoI)
Hc(s, t) = heat of vaporization of
the least volatile compo-
nent (J ImoI)
Evidently, in the enthalpy-composition hy-
perspace this corresponds to parallel hyper-
planes for liquid and vapor enthalpy.
Robinson and Gilliland (1950) have ex-
tended the scope by introducing pseudo-
molecular weights. In this way, unequal sen-
sitivity coefficients for vapor and liquid can
be equalized, The accuracy of the method
has been demonstrated for the design of a
column separating the nonideal mixture of
ammonia and water. Chen (1969) has shown
that this method is also consistent for the
dynamic case.
In actual trays, changes in liquid hold-
up will influence the pressure drop across
the tray. For trays with downcomers within
the normal operating range (where most of
the liquid holdup variations appear in the
downcomer), the liquid height on the tray
proper does not vary much, so the influence
of the vapor flow is most important. Of
course this no longer holds when the trays
are overloaded, that is, when the downcom-
ers start to overflow.
The third assumption even holds for vac-
uum operation, possibly with exception for
trays near the top where the absolute pres-
sure is very low. Finally, the fourth assump-
tion is warranted for well insulated big
columns and for small columns provided
with a compensatory heating system.
On the basis of the assumptions, the en-
ergy balance of the tray can be reduced to
the following, very simple balance for the
vapor flows (Rademaker, Rijnsdorp, and
Maarleveld, 1975):
dP(s, t)
B(s,t) dt =V(s-1,t)-V(s,t)
(4-15)
where B(s, t) = effective vapor capacity of
the tray
P(s, t) = pressure on the tray
The physical interpretation is that an in-
creasing pressure requires compression of
the vapor holdup and condensation of vapor
for heating the tray to the new equilibrium
temperature, both resulting in a difference
between the vapor flow entering the tray
and that leaving the tray.
The effect of resistance to mass transfer
between vapor and liquid can be accounted
for by a more refined version of Equation
4-15:
dP*(s, t) dP(s, t)
B*(s,t) dt +Bv(s,t) dt
= V(s - 1, t) - Yes, t) (4-16)
where Bv(s, t) = vapor compression term
of B(s, t)
B*(s, t) = remainder of B(s, 0
P*(s, t) = pressure in equilibrium
with the liquid tempera-
ture, determined by
dP*(s, t)
B*(s, t) dt
[pes, t) - P*(s, t)]
Rp(s, t)
(4-17)
where R/s,O is the resistance to mass
transfer between the phases on tray s
(Pa s/mol).
4-4-3 Propagation of Vapor Flow and
Pressure Responses
The equations for the vapor flow balances
across the trays (see Equation 4-15, or
Equation 4-16 together with Equation 4-17)
form the basis of the submodel for the pres-
Approximate and Simplified Models 53
sures and the vapor flows. We just have to
extend it with an equation for the tray pres-
sure drop. According to the second assump-
tion, the latter only depends on the vapor
flow rate. Now Equation 4-16 can be ap-
proximated by
dP*(s, t) dP(s, t)
B*(s, t) dt + Bv(s, t) dt
= fv{P(s - 1, t) - pes, t)}
-fv{P(s, t) - pes + 1, t)} (4-18)
where f v{ } represents the relationship be-
tween vapor flow rate and pressure drop.
For n trays, Equations 4-17 and 4-18 lead
to a set of state equations with the 2n tray
pressures [the pes, Os and the P*(s, t}s] as
state variables. In analogy to electrical net-
works it can also be interpreted as a resis-
tance-capacitance ladder (see Figure 4-2)
with nonlinear resistances. Note that this
network does not tell the complete story. In
fact, the condenser reacts to the change in
top vapor flow and so provides a feedback
to the submodel. A similar influence occurs
at the bottom end via the reboiler.
4-4-4 Propagation of Liquid Flow and
Uquld Holdup Variations
The responses of liquid holdups and flow
rates are determined by the total mass bal-
(ft/AP)-1
T
FIGURE 4-2. Electrical network interpretation of propagation of vapor flow and
pressure responses.
54 Practical Distillation Control
ance of the tray:
dML(s, t)
dt =L(s+l,t)-L(s,t)
+V(s - 1,t) - V(s,t)
(4-19)
With Equation 4-15 and with the hydraulic
equation for the tray,
ML(s,t)
= fM[L(s, t), V(s - 1, t)] (4-20)
we find
dL*(s, t)
'TL(S,t) dt
=L*(s + 1,t) -L*(s,t)
dP(s, t)
+ [1 - KVL(s, t)]B(s, t) dt
(4-21)
where, by definition,
L*(s, t) = L(s, t) - KVL(s, t)V(s - 1, t)
(4-22)
(
iJML(S, t) )
'TL(s, t) = (4-23)
iJL(s, t) V(s-l,t)
KVL( s, t) = ) ML (4-24)
As K VL is poorly known, we have taken its
average value, which also helps to keep the
model simple. When the total pressure drop
across the trays is small compared to the
average (absolute) pressure, the propaga-
tion of pressure variations is fast compared
to the propagation of liquid flow variations
(Rademaker, Rijnsdorp, and Maarleveld,
1975). Then pes, t) can be replaced by its
average pet).
4-4-5 Propagation of Vapor and Liquid
Concentrations
Finally, we come to the concentration re-
sponses. These are determined by the par-
tial mass balances of the trays:
d[ ML(s, t)xj(s, t)]
dt
d[ Mv(s, t)Yj(s, t)]
+ dt
= L(s + 1, t)Xj(s + 1, t)
-L(s,t)xis,t)
+ V(s - 1, t)Yis - 1, t)
- V(s, t)Yj(s, t) (4-25)
where xj(s, t) = mole fraction of compo-
nent j in the liquid on
tray s
Yj(s, t) = mole fraction of compo-
nent j in the vapor on
tray s
Subtracting the total mass balance, multi-
plied by xj(s, t), yields
dxj(s, t) dYj(s, t)
ML(s, t) dt + Mv(s, t) dt
= L(s + 1, t)[ Xj(s + 1, t) - xj(s, t)]
+ V(s - 1, t)[y/s - 1, t) - Yj(s, t)]
+[Yj(s,t) -xj(s,t)]
dP(s, t)
XB*(s, t) dt
(4-26)
where use has been made of Equation 4-15.
Note that the liquid flow from above, the
vapor flow from below, and the rate of
change of the tray pressure constitute the
inputs to this tray model for the concentra-
tions. In their tum the top concentration
influences the top pressure by way of the
condenser, and the bottom concentration
influences the bottom pressure by way of
the reboiler.
For the case of theoretical trays, the va-
por concentrations are related to the liquid
composition, so Equation 4-26 constitutes a
set of (n
j
- 1) (the number of components
minus 1) times n (the number of trays)
simultaneous differential equations. In the
binary case, there is just one differential
equation per tray.
45 LINEAR MODELS
The linear models for distillation columns
are essentially linearized versions of the
original nonlinear models that are devel-
oped based on mass and energy balances
around each equilibrium stage. In the vast
majority of published work, only composi-
tion dynamics are considered, and in very
few cases the flow dynamics have been su-
perimposed to obtain a more accurate de-
scription. Usually, the nonlinear model is
linearized around some steady state (Le.,
the operating state) of the column, and the
study of the resulting linear model reflects
the dynamics of the process around this
steady state. For a more comprehensive ex-
amination of the existing linear models of
distillation columns, we distinguish them ac-
cording to the domain they are presented
in:
1. Linear models in the time domain
2. Linear models in the Laplace domain
3. Linear models in the frequency domain
Because a model presented in one domain
can be transformed to a model of another
domain, there is an ambiguity in such a
classification. However, as we will see in the
following paragraphs, there are certain do-
main-dependent properties of these models.
4-51 Linear Models in the Time
Domain
Modal reduction techniques can be used for
the approximation of high-order linear sys-
tems in the time domain (Bonvin and
Mellichamp, 1982a). Several techniques are
available in the literature, and a critical
review of six of them is given by Bonvin and
Approximate and Simplified Models 55
Mellichamp (1982b). These techniques can
be applied to chemical engineering pro-
cesses in general, and to distillation in par-
ticular.
A modal reduction technique especially
suited for staged processes was developed
by Georgakis and Stoever (Georgakis and
Stoever, 1982; Stoever and Georgakis, 1982).
Their method of nonuniform lumping (and
uniform lumping as a special case) can be
used for the order reduction of tridiagonal
dynamic models of staged processes. The
advantages of their method are good reten-
tion of the dominant time constants, small
deviation of the eigenrows, physical signifi-
cance of the states of the reduced model,
and flexibility in the choice of lumping. The
disadvantages of the method include the
inaccurate simulation of the dead times, as
well as that the initial response of the re-
duced model can be fundamentally different
from that of the original model. Also, the
requirement for a tridiagonal form of the
Jacobian matrix of the system limits the
applicability of their method.
452 Linear Models in the Laplace
Domain
It has been observed by many distillation
control practitioners that the nonlinear dy-
namic composition response of distillation
columns often resembles a linear first-order
response. It is also known from theoretical
studies (Levy, Foss, and Grens, 1969) that
the dominant time constant of a distillation
column is well separated from the rest and
is nearly the same regardless of where a
disturbance is introduced or where compo-
sition is measured (Skogestad and Morari,
1987b). In addition, if constant molar flows
are assumed, then the flow dynamics are
decoupled from the composition dynamics,
and the dominant part of the dynamics can
be captured by modelling only the composi-
tion dynamics.
The preceding realizations have had a
twofold effect. First, they have led many
investigators to attempt representation of
56 Practical Distillation Control
the composition dynamics of distillation
columns using models of the simple forms
or
K( TOS + 1)e-
dS
(TIS + 1)(T
2
S + 1)
where T 1 is the dominant time constant of
the column. Many investigators focused their
attention on developing such transfer func-
tions for distillation columns, assuming that
a linearized model for the column is avail-
able. Kim and Friedly (1974) used the fact
that distillation columns exhibit a dominant
eigenvalue well separated from the rest (i.e.,
Al A2 > ... > An) to approximate a
normalized transfer function of the form
1
(4-27)
by a second-order-plus-deadtime model,
where the two dominant eigenvalues are
used as the time constants and the rest of
the eigenvalues are lumped as the dead-
time:
with
1
T=--
I A '
1
1
T---
2 - A'
2
(4-28)
(4-29)
The transfer function given by Equation
4-27 has no numerator dynamics. Such
transfer functions arise in staged systems
when considering the response of a variable
at one end of a cascade to a disturbance
introduced at the opposite end. In cases
where the transfer function of the column
contains numerator dynamics (i.e., zeros),
the numerator and denominator dynamics
of the column can be approximated sepa-
rately (Kim and Friedly, 1974; Celebi and
Chimowitz, 1985; Papadourakis, Doherty,
and Douglas, 1989). The approximation of
the denominator dynamics provides the
dominant time constants (poles) and a con-
tribution to the deadtime, whereas the ap-
proximation of the numerator dynamics
provides the zero(s) and the remainder of
the deadtime. In other words, a normalized
transfer function of the form
n?!.1(1 - s/,.,,;)
n;=1(1 - s/Aj )
can be approximated by
where
1 1
TO=--'
""1
T=--
1 A '
1
d=
j=3 Aj i=2 ""j
(4-30)
(4-31)
1
T---
2 - A2'
(4-32)
with Al A2 > ... > An and ,.". > ""2 >
... >""m
Celebi and Chimowitz (1985) provide a
method for obtaining the rigorous transfer
function (Equation 4-30) for a distillation
column, but their method is useful only
when considering the response of the col-
umn to changes in feed composition.
In general, if the full linearized model of
a distillation column is available in analyti-
cal form in the Laplace domain, many or-
der-reduction methods can be used to pro-
duce an approximate model similar to the
one given by Equation 4-31. Surveys of these
methods (i.e., moment matching methods,
cumulant matching methods, Routh approx-
imations, impulse energy approximations,
least squares approximations, etc.) are given
by Genesio and Milanese (1976) and by
Papadourakis, Doherty, and Douglas (1989).
However, the applicability of the preceding
methods is limited because the full transfer
function model of a distillation column is
seldom available in analytical form.
The second effect of the realization that
the nonlinear distillation dynamics are dom-
inated by a time constant well separated
from the rest is a large number of attempts
to relate the dominant time constant to
steady-state parameters. Many investigators
(Moczek, Otto, and Williams, 1965; Bhat
and Williams, 1969; Wahl and Harriot, 1970;
Toijala (Waller) and his co-workers,
1968-1978; Weigand, Jhawar, and Williams,
1972; Skogestad and Morari, 1987b, 1988)
have tried to estimate this time constant,
which appears to be equal to "the change in
column holdup of one component over the
imbalance in supply of this component,"
that is,
change in total column holdup
of one component (mol)
T = ----------------------
e imbalance in supply of this
component (mol/min)
Skogestad and Morari (l987b) developed an
analytical expression for the calculation of
the dominant time constant based on
steady-state parameters only. Such a formu-
lation also had been suggested by Rijnsdorp
(see discussion in Moczek, Otto, and
Williams, 1963). According to Skogestad and
Morari's work, the dominant time constant
of the composition dynamics of a two-prod-
uct distillation column is given by
(4-33)
where x = X initial - X final
NT = total number of trays (exclud-
ing the condenser and the re-
boiler)
M L = liquid holdup of tray i
= liquid holdup of the reboiler
MLD = liquid holdup of the con-
denser
Approximate and Simplified Models 57
the subscript f denotes the final steady state
reached by the column after a step change
has been introduced and subscript initial
refers to the initial steady state of the col-
umn.
The derivation of Te is based on two
assumptions:
Assumption I
The flow dynamics are immediate (i.e., the
holdups and the flows change instantane-
ously).
Assumption II
All the trays have the same dynamic respon-
se, which corresponds to treating the col-
umn as a large mixing tank.
The following comments apply on the
derivation of Te:
1. The column model, on which the deriva-
tion of T e is based, is not linearized.
2. Te depends on the magnitude and direc-
tion of change.
3. Equation 4-33 applies to any component
in a multicomponent mixture.
4. Equation 4-33 applies to any change that
changes the external material balance of
the column:
5. To compute Te' a steady-state model of
the column is needed in order to calcu-
late the compositions on all stages for
both the initial and the final steady states.
6. For small perturbations to columns with
both products of high purity, very large
time constants are found.
7. There are three contributions to Te
(caused by the holdup inside the column,
in the condenser, and in the reboiler).
As follows from the first comment, the
time constant given by Equation 4-33 is the
nonlinear time constant of the composition
58 Practical Distillation Control
dynamics of the column. The main advan-
tages of this method of evaluating the time
constant is that it depends on steady-state
data only and that it can be applied to
multicomponent separations. The disadvan-
tage of the method is that it requires the
availability of a nonlinear steady-state model
(not a great requirement in today's world)
and the calculation of compositions on all
trays.
A simple formula for a linearized time
constant limited to binary separations was
also developed by Skogestad and Morari
(l987b). The derivation of this formula was
based on three more assumptions, in addi-
tion to the two assumptions made for the
derivation of Tc:
Assumption III
All the trays (except the condenser and the
reboiler) have equal and constant holdup
ML
Assumption IV
The average composition inside the column
is given by
Assumption V
The steady-state gains may be estimated by
assuming that the separation factor S =
YD(1 - x
B
)/(1 - YD)X
B
is constant for any
given change.
Based on these assumptions, the lin-
earized value of Tc is given by
M[ M
LD
(1 - YD)YD
Tsc = I In S + I
s s
ML (1 - XB)X
B
+ B (4-35)
Is
where M[ is the total holdup of liquid in-
side the column,
M[ = NT * ML (4-36)
Is is the "impurity sum,"
and S is the separation factor,
YD(1- x
B
)
S = (4-38)
(1 - YD)X
B
T is expected to be in very good agreement
T for cases where Assumptions IV and
V hold (i.e., for high-purity columns with
high reflux). The advantages of the analyti-
cal formula for Tsc is that it depends only on
readily available steady-state data of the
initial operating state of the column, it does
not require the compositions on all stages,
and it does not require the availability of a
nonlinear steady-state model.
The main value of Tsc is that it can
provide a good understanding of the nonlin-
ear dynamics of the column and their de-
pendence on steady-state parameters. For
example, and taking into account the fact
that the contribution to Tc from the holdup
inside the column (MIlls In S) is dominant,
it is easy to see that the time constant is
determined mainly by Is (because In S does
not change significantly with the operating
conditions).
The usefulness and simplicity of the
method is illustrated by the following exam-
ple.
Example 41
Consider the binary column shown in Fig-
ure 4-3 with steady-state data given in Table
4-2. A disturbance was introduced by reduc-
ing the distillate flow rate by 0.01%. The
nonlinear column response (obtained by ap-
plying material balances on each equilib-
rium stage) was fitted by eye to a second-
order linear response 1/(1 + T
1
SX1 + T
2
S).
Both responses are shown in Figure 4-4.
The dominant time constant T 1 of the fitted
response is 4.3 h, whereas the time constant
Approximate and Simplified Models 59
TABLE 4-2 Steady State nata for Column
D,.., of Example 4-1
Parameter Symbol Value Units
Feed composition
ZF
0.4
Distillate composition
xD
0.99285
Bottoms composition
XB 0.00048
Feed quality q 1.0
Relative volatility a 2.0
yl Number of trays
NT
23
Feed tray location
NF
12
Distillate/feed flow D/F 0.4
"
.,
Reflux/feed flow L/F 1.6
Tray holdup/feed flow MTiF 0.01 hours
Cond. holdup/feed flow MD/F 0.1 hours
Reb. holdup/feed flow
MB/F 0.1 hours
.,-l


Tc calculated by Equation 433 was found to
be 3.94 h (i.e., an error of only 8%). The
:I value of Tsc calculated by Equation 435 for
1
the same column was found to be 4.15 h
(Le., an error of only 3.5%).
., sa
4-53 linear Models In the Frequency
FIGURE 43. Binary distillation column used in Ex- Domain
ample 4-1.
For small variations about an equilibrium
point, Equation 4-18 can be linearized. Us-
NORMALIZED LIQ MOL FRAC IN CONDENSER
1r----------------------------------------,
0.8
0.8 Line 1
0.4
0.2
2.8 8 7.8 10 12.8
TIME (HOURS)
FIGURE 4-4. Transient response of the condenser of the distillation column of
Example 4-1 to a step decrease of the distillate flow rate by 0.01%. Line 1: full
model; line 2: fitted modell/(l + 7'tsX1 + 7'2S).
60 Practical Distillation Control
ing average values for B*(s, t), Bv(s, t),
Re(s, I), and the resistance value Rv'(s, I),
and transforming to the frequency domain
gives
. jWTp,L .)
jWTp, V + . (I S, jW
1 + jWTe
where
= SP(s - 1,jw) - 2 SP(s,jw)
+ SP(s + 1,jw) (4-39)
(4-40)
(4-41)
linearized equation for the pressure drop
across the tray:
SV(s - 1,jw)
SP(s - 1,jw) - SP(s,jw)
(4-45)
For small deviations from equilibrium and
equal parameter values for the various trays,
the equation for the liquid flow variations,
Equation 4-21, is linearized to
jWTL SL*(s,jw)
= 8L*(s + 1,jw) - 8L*(s,jw)
+ (1 - KVL)jwB* SP(jw) (4-46)
Tp,V = BvRv
Tp,L = B*Rv
Te = ReB*
(4-42) Repeated application for a series of trays
yields
the average time constant for the condensa-
tion-vaporization phenomenon. This for-
mula represents a set of linear difference
equations, which can be solved analytically.
The result is
. (I: - J2) SP( m, jw)
8P(s,jw) = Jm_Jm
1 2
(4-43)
where 8P(m, jw) = pressure variation
above the top tray
(Pa)
SP
o
= pressure variation
below the bottom tray
(Pa)
J
1
, J
2
= roots of the charac-
teristic equation
Tp,.L ])J+1=0
1 + jWTe
(4-44)
The transfer functions for the vapor flow
responses can be found simply from the
8L*(1, jw)
8L*(M + 1,jw)
= . m + 8P(jw)(1 - K
VL
)
(1 + }WTL)
x ( )[1 - (1 + ] (4-47)
with Equation 4-41;
aL(l, jw)
SL*(m + 1,jw)
(1 +jwTL)m
[
. 8V(m,jw) ]
+KVL SV(O,jw) - . m
(1 + }WTd
+SP(jw)(1 - Kvd ( )
x [1 - __ 1_---::-]
(1 + jWT
L
) n
(4-48)
The tray parameter K VL has a strong im-
pact on the response of the bottom accumu-
lator level to the vapor flow. If it equals 1,
then the term KVL SV(O, jw) just cancels
the direct effect of the vapor flow variation,
so the bottom level has to wait for the term
with SV(m, jw), which is delayed by n times
the tray time constant Tv This makes bot-
tom level control by vapor flow as slow as by
reflux flow. In practice this can force col-
umn designers to increase the size of the
bottom holdup, and so make the column
more expensive.
Similar to the preceding submodels, one
can develop a linearized version of the con-
centration submodel with average values for
tray holdups, capacities, and pressures. Ap-
proximating the equilibrium relationship be-
tween vapor and liquid by a straight line,
Substitution into Equation 4-26 yields
dx(s, t)
[ML(s,t) + elMv(s,t)] dt
=L(s + 1,t)[x(s + 1,t) -x(s,/)]
+elV(s - 1,/)[X(S -1,/) -X(S,/)]
dP(s, I)
+ [eel - l)x(s, I) + eo]B(s, t) dt
(4-50)
By linearization and transforming Equation
4-26 to the frequency domain,
jW7'x 5x(s, jw)
= [5x(s + 1,jw) - 5x(s,jw)]
+E[5x(s - 1,jw) - 5x(s,jw)]
+jw[(e
l
- 1) 5x(s) + eo]B* 5P(s,jw)
+ [xes + 1) - xes)]
x[ 5L(S: 1,jw) _ 5V(s
where
V
E = el-
L
ML + elM
v
7' =
x L
( 4-51)
(4-52)
(4-53)
and use has been made of the static partial
Approximate and Simplified Models 61
mass balance of the tray,
L[x(s + 1) - x(s)]
+ V[y(s - 1) - y(s)] = 0 (4-54)
The pressure-dependent term in Equation
4-53 is rather awkward to handle ana-
lytically (Rademaker, Rijnsdorp, and
Maarleveld, 1975, state the formidable ana-
lytical expression for its effect on the con-
centration), but solving for the vapor and
the liquid flow is straightforward.
The analytical expressions for the re-
sponses of tray pressures, vapor flows, liquid
flow, liquid holdups, and liquid and vapor
concentrations are very convenient for the
design of regulatory control structures.
4-6 NONLINEAR MODELS
The number of equations constituting a
nonlinear model for a distillation column
depends primarily on three factors:
1. Simplifying assumptions
2. Number of components
3. Number of stages
In what follows, we will try to examine the
effect of these three factors on the complex-
ity and size of the models.
4-6-1 Simplifying Assumptions
The most rigorous model for describing the
transient behavior of stagewise distillation
processes should include a large number of
variables as state variables (i.e., tray liquid
compositions, liquid enthalpies, liquid
holdups, vapor enthalpies, vapor holdups,
etc.). By employing certain assumptions,
which we will call simplifying assumptions,
some of the state variables can be elimi-
nated, and the order of the model can be
significantly reduced. Such simplifying as-
sumptions are:
a. Negligible vapor holdup
b. Fast vapor enthalpy dynamics
c. Fast flow dynamics
d. Fast liquid enthalpy dynamics
62 Practical Distillation Control
Assumptions a and b are more or less stan-
dard in the literature of distillation dynam-
ics because interest is focused on composi-
tion dynamics and control. Assumption c
has been employed by many investigators
and is only reasonable if the measurement
lags in quality analysis are dominant (that is,
the flow dynamics are fast compared with
the composition dynamics). In case the flow
dynamics are significant, they can be treated
separately from the composition dynamics,
and then they can be superimposed on them
(see Section 4-4). Assumption d implies fast
enthalpy changes, turning the energy bal-
ances on each tray into a system of alge-
braic equations that must be satisfied at all
times (see Section 4-4).
4-6-2 Number of Components
After some simplifying assumptions have
been made, the resulting model can still be
of high order, due to the possible presence
of a very large number of components. Two
approaches have been suggested in the lit-
erature to overcome this problem. The first
consists of lumping some components to-
gether (pseudocomponents) and modelling
the column dynamics by applying energy
and material balances on these pseudocom-
ponents only. This approach has been criti-
cized as crude and arbitrary (Kehlen and
Ratzsch, 1987), mainly because there are no
good rules to guide the component-lumping
decisions.
The second approach is more recent
and is related to the development and ap-
plication of continuous thermodynamics
(Cotterman, Bender, and Prausnitz, 1985;
Cotterman and Prausnitz, 1985; Cotterman,
Chou, and Prausnitz, 1986) to chemical pro-
cess modelling. Application of these meth-
ods to steady-state multicomponent distilla-
tion calculations was performed by Kehlen
and Ratzsch (1987); it is expected that this
kind of work will extend the application of
the method to the modelling of distillation
dynamics. Recently, Shibata, Sandler, and
Behrens (1987) combined the method of
pseudocomponents with continuous ther-
modynamics for phase equilibrium calcula-
tions of "semicontinuous mixtures." Their
approach could be used for distillation
modelling applications.
In general, the methods of continuous
thermodynamics are suited for mixtures
containing very many components (i.e.,
petroleum fractions, shale oils, etc.), but not
for mixtures with a moderate number of
components (i.e., less than 10). For the lat-
ter cases, the lumping of some components
to pseudocomponents can be used to re-
duce the number of equations, but a lot
remains to be done (i.e., development of
lumping rules) before such an approach can
be successful.
4-6-3 Number of Stages - Orthogonal
Collocation
The third factor that affects the number of
equations that describe the dynamic model
of a distillation column is the number of
stages. There are several methods for reduc-
ing the order of the system in this context,
mainly by attempting to approximate the
dynamics of a number of stages by the dy-
namics of a fewer number of "pseudo-
stages." There are two approaches for such
an approximation. The first is the method of
compartmental models (Benalou, Seborg,
and Mellichamp, 1982). These investigators
considered a distillation column as a com-
partmental system, in which a number of
stages are lumped to form an equivalent
stage. The resulting model is low order,
nonlinear, and preserves both material bal-
ances and steady states for arbitrary changes
in the input variables. The advantages of
this model are that it retains the nonlinear
form and the gain of the original model, the
model parameters are related to the process
parameters, and changing the order of the
reduced model does not require any addi-
tional calculations. The disadvantages are
that there are no rules on how to divide a
given column into compartments and that
the simplified model can exhibit dynamic
behavior fundamentally different from that
of the original model (i.e., the compartmen-
tal model can exhibit inverse response
whereas the full model does not).
The second approach involves the reduc-
tion of the models using orthogonal colloca-
tion methods. There are several papers on
this topic (Wong and Luus, 1980; Cho and
Joseph, 1983a, 1983b, 1984; Stewart, Levien,
and Morari, 1985; Srivastava and Joseph,
1987a, 1987b). Their differences are mainly
in the selection of the polynomials, the ze-
ros of which are used as the grid points, and
in the selection of the polynomials used as
the weighting functions. One advantage that
the model of Stewart, Levien, and Morari
seems to have over the rest is the fact that it
uses the zeros of the Hahn polynomials,
which are more suited for discrete (staged)
systems, as the location of the collocation
points. Most of the other models use the
zeros of the Jacobi polynomials, which are
more suited for continuous descriptions.
Stewart, Levien, and Morari's model con-
verges to the full stagewise solution as the
number of collocation points approaches the
number of stages.
The collocation models are, in a sense,
compartmental models, and they are good
for simulation purposes because they re-
duce the order of the system significantly.
The advantages of these models (especially
of the one proposed by Stewart, Levien, and
Morari) is that they retain the nonlinear
form of the original model, they can be used
for muiticomponent distillation, they allow
free choice of thermodynamic subroutines,
they can be implemented without a full
order solution, and they can reduce the
computational time significantly. Their main
disadvantage is the nonretention of the orig-
inal model's gain in an exact manner. How-
ever, the gain predicted by the collocation
models is usually in very good agreement
with that of the full model. The basic
method for using orthogonal collocation to
obtain a reduced-order model for a distilla-
tion column is illustrated in succeeding text
Approximate and Simplified Models 63
------11+1
L v
II
2
L v
------ 0
FlGURE 4-5. A sequence of M process stages used
for the separation of a muiticomponent mixture.
and follows along the lines of the work of
Stewart, Levien, and Morari (1985).
Consider a sequence of M process stages
used for a multicomponent separation as
depicted in Figure 4-5. Assuming constant
molar overflow, the dynamic component
material balances on each tray s are given
by
dx(s, t)
ML dt =Lx(s+l,t)-Lx(s,t)
+ Vy(s - 1, t) - Vy(s, t)
(4-55)
where s = 1, ... , M. The compositions in
the module can be approximated by polyno-
mials, using n M interior grid points
S1' S2' ' sn plus two entry points sn+l =
M + 1 for the liquid and So = 0 for the
vapor compositions. (The reader should be
aware of the fact that in the original paper
of Stewart, Levien, and Morari the stages
are numbered from top to bottom, whereas
in this book the stages are numbered from
the bottom to the top. The equations pre-
sented here comply fully with the notation
and numbering used in this book and not in
the paper.)
64 Practical Distillation Control
TABLE 4-3 Collocation Points for a Module Consisting of M Stages
Number of
Collocation
Points n
1
2
3
Points Sj
M+ 1
2
M+1
-2 - V ----u--1-2 -
M + 1 M + 1 V 3M
2
- 7
-2-'-2- 20
4
M + 1 _ / 3M
2
- 13 _ /6M
4
- 45M2 + 164
-2- V 28 Y 980
The corresponding equations of the col-
location model for the module of Figure 4-5
are
d
ML d/(Sj' t).= Li(Sj + 1, t) - Li(sj' t)
+ Vj(Sj - 1, t) - Vj(Sj' t)
(4-56)
for j = 1, ... , n, where the tilde denotes an
approximate value. Therefore, the number
of equations describing the component ma-
terial balances for the module is reduced
from M (number of stages in the module)
to n (number of collocation points in the
module). The locations of the collocation
points S1"'" sn are the zeros of the Hahn
polynomials (Hahn, 1949) with (a, b) =
(0,0). The zeros of the Hahn polynomials
for n 4 are given in Table 4-3.
Srivastava and Joseph (1985) have shown
that using Hahn polynomials with a = 0
and b > 0 leads to better approximations
where the error is more evenly distributed
throughout the column.
The compositions in each stage of the
module are given by
n+l
xes, t) = E wh(s)x(Sj' t)
j=1
n
ji(s, t) = E Wjv(S)ji(Sj' t)
j=O
(4-57)
where the W functions are the Lagrange
polynomials
j = 1, ... , n + 1
(4-58)
j = 0, ... , n
Note again that So = 0 and sn+1 = M + 1.
The method is applicable for the solution
of both steady-state and dynamic models of
multicomponent distillation columns with
Di.tillate product
Side product
Peed atream 1
Peed atream 2
Bouam. product
nGURE 4-6. A complex distillation column with
two feed streams and three product streams.
multiple feeds and/or product streams, and
it can accommodate columns with tray-
dependent temperatures, liquid and vapor
flows, mixtures with nonideal thermodynam-
ics, trays with efficiencies less than 1, and so
on.
One issue related to the implementation
of the method is associated with the way in
which a given column is divided into mod-
ules. Consider the column shown in Figure
4-6; it consists of two feed streams and
three product streams. A modular decom-
position of this column is shown in Figure
4-7a (decomposition scheme A). An alterna-
Approximate and Simplified Models 65
tive decomposition (scheme B) is shown in
Figure 4-7b.
There is a trade-off between the two
decomposition schemes. Scheme A leads to
a smaller number of equations (Le., equa-
tions for modules plus equations for con-
denser and reboiler), but the Jacobian of
the resulting model is a full matrix (Le., very
few nonzero elements). In contrast, scheme
B requires more equations (Le., equations
for modules plus equations for condenser
and reboiler plus equations for feed-prod-
uct trays), but the corresponding Jacobian
has a block diagonal form that facilitates
the solution of the system (Papadourakis,
1985). The superiority of scheme B, which
corresponds to treating all the feed-product
D trays as collocation points, was also recog-
nized (for different reasons) by Pinto and
Biscaia (1988). In addition, the modular de-
composition of a column according to
scheme B is easier to implement. The appli-
cation of the preceding method is illustrated
s in the following example.
D
S
FI
Modale 2
FI
Module 3
F2
F2
Modul.4
B B
ScIIemeA ScbImeB
<eI) (b)
FIGURE 4-7. Modular decomposition schemes for
the column shown in Figure 4-6. (a) Decomposition
scheme A in which the feed and intermediate product
trays are lumped inside a module. (b) Decomposition
scheme B in which all feed and product trays are used
as collocation points.
Example 4-2
Consider again the binary column described
in Example 4-1. Applying the decomposi-
tion scheme B, the stripping section trays
(excluding the feed tray) constitute one
module, the rectifying section trays consti-
tute a second module, and the feed tray, the
condenser, and the reboiler are used as
collocation points by themselves. Applying
this decomposition scheme and the colloca-
tion method of Stewart, Levien, and Morari
(1985), a reduced-order model of the col-
umn was developed.
The first equation is' the mass balance
around the reboiler:
d
MLBd/(Sl,t) = (L + F)i(Sl + 1,t)
-Bi(Sl' t) - W(Sl' t)
(4-59)
where S 1 is the collocation point corre-
66 Practical Distillation Control
sponding to the reboiler and from Table 4-2
for n = 1 and M = 1, we get S 1 = 1.
Assuming that the dynamics of the strip-
ping section can be described by using only
two collocation points, the corresponding
mass balance equations become
d
M
Ld
/(S2' t)
= (L + F)i(S2 + 1, t) - (L + F)i{S2' t)
+VY(S2 -1,t) - Vy(S2,t) (4-60)
= (L + F)i(S3 + 1, t) - (L + F)i(S3' t)
+Vy(s3 -1,t) - Vy(S3,t) (4-61)
The locations of the collocation points S2
and S3 are found from Table 4-2 for n = 2
(number of collocation points in the section)
and for M = 11 (number of trays in the
section). It turns out that S2 = 2.84 and
S3 = 9.16. However, because there is an ad-
ditional stage (the reboiled below the strip-
ping section, then S2 = 3.84 and S3 = 10.16.
The mass balance for the feed tray is
given by
d
M
Ld
/{S4,t)
= Li(S4 + 1, t) - (L + F)i(S4' t) + FXF
(4-62)
where 54 is the collocation point for the
feed tray. Again, from Table 4-2 we get (for
n = 1 and M = 1) that S4 = 1. However,
because there are 13 stages below the feed
tray (reboiler + 12 trays of the stripping
section), we set S4 = 14.
Assuming that the dynamics of the recti-
fying section can also be described by using
only two collocation points, the correspond-
ing mass balance equations become
d
M
Ld
/(S5,t)
= Li(S5 + 1, t) - Li(S5' t)
+ V)i(S5 - 1, t) - Vy(S5' t)
d
M
Ld
/(S6,t)
= Li(S6 + 1, t) - Li(S6' t)
+ VY(S6 - 1, t) - V)i{S6' t)
(4-63)
( 4-64)
The location of the collocation points S5
and S6 is found from Table 4-2 for n = 2
(number of collocation points in the section)
and for M = 11 (number of trays in the
section). It turns out that S5 = 2.84 and
S6 = 9.16. Again, because there are 14 trays
below the rectifying section, we set S5 =
16.84 and S6 = 23.16.
Finally, the mass balance for the total
condenser takes the form
d
M
LDd
/(S7' t) = -Li(S7' t) - Di(S7' t)
+ V)i{ S7 - 1, t) (4-65)
where S7 = 25.
Equations 4-59 to 4-65 constitute a sys-
tem where the unknowns are i(sj, t), i =
1,7. Note that a full scale model for the
column (using the same assumptions) would
consist of 25 equations (Le., number of
equations = number of stages).
In Equations 4-59 to 4-65 the corre-
sponding Y(Si' t) can be found from the
equilibrium relationship
_ ai(sj,t)
Y(Sj, t) = 1 + (a _ 1)i(sj, t) (4-66)
The remaining variables in the preceding
equations are functions of the liquid and
vapor mole fractions of the collocation
points as described by Equations 4-57 and
4-58. For example, the variable i(S3 - 1, t)
= i(9.16, t) is given by
4
i(9.16, t) = L w
h
{9.16)i(Sj, t) (4-67)
j=2
Approximate and Simplified Models 67
where the w functions are the Lagrange the stage (collocation point) immediately
polynomials below the module.
j = 2, ... ,4
(4-68)
Note that (9.16, t) is the liquid mole frac-
tion of a fictitious tray that is inside the
stripping section (module). Therefore, and
according to Equations 4-57 and 4-58, it will
depend on the liquid mole fractions of the
collocation points of this module only, plus
on the liquid mole fraction of the stage
immediately above the module. Using the
decomposition scheme shown on Figure 4-
7b, a tray immediately above or below a
module is a feed or product stage and,
therefore, it constitutes the location of a
lone collocation point.
Note also that the vapor mole fraction of
a fictitious (or real) tray that belongs to a
certain module depends only on the vapor
mole fractions of collocation points of the
module and on the vapor mole fraction of
By setting the derivatives of the coll-
ocation dynamic model equal to zero, an
approximate steady-state model for the col-
umn was obtained. The steady-state compo-
sition profile of the column was calculated
by first solving the full model, and also by
solving the reduced-order model. The re-
sults are shown in Figure 4-8. In the same
figure, the results are shown from approxi-
mate models developed by using 3 and 4
collocation points in each module (resulting
in 9 and 11 collocation points for the entire
column). As can be seen in the figure, the
agreement between the full model and the
reduced-order model is excellent when 11
collocation points are used in the column,
very good when 9 points are used, and poor
when only 7 points are used for the entire
column. It should be noted here that the
solution that exhibits very good agreement
with the full solution (i.e., when 9 colloca-
tion points were used for the entire column)
results from the solution of 9 equations,
whereas the full model required the solu-
tion of 25 equations.
LIQUID MOLE FRACTION OF LIGHT COMPONENT
1.2.-----------------------------------------,
0.8
0.6
0.4
.0.2



STAGE NUMBER (O-REBOILER, 24-CONDENSER)
FIGURE 4-8. Steady-state composition profile of the column of Example 4-2 . full
model; + reduced model with 7 collocation points; * reduced model with 9 colloca-
tion points; 0 reduced model with 11 collocation points.
68 Practical Distillation Control
LIQUID MOLE FRACTION IN CONDENSER

0.888
0.888 Line 1
Line 2
0.88 t-U-.L.L.L..L.L.L."+I-......... ..r...u....I..I..j...L.U...L.UU-I..I"+L'U../..,u..L..L.4.L.L..1..L.L..I...&..I.J.+U..&..I.J...u.u
o 2.5 5 T.5 10 12.5
TIME (HOURS)
FIGURE 4-9. Transient response of the condenser of the distillation column of
Example 4-2 to a step decrease of the distillate flow rate by 1%. Une 1: full model;
line 2: reduced model with 9 collocation points; line 3: reduced model with 11
collocation points.
L ________________ -,
0.012.-
0.01
Line :3
0.004
0.002
.........
o 2.5 5 T.5 10 12.5
TIME (HOURS)
FIGURE 4-10. Transient response of the reboHer of the distillation column of
Example 4-2 to a step decrease of the distillate flow rate by 1%. Une 1: full model;
line 2: reduced model with 9 collocation points; line 3: reduced model with 11
collocation points.
The method was also applied to obtain a
reduced-order dynamic model for the col-
umn. The predicted column response for a
1 % reduction in the distillate flow rate is
shown in Figures 4-9 and 4-10. Again, the
agreement between the reduced-order mod-
els and the full model ranges from excellent
(model with 11 collocation points in the
column) to very good (model with 9 colloca-
tion points in the column).
The degree of order reduction achievable
for a given column (using orthogonal collo-
cation) was dealt with in a paper by Srivas-
tava and Joseph (1985). They developed an
index called order-reduction parameter that
can be used to predict the extent of order
reduction achievable and to provide guide-
lines for choosing an appropriately sized
reduced-order model.
In general, reduced-order models ob-
tained from the application of the previ-
ously mentioned orthogonal collocation
methods can be used for both steady-state
and dynamic simulation purposes with ex-
cellent results. Use of these approximate
models can lead to reductions in computa-
tional times proportional to (N
c
IN)2 where
Nc is the total number of collocation points
in the column and N is the total number of
equilibrium stages in the column (Pinto and
Biscaia, 1988).
References
Arkun, Y. (1987). Dynamic block relative gain
array and its connection with the perfor-
mance and stability of decentralized control
structures. Int. J. Control 46, 1187-1193.
Bamberger, W. and Isermann, R (1978). Adap-
tive steady-state optimization of slow dynamic
processes. Automatica 14,223-230.
Benalou, A, Seborg, D. E., and Mellichamp,
D. A (1982). Low order, physically lumped
dynamic models for distillation column con-
trol. Presented at the AIChE Annual Meet-
ing, November 1982, Los Angeles, CA.
Betlem, D. H. L. (1991). Personal communica-
tion.
Bhat, P. V. and Williams, T. J. (1969). Approxi-
mate dynamic models of distillation opera-
tions and their relation to overall column
Approximate and Simplified Models 69
control methods. Int. Chem. Eng. Symposium
Series 32,22-32.
Bonvin, D. and Mellichamp, D. A. (1982a). A
generalized structural dominance method for
the analysis of large-scale systems. Int. J.
Control 35, 807-827.
Bonvin, D. and Mellichamp, D. A (1982b). A
unified derivation and critical review of modal
approaches to model reduction. Int. J. Con-
trol 35, 829-848.
Celebi, C. and Chimowitz, E. H. (1985). Analytic
reduced order dynamic models for large equi-
librium staged cascades. AIChE J. 31,
2039-2051.
Chen, C. H. (1969). M.Sc. thesis, Polytechnic
Institute of New York.
Cho, Y. S. and Joseph, B. (1983a). Reduced-order
steady-state and dynamic models for separa-
tion processes Part I. Development of the
model reduction procedure. AIChE J. 29,
261-269.
Cho, Y. S. and Joseph, B. (1983b). Reduced-
order steady-state and dynamic models for
separation processes Part II. Application to
nonlinear multicomponent systems. AIChE J.
29,270-276.
Cho, Y. S. and Joseph, B. (1984). Reduced-order
models for separation columns-III. Applica-
tion to columns with multiple feeds and
sidestreams. Comput. Chern. Eng. 8, 81-90.
Cotterman, R L. and Prausnitz, J. M. (1985).
Flash calculations for continuous or semicon-
tinuous mixtures using an equation of state.
Ind. Eng. Chem. Proc. Des. Dev. 24,434-443.
Cotterman, R L., Bender, R, and Prausnitz,
J. M. (1985). Phase equilibria for mixtures
containing very many components. Develop-
ment and application of continuous thermo-
dynamics for chemical process design. Ind.
Eng. Chem. Proc. Des. Dev. 24, 194-203.
Cotterman, R L., Chou, G. F., and Prausnitz,
J. M. (1986). Comments on "Flash calcula-
tions for continuous or semicontinuous mix-
tures using an equation of state." Ind. Eng.
Chem. Proc. Des. Dev. 25, 840-841.
Genesio, R M. and Milanese, M. (1976). A note
on the derivation and use of reduced-order
models. IEEE Trans. Automat. Control AC-
21, 118-122.
Georgakis, C. and Stoever, M. A. (1982). Time
domain order reduction of tridiagonal dynam-
ics of staged processes-I. Uniform lumping.
Chem. Eng. Sci. 37, 687-697.
70 Practical Distillation Control
Hahn, W. (1949). Uber Orthogonalpolynome die
q-Differenzengleichungen genugen. Math.
Nachrichten 2,4-34.
Hawkins, D. E., Tolfo, F., and Chauvin, L. (1987).
Group methods for advanced column control.
In Handbook of Advanced Process Control
Systems and Instrumentation, L. Kane, ed.
Houston: Gulf Publishing Co.
Kehlen, H. and Ratzsch, M. T. (1987). Complex
multicomponent distillation calculations by
continuous thermodynamics. Chem. Eng. Sci.
42,221-232.
Kim, C. and Friedly, C. F. (1974). Dynamic mod-
eling of large staged systems. Ind. Eng. Chem.
Proc. Des. Dev. 13, 177-181.
Levy, R. E., Foss, A S., and Grens, E. A, II
(1969). Response modes of a binary distilla-
tion column. Ind. Eng. Chem. Fundam. 8,
765-776.
Moczek, J. S., OUo, R. E., and Williams, T. J.
(1963). Proceedings of the Second IFAC
Congress, Basle. New York: Pergamon.
Moczek, J. S., OUo, R. E., and Williams, T. J.
(1965). Approximation models for the dy-
namic response of large distillation columns.
Chem. Eng. Prog. Symp. Ser. 61, 136-146.
Papadourakis, A (1985). Stability and Dynamic
Performance of Plants with Recycle. Ph.D. Dis-
sertation, University of Massachusetts,
Amherst, MA.
Papadourakis, A., Doherty, M. F., and Douglas,
J. M. (1989). Approximate dynamic models
for chemical process systems. Ind. Eng. Chem.
Res. 28, 546-552.
Perkins, J. D. (1990). Interactions between pro-
cess design and process control. In Dynamics
and Control of Chemical Reactors, Distillation
Columns and Batch Processes. IFAC Sympo-
sium Series 1990, No.7, J. E. Rijnsdorp et aI.,
eds. Oxford: Pergamon, 195-203.
Pinto, J. D. and Biscaia, E. c., Jr. (1988). Order
reduction strategies for models of staged sep-
aration systems. Comput. Chem. Eng. 12,
812-831.
Rademaker, 0., Rijnsdorp, J. E., and Maar-
leveld, A (1975). Dynamics and Control of
Continuous Distillation Units. Amsterdam: El-
sevier Science Publishers.
Rijnsdorp, J. E. (1991). Integrated Process Con-
trol and Automation. Amsterdam: Elsevier
Science Publishers.
Robinson, C. S. and Gilliland, E. R. (1950).
Elements of Fractional Distillation. New York:
McGraw-Hill.
Shibata, S. K., Sandler, S. I., and Behrens, R. A
(1987). Phase equilibrium calculations for
continuous and semicontinuous mixtures.
Chem. Eng. Sci. 42, 1977-1988.
Shinskey, F. G. (1967). Process Control Systems.
New York: McGraw-Hill.
Skogestad, S. and Morari, M. (1987a). Effect of
disturbance directions on closed loop perfor-
mance. Ind. Eng. Chem. Res. 26, 2029-2035.
Skogestad, S. and Morari, M. (1987b). The domi-
nant time constant for distillation columns.
Comput. Chem. Eng. 11,607-617.
Skogestad, S. and Morari, M. (1988). Under-
standing the dynamic behavior of distillation
columns. Ind. Eng. Chem. Res. 27,1848-1862.
Srivastava, R. K. and Joseph, B. (1985). Re-
duced-order models for separation
columns-V. Selection of collocation points.
Comput. Chem. Eng. 9, 601-613.
Srivastava, W. E. and Joseph, B. (1987a). Re-
duced-order models for staged separation
columns-IV. Treatment of columns with
mUltiple feeds and sidestreams via spline fit-
ting. Comput. Chem. Eng. 11, 159-164.
Srivastava, W. E. and Joseph, B. (1987b). Re-
duced-order models for staged separation
columns-VI. Columns with steep and fiat
composition profiles. Comput. Chem. Eng. 11,
165-176.
Stewart, W. E., Levien, K. L., and Morari, M.
(1985). Simulation of fractionation byorthog-
onal collocation. Chem. Eng. Sci. 40, 409-421.
Stoever, M. A and Georgakis, C. (1982). Time
domain order reduction of tridiagonal dynam-
ics of staged processes-II. Nonuniform
lumping. Chem. Eng. Sci. 37, 699-705.
Toijala, K. V. (1968). An approximate model for
the dynamic behavior of distillation columns.
Part I. General theory. Acta Acad. Aboensis
(Math. Phys.) 28(8), 1-18.
Toijala, K. V. (1969). An approximate model for
the dynamic behavior of distillation columns.
Part II. Composition responses to changes in
feed composition for binary mixtures. Acta
Acad. Aboensis (Math. Phys.) 29(10), 1-34.
Toijala, K. V. (1971a). An approximate model
for the dynamic behavior of distillation
columns. Part IV. Comparison of theory with
experimental results for changes in feed com-
position. Acta Acad. Aboensis (Math. Phys.)
31(3), 1-25.
Toijala, K. V. (1971b). An approximate model
for the dynamic behavior of distillation
columns. Part V. Comparison of theory with
experimental results for changes in reflux flow
rate. ActaAcad. Aboensis (Math. Phys.) 31(5),
1-19.
Toijala, K. V. (1978). Simple models for distilla-
tion dynamics. Paper presented at the 86th
National AIChE Meeting, April 1-5, 1978,
Houston, Texas.
Toijala, K. V. and Fagervik, K. (1972). An ap-
proximate model for the dynamic behavior of
distillation columns. Part VII. Comparison of
model with experimental results for con-
trolled columns by the use of digital simula-
tion. Acta Acad. Aboensis (Math. Phys.) 32(1),
1-17.
Toijala, K. V. and Gustafsson, S. (1971). Process
dynamics of multi-component distillation.
Kemian Teollisuus 28, 113-120.
Toijala, K. V. and Gustafsson, S. (1972a). Pro-
cess dynamics of multi-component distillation.
Comparison of theoretical model with pub-
lished experimental results. Kemian Teollisuus
29,95-103.
Toijala, K. V. and Gustafsson, S. (1972b). On the
general characteristics of component distilla-
tion dynamics. Kemian Teollisuus 29, 173-184.
Approximate and Simplified Models 71
Toijala, K. V. and Jonasson, D. (1970). An ap-
proximate model for the dynamic behavior of
distillation columns. Part III. Composition re-
sponse to changes in reflux flow rate, vapour
flow rate and feed enthalpy for binary mix-
tures. Acta Acad. Aboensis (Math. Phys.)
30(15), 1-21.
Toijala, K. V. and Jonasson, D. (1971). An ap-
proximate model for the dynamic behavior of
distillation columns. Part VI. Composition re-
sponse for changes in feed flow rate for bi-
nary mixtures. Acta Acad. Aboensis (Math.
Phys.) 31(11), 1-27.
Wahl, E. F. and Harriot, P. (1970). Understand-
ing and prediction of the dynamic behavior of
distillation columns. Ind. Eng. Chem. Proc.
Des. Dev. 9,396.
Weigand, W. A., Jhawar, A. K., and Williams,
T. 1. (1972). Calculation method for the re-
sponse time to step inputs for approximate
dynamic models of distillation columns.
AIChE 1. 18, 1243-1252.
Wong, K. T. and Luus, R. (1980). Model reduc-
tion of high-order multistage systems by the
method of orthogonal collocation. Can. 1.
Chem. Eng. 58, 382.
5
Object-Oriented Simulation
B. D. Tyreus
E. I. Du Pont de Nemours & Co.
5-1 INTRODUCTION
Steady-state process simulation is used rou-
tinely in the evaluation, selection, and de-
sign of new processes (Biegler, 1989). In the
past, all the pertinent design factors could
be addressed with steady-state programs.
Today, however, there are new factors to be
considered, and some of these have more to
do with the operation and control of the
plant than with steady-state aspects. For
example, product quality demands, increas-
ing safety concerns, stringent environmental
requirements, process flexibility, minimiza-
tion of capital expenditure through process
simplification and integration, and elimina-
tion of overdesign are all important factors
that tend to influence process operations.
However, the operational aspect of process
design is difficult or impossible to glean
from steady-state simulations, and we must,
therefore, consider dynamic simulations for
answers. How then are dynamic simulations
used in process design? Traditionally, they
are used in control system evaluation and
tuning but also to find the appropriate con-
troller structure for a particular process. To
date, such simulations have mostly involved
one or two unit operations with little em-
phasis on plantwide control.
72
The operational factors cited in the fore-
going text suggest a broader use for dy-
namic simulations than for just control stud-
ies of unit operations. For example, with
current integrated plant designs, the great-
est benefit of a dynamic analysis might be in
solving the plantwide control problem. This
idea is further reinforced in Chapter 6. In
that context, the management of invento-
ries, recycle paths, product transitions,
startup, and shutdown can be studied as
they impact safety, environment, and prod-
uct quality. Furthermore, when a plantwide
dynamic simulator has been developed for
process design, it can later be used for oper-
ator training and process improvement stud-
ies.
Are we ready then to meet an increased
demand of dynamic simulations? Unfortu-
nately not, due to the prevailing paradigm
used when considering such simulations.
Under this paradigm, dynamic simulations
are viewed as control system evaluation
tools, and because process control is a spe-
cialized field mastered by control engineers,
we let the specialists create and use the
simulations according to their needs. Some-
times this means that the model is a set of
transfer functions used in a linear analysis;
sometimes a more detailed nonlinear dy-
namic model is written from first principles
and implemented in FORTRAN or BASIC
(see Chapter 3 for example). The models
are almost always tailored to the specific
problem at hand; seldom extensible to other
problems and almost never usable by oth-
ers.
The tools we know today for creating
dynamic simulations are consistent with the
prevailing paradigm. Whether they are inte-
grators like LEANS, libraries of FOR-
TRAN subroutines like DYFLO (Franks,
1972), or equation-based numerical solvers
like SPEEDUP (Perkins, 1986), these tools
are aimed at the specialist who has consid-
erable experience in using the tool, who
knows how to model various processes in
terms of their fundamental equations, and
who is willing to spend considerable time in
entering code and data into input files,
which are compiled, edited, and debugged
before they yield results in terms of time
plots of pre specified variables over fixed
time periods.
I feel that in order to meet the broader
need for dynamic simulations, a paradigm
shift is required. Under a different para-
digm, dynamic simulations would go beyond
control system studies, and all process engi-
neers should be able to create and use such
simulations. I also feel that today's tools
and techniques for generating dynamic sim-
ulations will be inadequate under a differ-
ent paradigm. Instead, I suggest that new
software approaches, such as object-ori-
ented programming (OOP), will replace
current tools in order to support a new
paradigm (Stephanopoulos, Henning, and
Leone, 1990). I am also aware of a couple of
simulation software companies that have
reached the same conclusion.
The purpose of this chapter is thus to
familiarize you with an emerging software
technology for dynamic simulations. I will
describe the concepts of objects and mes-
sages and contrast them to their more fa-
miliar counterparts: data structures and
subroutines. The important features of in-
heritance and polymorphism are also cov-
Object-Oriented Simulation 73
ered. All this is put in the context of
creating dynamic simulation models for dis-
tillation systems.
This chapter is, however, not detailed
enough to show you how to write your own
object-oriented dynamic simulator, nor does
it give enough insight into how I have writ-
ten mine. Instead I merely hope to remove
some of the mystique around object-ori-
ented programming and to spark enough
interest that you might consider experiment-
ing with an object-oriented compiler next
time you write a simulation model. In case
you do not like writing your own simulations
from scratch but rely on vendor tools, this
chapter will hopefully interest you a little
more in emerging simulation packages built
on object-oriented principles.
52 IDEAL DYNAMIC SIMULATOR
Before delving into the principles of OOP,
we will take a closer look at a vision of a
suitable tool for creating dynamic simula-
tions.
In order to shift to a different paradigm,
one has to ask the question: "What is it that
we cannot do now that if it could be done
would have a significant impact on our busi-
ness?" When I asked this question about
dynamic models and the simulation tools we
use, I found the following to be true:
Most process engineers cannot create or
use a dynamic simulation because the tools
and techniques are designed for special-
ists.
Dynamic simulators are not user-friendly,
preventing the casual user from consider-
ing them.
Dynamic simulations are expensive to de-
velop because it is generally difficult to
reuse previously written code in new ap-
plications, thereby forcing the developer
to rewrite substantial sections of the new
model.
Dynamic simulations are not readily ac-
cessible because the development envi-
74 Practical Distillation Control
ronments and final models require rela-
tively large, expensive computers.
Many simulators are not configurable and
require extensive programming and al-
most always compilation before the model
can be run.
Dynamic simulations are often slow.
The user cannot interact with the simula-
tion during a run.
Even for the sophisticated user, most dy-
namic simulation packages are limited be-
cause they are not portable, extensible, or
maintainable.
If instead we had simulation tools that were
configurable, user-friendly, flexible, port-
able, readily accessible, fast, and inexpen-
sive, we could dramatically change the way
we consider dynamics in process design and
control. Such an ideal dynamic simulator
would allow process engineers and control
engineers to routinely develop rigorous,
nonlinear, dynamic simulations for most
processes. It would allow us to study the
impact of process design and controller
structure on the operation of not just dif-
ferent unit operations, but entire processing
units and even plants (see also Chapter 6).
The ideal dynamic simulator would allow
process engineers to explore and under-
stand existing processes so the processes
could be simplified and improved. It would
allow us to quickly validate our control sys-
tems and to train operators with realistic,
validated models. In short, the ideal dy-
namic simulator would change the way we
do business; it would help create a paradigm
shift.
Once the vision of an ideal dynamic sim-
ulator has been created, we need to know
how such a tool could be developed.
Object-oriented programming provides an
excellent vehicle to take us to our goal.
5-3 OBJECT-ORIENTED
PROGRAMMING
To help explain the concepts of object-ori-
ented programming, I will contrast this
new software technique with traditional
programming approaches supported by pro-
cedural languages, such as BASIC, FOR-
TRAN, Pascal, and C. In procedural pro-
gramming, the focal point is data. Data are
transformed to results by letting procedures
or subroutines operate on them. For exam-
ple, in a simulation package, such as
DYFLO (Franks, 1972), there are data
structures defined that specify how compo-
nent compositions, temperature, and pres-
sure are stored, how parameters enter the
simulation, and how the results of chemical
reactions are made available when they are
needed. DYFW then consists of a library
of subroutines that either take the data
structures as parameters or use the data
directly from COMMON areas to produce
results. Similarly, when writing a Pascal pro-
gram, the first thing that has to be done is
to define the data types and variables that
will be used by the procedures and func-
tions that get invoked in a predefined, pro-
cedural fashion.
If in procedural programming the key
words are data and subroutines, then in
object-oriented programming the key words
are objects and messages (Pascoe, 1986).
Although there are some similarities be-
tween data and objects on the one hand and
subroutines and messages on the other,
there are several significant differences. The
most significant difference is between data
and objects. Data structures are passive
storage places for information that are
transformed by subroutine calls in the pro-
gram. Objects, on the other hand, are active
entities that exhibit behavior when they re-
ceive a message. Objects are data structures
and subroutines packaged into one unit (en-
capsulations) where the data are globally
known to all the subroutines that belong to
that object. A message is then the mecha-
nism to trigger the appropriate subroutine
within the object. Because objects send
messages to each other and because the
particular object to receive a particular mes-
sage can change during execution of the
program, there is no counterpart to the
sequence of subroutine calls found in proce-
dural programs. Here lies part of the diffi-
culty in transitioning from the procedural
paradigm to an object-oriented approach,
but it also contains the real power of
object -oriented programming.
5-3-1 Classes and Objects
Sometimes terminology can be an obstacle
when describing a new concept. Object-ori-
ented programming has its share of special
nomenclature such as classes, encapsula-
tions, data abstraction, objects, instantia-
tions, inheritance, polymorphism, and dy-
namic binding. If these words were mere
synonyms for well known concepts, I could
simply provide a glossary and be done. Un-
fortunately, this is not the case. Each of
these words stands for a concept that is
unique to object-oriented programming, and
for that reason, it is probably a good idea
that they do not sound like things we are
familiar with. I will describe each concept as
I introduce it and give some examples, but
for now, we will focus on classes, instances,
and objects. A class is the template for an
object. A class defines what data the object
contains, what behavior the resulting ob-
jects can exhibit (what subroutines are con-
tained in the object), and what messages the
object responds to. The declaration and
definition of classes look and feel like pro-
gramming in just about any language; there
are data declarations and subroutine defi-
nitions. Given the declaration of a class (the
template), one or more objects can be in-
stantiated (cloned) from that class anywhere
and anytime in the program or during pro-
gram execution. This does not have a simple
analog in conventional programming. The
working program then revolves around ac-
tive objects that send and receive messages
that cause desirable actions and produce
results.
532 Modelling a Column Tray
In order to illustrate some of the important
concepts in object-oriented programming,
we will model the dynamics of the fluid on a
tray in a distillation column. We will com-
Object-Oriented Simulation 75
FIGURE 51. nth tray of multicomponent column.
pare an object-oriented solution to the more
familiar procedural approach. Figure 5-1
shows a section of a distillation column with
a single tray. The liquid hydraulics and
multicomponent composition dynamics can
be described by the following equations
(Luyben, 1990; see also Chapter 3 for more
details on modelling a distillation column):
(5-1)
+ v,,-IYn-l,j - VnYn,j (5-2)
(5-3)
(5-4)
Procedural Approach
We will first study a procedural approach to
implementing the tray model. The equa-
tions are conveniently coded just as they
appear in a procedural language like FOR-
TRAN or Pascal. In writing the code, we
would, however, be responsible for defining
data areas or storage locations for the xs,
76 Practical Distillation Control
c****************** COLUMN MODULE***********************************
SUBROUTINE COLUMN(NT,NF,DT)
S
10
1S
COMMON NC
COMMON DM,DXM,XM,X,Y,T,P
DIMENSION DM(SO),DXM(SO,S),XM(SO,S),X(SO,S),Y(SO,S),
T(SO),XX(S),YY(S)
DO 1S N=1,NT
DO S J = 1,NC
XX(J) = X(N,J)
CALL BUBPT(T(N),XX,YY,P)
DO 10 J=1,NC
Y(N,J) = YY(J)
CONTINUE
C CODE TO CALCULATE VAPOR AND LIQUID TRAFFIC HERE
DO 20 N=2,NF-1
20 DM(N) = UN + 1) + V(N- 1)- UN)- yeN)
C DM FOR FEED TRAY AND REST OF COLUMN HERE
DO 100 J=1,NC
DO SO N=2,NF-1
SO DXM(N,J) = X(N + 1,J)*UN + 1) + Y(N-1,J)*V(N-1)- X(N,J)*UN)
-Y(N,J)*V(N)
c DXM FOR FEED TRAY AND REST OF COLUMN HERE
100 CONTINUE
c INTEGRATE A LA EULER
200
300
400
DO 200 N = 1,NT
M(N) = M(N) + DM(N)*DT
DO 400 J = 1,NC
DO 300 1,NT
XM(N,J) = XM(N,J) + DXM(N,J)*DT
X(N,J) = XM(N,J) I M(N,J)
IF (X(N,J) .GT. 1.0)X(N,J) = 1.0
IF (X(N,J) .LT. O.O)X(N,J) =0.0
CONTI NUE
CONTINUE
RETURN
END
c************************* VLE MODULE *************************
SUBROUTINE BUBPT(T,X,Y,P)
COMMON NC
COMMON ALPHA
DIMENSION ALPHA(S),X(S),Y(S)
SUM=O.O
D01K=1,NC
SUM = SUM + ALPHA(K)*X(K)
D02J=1,NC
Y(d) = ALPHA(d)*X(d) I SUM
RETURN
END
FIGURE S-2. Procedural implementation of distillation column dynamics.
ys, alphas, Ls, and Ms. Typically, these
variables would be stored in single or multi-
dimensional arrays.
As convenient as it might be to directly
code the equations into a program, we would
introduce a serious limitation in our code by
mixing the tray equations with the details of
the equilibrium calculations. If we later want
to change the way we calculate equilibrium,
we would have to go right to the heart of
the program and make changes. Such edit-
ing is always dangerous because there is a
risk of breaking or invalidating a perfectly
fine program. We could, of course, make a
copy of the program and keep the original
as version} and edit the copy and call it
version2. Such proliferation of versions has
its own problems because it quickly leads to
an unmanageable set of programs that not
even the owner can keep track of. A better
approach is to separate the programs into
different modules by use of subroutines.
Equations 5-1 through 5-3, which describe
the composition dynamics on the tray, are
not going to change much and should be
kept in their own module. The equilibrium
calculation (Equation 5-4), which is likely to
change more, should be in a different place.
In order to communicate between the two
modules, we use a subroutine call such as:
CALL BUBPT (T,X,Y,P) in place of Equa-
tion 5-4. Strictly speaking, we only need X
and Y as arguments when we use relative
volatilities, but it might be a good idea to
provide temperature (T) and pressure (P) in
case we later plan to perform bubble point
calculations based on vapor pressures.
The procedural implementation of the
tray model now seems to be in good shape.
We have one module that describes the tray
dynamics and another module that de-
scribes the equilibrium (see Figure 5-2). We
have eliminated the need to make changes
in the tray section code and we have re-
moved the multiple version syndrome. Or
have we? There are still some loose ends. In
return for removing Equation 5-4 from the
tray section code, we had to permanently
introduce the call to BUBPT (T,X,Y,P). We
made the call general enough to cover many
Object-Oriented Simulation 77
cases, but how do we keep track of the
cases? If we decide to always use relative
volatilities (Equation 5-4) in the body of
BUBPT, then that means that all simula-
tions using the tray section module will use
relative volatilities for equilibrium. That is
not what we had in mind. We could of
course implement a different equilibrium
method in BUBPT, but even then, all simu-
lations using the tray module would be
forced to use the same equilibrium calcula-
tion. For dynamic simulations, this is most
undesirable because different columns have
different modelling requirements. For ex-
ample, columns separating components with
low relative volatilities usually require many
trays (80 to 120). These systems are almost
always ideal, meaning that relative volatili-
ties adequately describe their equilibrium.
Using Equation 5-4 in dynamic simulations
of such columns is very desirable because
the calculation of equilibrium is explicit and
fast. However, not all systems are ideal and
they should therefore not be modelled by
relative volatilities. Wide boiling fluid sys-
tems are best described by vapor pressures,
possibly in combination with activity coef-
ficients. The bubble point calculation for
liquids modelled by vapor pressures and
liquid activity coefficients is extremely time
consuming and can easily make dynamic
simulations of columns with many trays im-
practical. Fortunately, the separation of
wide boiling systems requires only a few
trays, making these columns as feasible to
simulate as the low relative volatility
columns provided we use the appropriate
bubble point calculation in each case. How-
ever, we just finished saying that all our
simulations will use the same version of
BUBPT, which implies that some of our
applications using the tray module will be
incorrect or hopelessly slow depending on
which method we use in BUBPT. To resolve
this dilemma, we could of course create a
couple of different tray modules that make
calls to different BUBPT routines (BUBPTl,
BUBPT2, etc.), but now we are back where
we started in version management and pro-
liferation of similar looking code.
78 Practical Distillation Control
Problems like these, which only magnify
as we go beyond a single tray and into more
complex models, have prevented develop-
ment of the ideal dynamic simulator. It is
virtually impossible to come to grips with
these problems when using a procedural
approach. On the other hand, object-ori-
ented programming elegantly removes these
obstacles. To see how this is possible I need
to describe how a tray is modelled in an
object-oriented style.
Object-Oriented Approach
In the object-oriented approach we start by
creating two different classes. One class
would model the tray itself and its behavior,
whereas another class would model the
properties and behavior of the fluid system
on and around the tray. For example the
fluid class might look like this:
class Fluid
{
bubblePointO;
getXG);
getYG);
molesG);
getNumberOfComponentsO;
protected:
/ /data area
};
where the messages bubblePoint(), getX{j) ,
getY{j), moles{j), and getNumberO/Compo-
nents() would invoke specific behavior from
all objects instantiated from class Fluid. The
nomenclature I am using to describe the
classes is similar to that of the object-ori-
ented programming language C+ +. The
data area where compositions, component
quantities, temperature, pressure, and phys-
ical property parameters are stored is pro-
tected from outside use or inspection; only
the message routines have complete access
to these data even though the data are not
formally passed as parameters to the rou-
tines. It is as if the data were present in
invisible COMMON blocks in each of the
message routines.
Next we specify a class for the tray itself.
It might look like this:
class Tray
{
update(dt);
Fluid* liquidIn;
Fluid* liquid Out;
Fluid* vaporIn;
Fluid* vaporOut;
protected:
};
Fluid* holdup;
/ /more data
The Tray class would typically have several
messages that tray objects react to, but only
one of these messages, update, is shown
here. In the protected data section, tray is
referring to a fluid object called holdup. In
the C+ + nomenclature F1uid* designates
the variable type and holdup the variable
name. Our object-oriented model closely
mimics the real world; a tray is a physical
entity that holds a fluid phase. A tray also
knows about its environment such as the
liquid coming in from the tray above and
the vapor leaving the tray. These are all
fluid objects that the tray object can refer
to.
Now let us see what the update(dt) mes-
sage might cause a tray object to do.
update(dt)
(
holdup -+ bubblePointO;
n = holdup -+ getNumberOfComponentsO;
for all n do: (5-5)
holdup -+ moles(j) = holdup -+ moles(j) +
OiquidIn -+ moles(j) + vaporIn -+ moles(j) -
holdup -+ getX(j)* L -
holdup -+ getY(j)*V)*dt;
The message update(dt) takes one external
parameter dt. The tray object receiving the
update message responds by first sending
the message bubblePoint() to its holdup ob-
ject [holdup -+ bubblePointO means "send
the message bubblePoint to the object iden-
tified as holdup"]. The bubblePoint message
requests holdup to calculate its bubble
point. Then update calculates the liquid (L)
and vapor (V) flow rates leaving the tray
(code not shown here). Next, update asks
holdup how many components it has and in
return informs holdup how each of these
components will change as a result of inte-
grating the component balance. The compo-
nent balance is derived from Equation 5-2
integrated with a simple Euler integration
step dt.
Comparison of Approaches
Equation 5-5 implements the object-ori-
ented form for tray component dynamics,
whereas Figure 5-2 shows the procedural
version. There are several important dif-
ferences between these implementations. In
Figure 5-2, we are explicitly declaring and
referring to the mole fractions X and Y that
are stored as part of the tray module. In
Equation 5-5, the composition data are not
stored as part of the tray object but instead
as part of each fluid object (e.g., holdup).
However, even in the fluid object, the data
are not explicit because we are not directly
referring to the mole fractions in holdup
(e.g., holdup.x[j)). Instead, we are referring
to the mole fractions in an abstract way by
sending a message to holdup asking it to get
the appropriate mole fraction for us [e.g.,
holdup -+ getXG)]. This concept is called
data abstraction and is another key idea in
object-oriented programming. Data abstrac-
tion allows us to change the data structures
in classes without affecting any code that
uses objects instantiated from those classes.
For example, initially we might have chosen
to store the mole fractions X and Y as
arrays in the fluid class. Later we can change
our mind and decide not to store the mole
fractions explicitly but to store the number
of moles of each component and calculate
Object-Oriented Simulation 79
X and Y as we need them. With data ab-
straction such a change does not impact
users of the Fluid class (e.g., tray objects).
5-3-3 Inheritance and Polymorphism
We are now ready to tackle the problem of
different equilibrium subroutines for dif-
ferent column applications. First, assume
that fluid objects calculate their bubble
points from relative volatilities (Equation
5-4). Furthermore, we have built a general
purpose distillation column from the tray
class. Now, we consider an application
where vapor pressures and activity coeffi-
cients should be used to correctly calculate
the bubble point. This is where we got stuck
before with procedural programming. Ob-
ject-oriented programming allows us to pro-
ceed. First we use the object-oriented con-
cept of inheritance to derive a new fluid
class from Fluid. Call the new fluid class
VaporPressureFluid:
class VaporPressureFluid : public Fluid
{
bubblePointO;
};
Creating this new class was simple enough:
All we had to do was to give the class a
distinct name and mention from what other
class it should inherit properties and behav-
ior. A VaporPressureFluid object will thus
have all the properties and behavior of a
Fluid object. They will only differ in one
aspect; they will calculate their bubble points
using different methods. A Fluid object still
calculates bubble points from relative
volatilities, whereas the VaporPressureFluid
will use vapor pressures and activity coeffi-
cients to complete its bubble point cal-
culation. The interesting part is that in both
cases, the message is the same: bubble-
Point(). We can thus keep our tray class and
its update message intact because update
was sending that very same message bubble-
Point() to its holdup. We simply have to
provide tray objects with Fluid objects when
we want to continue calculating bubble
80 Practical Distillation Control
points with relative volatilities or provide
the tray objects with VaporPressureFluid
objects if we want the bubble point to be
calculated using vapor pressures. This state-
ment might seem puzzling at first
because the Tray class is declared with
references to Fluid and not to Vapor-
PressureFluid. How then can we arbitrarily
substitute a Fluid object with a VaporPres-
sureFluid object? This is where the most
powerful concept of object-oriented pro-
gramming enters, namely, polymorphism.
Polymorphism means many shapes and
refers to the principle that any object in-
vaporla IiquiclOat
stantiated from a derived class (e.g., Vapor-
PressureFluid) can replace an object from a
more basic class (e.g., Fluid) without telling
the user class (Tray) about the substitution.
Tray objects send the messages as before
assuming that they are referring to Fluid
objects when in reality all those objects re-
ceiving the messages are VaporPressure-
Fluid objects and as such respond in their
own way to the same messages. Not only
can we perform this powerful substitution
of base classes with derived classes when we
link our modules together, but we can also
make the substitution while the program is
_----.... lIamer-l
so
ColumnS
-----.... lIamer-2
Update(dl)
Tray object
vaporla IiquiclOat
OGURE 5-3. Graphical representation of polymorphism.
executing. The run time assignment of ob-
jects from derived classes is called dynamic
binding and is probably the most intriguing
aspect of object-oriented programming
(Cornish, 1987).
Figure 5-3 summarizes the concepts of
inheritance, polymorphism, and dynamic
binding as applied to the fluid objects
around the trays in a distillation column.
Two different columns are part of a simula-
tion separating several components. Both
columns are instantiated from a generic
Column class built from the Tray class by
connecting Tray objects together. When a
Column object is instantiated from a Col-
umn class we specify how many stages it will
have. This can be done by the user during
execution of the program and does not re-
quire separate compilation. Column A is
removing light ends from a mixture of mostly
isomers and requires only 10 stages. Col-
umn A should be simulated with a rigorous
bubble point calculation involving vapor
pressures and activity coefficients. Column
B on the other hand separates the isomers
and needs 80 stages to do the job. For
column B, relative volatilities are adequate
to describe the vapor-liquid equilibrium.
The two columns can be simulated together
using the same Column class with different
fluid objects on the trays. The different fluid
objects can be specified by the user, at run
time, when the user requests two different
instances of the same Column class.
5-4 DISTILLATION COLUMN
SIMULATION WITH
OBJECT-ORIENTED
PROGRAMMING
I have demonstrated how inheritance, poly-
morphism, and dynamic binding help build
a generic distillation column tray section
that can be used with many different fluid
systems and physical property calculation
methods. I will now describe how these same
concepts and the notions of encapsulation
and data abstraction help us build complete
Object-Oriented Simulation 81
simulation systems that are very flexible.
Before doing this, it is necessary to review
the principles of model structuring (Astom
and Kreutzer, 1986; Mattsson, 1988).
5-4-1 Structured Models
The first principle of model structuring is
that of hierarchical submodel decomposition,
which suggests that a complex system should
be built from a set of small, self-contained
units. The small units should in turn be
built from yet smaller units until there is a
smallest logical unit that corresponds to
some physical entity in the real world. In
the other direction, a complex unit can be
part of an even more complex system. A
distillation column is an excellent example
of the potential use of hierarchical submod-
els because a complete column with trays
and auxiliaries can be considered a complex
system. According to the hierarchical sub-
model principle, the column should be built
from small pieces, in this case a tray section
module, a reboiler module, and a condenser
module. Furthermore, the tray section mod-
ule should be a collection of individual tray
modules, which in turn contain different
fluid phase modules. In a different hierar-
chy, a column can be part of a larger, more
complex system, say a refining train or a
complete plant.
The second modelling principle we need
to consider is that of parametrization
(Mattsson, 1988; Nilsson, 1989). In design of
the modules for a hierarchical subsystem, it
is important that these modules be flexible
and modifiable. If we can modify a module's
behavior based on parameters we give it, we
are using parametrization. A simple exam-
ple is the number of trays in a distillation
column tray section. A generic tray section
of say 20 trays might be useful for many
simulations but not nearly as useful as a
generic tray section with a user-specified
number of trays. The substitution of differ-
ent fluid objects in the tray class can be
viewed as another example of parametriza-
tion (Nilsson, 1989). Here the tray is the
82 Practical Distillation Control
flexible module using the fluid system as a
parameter to modify its behavior.
The concept of terminals is the third
modelling principle we should consider
(Mattsson, 1989). A terminal on a software
module is analogous to the antenna termi-
nal on a regular television set or the serial
port on a personal computer. These termi-
nals provide agreed upon formats for com-
municating information to or from the unit.
The serial port on a personal computer
might be the best example of a terminal.
From the computer's point of view, we need
not know what device is at the other end of
the cable attached to the serial port. We
only need to know how to read from and.
write to the serial port itself. Examples of
terminals on a distillation column module
would be the feed nozzle, distillate nozzle,
and bottoms nozzle. If these terminals have
the characteristics of a fluid, we can model
and simulate a column without knowing
what is attached to the other ends of the
pipes connected to those nozzles.
The last modelling concept we will con-
sider is that of connections (Mattsson, 1989).
Connections are modules that connect units
together via their terminals. For a distilla-
tion column, the connections correspond to
the pipes that connect various submodules
together. For example, the reflux pipe on a
real column would be represented by a con-
nection module in a model of the column.
5-4-2 Structured Models and
Object-Oriented Programming
Hierarchical submodules, parametrization,
terminals, and connections are concepts that
have been used successfully in many process
simulation systems. They do not imply nor
are they synonymous with object-oriented
programming. For example, both DYFLO
and SPEEDUP have subunits that can be
connected together with terminals and
streams. Parametrization is also used in
these programs to modify structure and be-
havior. However, neither DYFLO nor
SPEEDUP uses an object-oriented para-
digm. So why do I advocate object-oriented
programming as the preferred way of imple-
menting dynamic simulation systems? For
the simple reason that object-oriented pro-
gramming makes it extremely easy to imple-
ment and fully exploit model structuring
principles.
For example, there is a very natural map-
ping between hierarchical submodels and
classes. I have already indicated how the
Tray class was used to build a tray section.
Similarly, Condenser and Reboiler classes
can be used with the tray section to make a
complete column. The object-oriented prin-
ciples of encapsulation and data hiding en-
courage making these units self-contained
and autonomous. Their only communication
with the outside world is through terminals
and connections that can be made very ver-
satile through data abstraction and poly-
morphism.
In this fashion I have used object-ori-
ented programming to build an extremely
flexible dynamic simulator for distillation
columns. One generic class represents the
tray section, one class represents a total
condenser, and one class represents a ther-
mosiphon reboiler. These three classes are
then connected to form a generic Column
class. From the Condenser class, several
other types of condensers, such as partial
condensers, condensers handling inerts, and
spray condensers are derived. Because all
the derived condensers can take the place
of a total condenser by way of polymor-
phism, different types of overhead systems
can be realized with the generic column
without ever having to modify or add to the
code describing the column. This way I did
not have to concern myself up front about
all the possible configurations for a column
because if I want to modify any part of the
column, I simply derive a new class or classes
from the basic ones, slip them into the
generic column class, and the new behavior
appears almost as magic. There is never a
risk of modifying and possibly breaking ex-
isting proven code or subunits. Because I
use the same principle to model the fluid
streams in the column, I can use the generic
column to process literally every possible
fluid system with any physical property
methods I care to consider. This even allows
me to simulate reactive distillation systems
with one and the same generic column
model. After all, chemical reactions on the
trays of a column are not a function of what
column we use but merely depend on the
properties and the state of the fluid we
process in the column.
Flexible dynamic simulators do not come
without a price. The major disadvantage in
working with object-based systems is the
lack of standard design principles for creat-
ing classes. This is bothersome because the
whole concept of object-oriented program-
ming critically hinges on the existence of a
few powerful base classes such as the Fluid
class and the Tray class described earlier.
Without these well thought out base classes,
object-oriented programming can turn into
a jungle of many small, specialized classes
that are hard to use and keep track of. It
takes considerable time and effort, includ-
ing many design iterations, to come up with
a good set of base classes. However, once
this front end loading effort is completed,
the downstream possibilities are virtually
endless.
5-5 EXPERIENCE IN USING
OBJECT-ORIENTED SIMULATION
FOR DISTILLATION
The vision of the ideal dynamic simulator
was created independently of any perceived
relation to object-oriented programming. So
when I originally started to explore object-
oriented programming for dynamic simula-
tions, I did it for very specific and less
visionary reasons. The first reason was to
provide a better framework for reusing the
code in the many custom simulations I had
developed over time. Often new simulation
models were sufficiently different from pre-
vious ones such that the old code could not
be used without substantial error prone
Object-Oriented Simulation 83
surgery. This forced me to start from scratch
many times. On the other hand, as I devel-
oped a new simulation, I always felt that
there were more than 50% common code
and concepts from simulation to simulation.
By using model structuring principles and
object-oriented programming, I have made
it possible to reuse substantial amounts of
proven code. For example, when it used to
take me a couple of months to write, debug,
and run a dynamic model of a distillation
column, I can today put together and evalu-
ate the same simulation in a few days. Fur-
thermore, I hand over an interactive model
to the plant personnel so they can make
their own assessment of what needs to be
done in the process. The second reason I
started to look at object-oriented program-
ming was the desire to have an interactive
simulator that engineers and operators could
use to learn the principles of process dy-
namics and process control. The object-ori-
ented features I have described make it
relatively easy to write programs that simu-
late different unit operations or even a
whole process. After all, the concept of
object-oriented programming grew out of
the language SIMULA, which had been
specifically developed in the 1960s to deal
with large scale dynamic simulations
(Magnusson,1990).
In relation to languages, it might be of
interest to mention which ones I have tried
and what I believe works. There are many
different high-level languages that can be
classified as object-oriented. Smalltalk is
considered the purest object-oriented lan-
guage. In Smalltalk everything is an object
and all computing is done by message pass-
ing. Because Small talk is an interactive lan-
guage, there is no distinction between the
development environment and the run time
simulation. New classes can therefore be
defined at run time. This makes Smalltalk
ideal for model building because hierarchi-
cal models can be built and tested without
having to explicitly compile and link the
program. Although run time interaction is
the strength of Smalltalk, it is also the ma-
84 Practical Distillation Control
jor weakness when it comes to simulation. I
have found Smalltalk to be unacceptably
slow for distillation simulation purposes.
Lisp is another language often men-
tioned in an object-oriented context. Al-
though Lisp itself is not object-oriented it is
. '
easIly extended to be so. Flavors is one of
the most used extensions to Common Lisp
whereas CLOS has been accepted by ANSI
as the standardized extension. Lisp and its
object-oriented extensions offer faster alter-
natives to Small talk when it comes to simu-
lation because they support numeric calcu-
lations without message passing. Lisp is,
however, difficult to learn and master and is
therefore not something I recommend for
practical simulation applications.
C++ is an extension to C that supports
object-oriented programming. It is also in-
tended to be a better C. C++ is a compiled
language and is therefore very fast during
execution, which is a most desirable feature
for large distillation simulations. Although
C+ + requires more work than Smalltalk to
provide interactive model building it is
clearly superior when it comes to program
size and execution speed. This has enabled
me to develop large scale simulations in
C++ that run many times faster than real
time even on a personal computer. I feel
that C++ is currently the most suitable
language for object-oriented simulation of
chemical systems.
Run time speed is extremely important
for dynamic simulations. That is primarily
why C++ is preferable to Smalltalk and
Lisp. However, because C+ + is compara-
ble to FORTRAN and Pascal in run time
speed, you might ask why I advocate C+ +
over these older and more common lan-
guages? The answer is that run time speed
is only one issue in dynamic simulation and .
the total time from problem formulation to
solution is what really counts. Here is where
object-oriented techniques have significant
advantages in that they encourage generic
module building. Once the generic modules
are available, complex dynamic models can
be constructed in extremely short time.
5-6 CONCLUSION
By using object-oriented programming I
have already realized the two major goals of
code reuse and a simple configurable train-
ing simulator as described. However, with
recent advances in computer hardware and
compilers I could use object-oriented pro-
gramming to move closer to the ideal simu-
lator. Needless to say this task is difficult to
achieve because such a simulator must not
only have a large dose of chemical process
simulation capability, but must also have a
flexible and user-friendly interface. Without
doubt, I see object-oriented programming
as the right way to get there.
Commercial products will soon be avail-
able to meet the need for the ideal simula-
tor. Some of these products will be based on
object-oriented principles. One could per-
haps argue that it is irrelevant what lan-
guage was used in creating a simulator that
is run time configurable; the user and the
computer do not care if the source code was
FORTRAN, Pascal, or C+ +. However
. '
there IS always a need to make new simula-
tion units or to modify existing ones. The
user should be able to do so with inheri-
tance and dynamic binding features that
encourage building libraries of small effi-
. '
clent, and well-tested modules that can be
shared across an organization. This ex-
tremely powerful aspect of a good simulator
has, to date, been fully explored only in a
few commercial products (Bozenhart and
Etra, 1989).
References
Astrom, K. J. and Kreutzer, W. (1986). System
representations. IEEE Third Symposium on
Control System Design, Ar-
lmgton, Virginia. New York: IEEE.
Biegler, L. T. (1989). Chemical process simula-
tion. Chem. Eng. Prog. 85,50-61.
Bozenhart, H. and Etra, S. (1989). Intelligent
simulation (Mercury ISIM). Product publica-
tion, Artificial Intelligence Technologies, Inc.,
Hawthorne, NY.
Cornish, M. (1987). What would you do with
object-oriented programming if you had it?
FYAI-T/ Technical Tips 1(10), 3-12.
Franks, R. G. E. (1972). Modeling and Simula-
tion in Chemical Engineering. New York:
Wiley Interscience.
Luyben, W. L. (1990). Process Modeling, Simula-
tion and Control for Chemical Engineers, 2nd
ed. New York: McGraw-Hill.
Magnusson, B. (1990). European perspective. 1.
Object-Oriented Program. 2, 52-55.
Mattsson, S. E. (1988). On model structuring
concepts. 4th IF AC Symposium on Com-
puter-Aided Design in Control Systems, P. R.
China.
Object-Oriented Simulation 85
MaUsson, S. E. (1989). Modeling of interactions
between submodels. 1989 European Simula-
tion Multiconference, ESM'89, Rome, Italy.
Nilsson, B. (1989). Structured modeling of chem-
ical processes with control systems. AIChE
fall meeting, San Francisco, Nov. 1989.
Pascoe, G. A. (1986). Elements of object-ori-
ented programming. BYTE Aug., 139-144.
Perkins, J. D. (1986). Survey of existing systems
for dynamic simulation of industrial process.
Modeling, Identification and Control 7, 2.
Stephanopoulos, G., Henning, G., and Leone, H.
(1990). MODEL.LA., a modeling language
for process engineering. Comput. Chem. Eng.
14(8), 813-869.
6
Plantwide Process Control
Simulation
Ernest F. Vogel
Tennessee Eastman Company
6-1 INTRODUCTION
Process simulation is a valuable tool for the
design of process control strategies, par-
ticularly for distillation columns. Typically,
control strategies use conventional single-
input-single-output (SISO) PID controllers.
Therefore, designing a control strategy nor-
mally includes determining which process
variables should be measured and con-
trolled and which control valve should be
paired with each controlled variable. Con-
trol strategy design also includes identifying
locations in the process where conventional
single-input-single-output PID controllers
may not work well.
Distillation columns are relatively com-
plex unit operations and it is often difficult
for an engineer to develop a satisfactory
distillation control strategy without the as-
sistance of a distillation column simulator.
In many cases, distillation columns operate
somewhat independently of other unit oper-
ations and the columns can be studied one
at a time. For example, for a process where
all unit operations are in series with no
recycle (Figure 6-1), a control strategy can
be developed for each column one at a
time. Additionally, tanks on column feed or
product streams isolate columns, allowing
column control strategies to be developed
86
independently (Figure 6-2). However, for
processes with recycles (Figure 6-3) or pro-
cesses with coupled distillation columns
(Figure 6-4), a distillation column control
strategy often must be developed in consid-
eration of associated unit operations. Chap-
ter 20 provides additional examples.
Studying the control and operation of
distillation columns and associated equip-
ment requires a plantwide process control
simulator. A plantwide process control sim-
ulator can simulate multiple unit operations
of various types (a ftowsheet simulator) and
can perform both steady-state and dynamic
simulation.
This chapter describes the use of a
plantwide process control simulator for con-
trol strategy design and for other problems
as well. The chapter also discusses model
accuracy, simulation environments, and nec-
essary features of a plantwide process con-
trol simulator.
6-2 APPLICATIONS OF A
PLANTWIDE PROCESS
CONTROL SIMULATOR
In addition to the development of process
control strategies, there are a number of
other related problems where a plantwide
process control simulator can be applied.
Plantwide Process Control Simulation 87
Feeds
B
o
FIGURE 6-1. Process with units in series.
Feeds
Tank 1
'---...L.-------+I Tank 2 I--------ot
o
FIGURE 6-2. Process with intermediate storage tanks.
88 Practical Distillation Control
Feeds
FIGURE 6-3. Process with recycles.
c
FIGURE 6-4. Coupled distillation columns.
The applications fall into three major cate-
gories: process control, process design, and
process safety. The following sections dis-
cuss the application of a plantwide control
simulator in all three of these areas.
B 6-21 Process Control
Control Strategy Design
Control strategy design is the procedure of
determining which process measurements
need to be controlled and which valves
should be used to control them. This proce-
dure assumes that conventional single-
input-single-output PID controllers will
work satisfactorily, which is normally the
case. This procedure also identifies cases
where the conventional control algorithm is
not sufficient. Steady-state simulation with
sensitivity analysis and interaction analysis
is valuable for identifying good potential
control strategies. Dynamic simulation pro-
vides a demonstration of how the candidate
control strategies will actually perform for
expected disturbances and setpoint changes.
Justification of Control Strategy Modifications
Changing the control strategy in an existing
plant typically requires convincing plant en-
gineers to agree to the change. Changing
the strategy also requires justifying any ex-
pense of implementation, process downtime
or hardware and software expenses. Dy-
namic simulations showing how both the old
and the new strategies perform for typical
process disturbances often can accomplish
both of these tasks.
On-line Analyzer Justification
On-line analyzers, which provide process
composition measurements either continu-
ously or frequently (every 10 min or less),
can provide significant improvements in
process control. However, many on-line an-
alyzers are more expensive and require more
maintenance compared to more common
measurements such as pressure and temper-
ature. Dynamic simulation can demonstrate
how much an on-line analyzer can improve
control through use in a feedforward or
feedback control scheme. Dynamic simula-
tion can also be used to determine the fre-
quency at which new composition measure-
ments must be available.
Advanced Control Algorithm Evaluation
Advanced control algorithms, such as pre-
dictive control, require significantly more
effort to implement and maintain than con-
ventional PID control. Implementation of
advanced control algorithms on realistic
process models provides an evaluation of
their performance compared to conven-
tional control and justification of their im-
plementation.
Batch Process Control and Operation
Dynamic simulation provides a method to
evaluate strategies and procedures for con-
trolling and operating batch processes. Ex-
amples of batch control and operation
issues include tracking temperature trajec-
tories and determining cycle times and op-
erating sequences.
6-22 Process Design
Surge Capacity Sizing
Intermediate holdup tanks provide attenua-
tion of flow rate or composition distur-
bances before they reach downstream
Plantwide Process Control Simulation 89
equipment. Such surge tanks are essential
for interfacing batch and continuous pro-
cesses. Simulation provides a way to deter-
mine how large surge tanks need to be and
where they should be located to sufficiently
attenuate harmful disturbances.
Flowsheet Design
Some process flowsheet designs that pro-
duce the same product are inherently easier
to control and operate than others. The
dynamic performance and operability of al-
ternate flowsheet designs can be compared
using a plantwide dynamic simulation.
Heat Integration Sysums
Although heat integration can result in sig-
nificant energy savings, it can also cause
process disturbances to propagate to parts
of a process that would not be affected
without heat integration. Plantwide process
control simulation with and without heat
integration provides an evaluation of the
consequences of additional disturbance
propagation.
6-23 Process Safety
Emergency Shutdown Evaluation
Most processes have emergency shutdown
systems to prevent the occurrence of a
catastrophe if the process moves to a dan-
gerous operating condition. Unfortunately,
the performance of these systems can only
be truly tested if an event occurs. However,
dynamic simulation provides a way to test
and compare emergency shutdown strate-
gies and to quantify how well they would
work.
AccUknt Evaluation
When a process accident occurs, there are
often many proposed explanations and sce-
narios for what happened. Usually, it is not
clear what actually caused the problem.
With dynamic simulation, the proposed ex-
planations can be tested to evaluate their
feasibility. The dynamic simulations often
rule out several of the possibilities and help
point to the real cause.
90 Practical Distillation Control
Operator Training
Good operator training is important with
processes that are relatively complex or po-
tentially hazardous. In these cases, operator
training simulators are very helpful. With an
operator training simulator, the operator's
interface to the "process" is the same con-
trol system operating console as used for
the real plant, but the process measure-
ments are provided by a real-time, dynamic
process simulator and the operator's com-
mands go to the simulator rather than the
actual process. A plantwide process control
simulator facilitates the building of such
operator training simulators.
Pre-Stan-up Knowledge
Running simulations of new processes pro-
vides a feel for how they will operate, how
quickly measurements respond, and how
sensitive they are to the manipulated vari-
ables. This insight is helpful during actual
process start-ups.
6-2-4 Example
Figure 6-5 shows an example process used
here to illustrate the role of a plantwide
process control simulator in the process and
control strategy design. This process has
two feeds, components A and B. A and B
Reactions:
A+B-+R
2B .... S
Feeds
A B
A&B
react together to form the desired product
R. B reacts with itself to form the undesired
by-product S. The reactions are exothermic.
The process consists of the reactor and the
two distilltion columns to separate the reac-
tant and products. The reactor includes in-
ternal cooling coils to remove the heat of
reaction. The first distillation column sepa-
rates the reactants (A and B) from the
products (R and S). The reactants are recy-
cled back to the reactor. The second distil-
lation column separates the products (R
and S). In the design of the process, the
control, and the safety system for this exam-
ple, there are several issues where simula-
tion with a plantwide process control simu-
lator is helpful.
Process Control
What are possible control strategies? How
well will each perform for expected distur-
bances? Where should the production rate
be set: the feeds to the reactor, the feed to
the first column, another valve in the pro-
cess?
What variables need to be controlled?
Should reactor temperature be controlled?
Which temperature or temperatures should
be controlled in the distillation columns?
Is temperature control on the columns
sufficient, or are composition measurements
N
C

8
s
FIGURE 6-5. Example process for plant-wide simulation.
needed? How much might they help? How
frequently are composition measurements
needed?
Are SISO PID controllers satisfactory?
Are there any severe interactions? Are they
bad enough to justify a more advanced algo-
rithm? How much will an advanced algo-
rithm help?
Which disturbances have the most effect?
What happens if feed A or feed B is in
excess?
Process Design
Is a surge tank needed between the
columns? The simulation can justify it if
needed or prove that one is not necessary if
that is the case.
Is the reflux drum on the first column big
enough so that disturbances in the column
are sufficiently attenuated and do not upset
the reactor through the recycle?
Process Safety
Is the exothermic reaction potentially dan-
gerous? If so, will stopping the feeds and
maximizing the cooling water flow stop a
temperature excursion? Could a change in
the recycle flow or composition initiate a
temperature excursion?
6-3 BENEFIT FROM PLANTWIDE
PROCESS CONTROL SIMULATION
The primary result or benefit from process
simulation is improved process insight and
understanding. The thinking process the
user goes through in developing and run-
ning a simulation is what leads to the im-
proved process insight and understanding.
Better process understanding results in the
formulation of better control strategy de-
signs, as well as better flowsheet designs
and better safety systems. A process simula-
tor does not explicitly provide such solu-
tions, but as a thinking tool it assists in the
development of solutions. A simulation re-
veals which process parameters are impor-
tant and which ones are not. Sometimes the
Plantwide Process Control Simulation 91
simulated dynamic behavior is easily ex-
plained, but counterintuitive. The mind does
not easily predict dynamic responses by su-
perimposing all of the simultaneous dy-
namic effects in a chemical process, particu-
larly if the process contains one or more
recycles.
The explicit results from process control
simulation tend to be more qualitative than
quantitative. Application of process control
simulation often results in a decision as
illustrated in the example above, for exam-
ple, use control strategy B rather than strat-
egy A, buy a continuous analyzer for use in
a feedback or feedforward strategy, a surge
tank is necessary, and so forth. To contrast,
process design simulations often provide ab-
solute numbers, for example, reboiler heat
duty, distillation column diameter, and so
forth.
6-4 DEFINING THE SCOPE
OF A PLANTWIDE PROCESS
SIMULATION
The first step in performing a plantwide
process control simulation is to define the
scope of the simulation. The scope includes
the size and the accuracy of the simulation.
The size of a simulation refers to how much
of the process is simulated. The accuracy of
a simulation refers to how well the steady-
state and dynamic responses of the simula-
tor match that of the actual process. To
minimize time and effort in building and
running a simulation, the engineer deter-
mines the size and accuracy actually needed
in order to avoid simulating more of the
process than necessary or simulating the
process more accurately than necessary.
Both the size and accuracy needed for a
simulation depend on the problem being
solved and normally must be determined on
a case-by-case basis.
In defining the size of the simulation, for
example, which unit operations to simulate,
the engineer looks for logical breaks in
the process. Large interprocess inventories
92 Practical Distillation Control
make good boundaries for simulations be-
cause most disturbances are significantly at-
tenuated at those points. Also, portions of a
process that are not connected by recycles
can be segregated.
The accuracy of a simulation is not easily
measured. In some cases, the plant does not
yet exist. For existing processes, perturbing
the plant to observe dynamic responses is
usually not desirable. Generally, to obtain
more accuracy, the simulation must use
more rigorous or complex models of unit
operations, valves, and so on.
In control strategy design, dynamic simu-
lation is typically used for making relative
performance comparisons between alter-
nate control strategies. When making rela-
tive comparisons, a high degree of accuracy
is not as important as when determining
absolute values. If two control schemes are
significantly different, simulations will dem-
onstrate that fact, even if the model re-
sponse does not exactly match the process
response. If the differences between two
schemes are small, other criteria should be
used to choose between them, for example,
ease of operator understanding and ease of
implementation and support.
As an accuracy guideline for control
strategy design, match the simulation's
steady-state base case to the plant opera-
tion or to the steady-state process design
conditions to within 5%. These conditions
include flows, temperatures, pressures, com-
positions, levels (inventories), and vessel
volumes. Note that with good estimates of
vessel volumes, and flow rates, residence
times (time constants) will be similar to
those of the process. Thus, validation of
dynamic responses is generally not neces-
sary for control strategy design. If the pro-
cess conditions are matched as previously
described, the process dynamics are usually
sufficiently accurate for control strategy
comparisons.
Note, however, that for other applica-
tions model accuracy is much more impor-
tant. For instance, when evaluating an
emergency shutdown system on an exother-
mic reactor and determining how high the
temperature will go during an upset and
how quickly it will rise, accuracy of the
model is very important.
65 BUILDING A PLANTWIDE
PROCESS SIMULATION
There are several simulation environments
an engineer can use to build a dynamic
flowsheet simulation. Three possibilities are
described in the following subsections.
6-51 Programming Environment
In this case, the engineer writes source code
for the differential and algebraic equations
that describe the process to be simulated.
The programmer must provide numerical
methods and interface routines as well.
Simulations built in this environment tend
to be customized for simulating one particu-
lar process and are not easily modified for
other problems. Writing and debugging sim-
ulations is likely to be time intensive. This
approach offers the most flexibility, but re-
quires a high level of expertise in both mod-
elling and programming.
6-52 Equation Solving Environment
The equation solving environment is a
higher level than the programming environ-
ment. Similar to the programming environ-
ment, the engineer must provide the differ-
ential and algebraic equations that describe
the process to be simulated. However, the
numerical method and interface routines
are provided. This approach is similar to the
programming approach in that simulations
tend to be customized and writing and de-
bugging the simulation can be time consum-
ing. Although this environment may not be
quite as flexible as the programming envi-
ronment, the total time required to build
the simulation and the programming exper-
tise required will likely be less.
6-5-3 Steady-State - Dynamic
Flowsheat Simulation Environment
A steady-state-dynamic flowsheet simulator
offers a significantly higher level simulation
environment compared to the previous two
options. A steady-state-dynamic flowsheet
simulator has preprogrammed unit opera-
tion models and control algorithms and the
engineer builds the simulation by interac-
tively linking the necessary pieces together.
The simulator also provides all the needed
interfaces and numerical methods. Thus,
this approach requires a lower level of ex-
pertise in the modelling and programming
areas. Also, because models are prepro-
grammed, significantly less time is necessary
for building and debugging simulations.
However, this approach offers less flexibility
for unusual cases because the engineer is
limited to the provided library of process
and controller models. To address this po-
tential problem, the simulator may offer a
way to build custom models, but again the
use of custom models is limited by the engi-
neer's time and modelling expertise.
6-5-4 Practical Considerations
The preceding three options offer a trade-off
between flexibility and ease of use. The first
two options offer more flexibility for creat-
ing custom models that may more closely
represent the actual process. Realistically,
however, these two choices are of little value
to most control strategy design engineers.
The time required to build custom simula-
tions starting from the point of writing dif-
ferential equations is often prohibitive ex-
cept in critical cases. The control strategy
designer is likely to be in an environment
where the schedules and workload make
quick turnaround of projects essential. De-
sign schedules allow little time for process .
simulation. Process control trouble shooting
type problems often need answers as soon
as possible, typically within two to three
days. Further, the control strategy design
engineer may not have the necessary model-
Plantwide Process Control Simulation 93
ling or programming expertise. Thus, the
interactive flowsheet simulator option ap-
pears to be the best choice for most control
strategy designers.
Given the control strategy designer's lim-
itations of time and expertise, it is impor-
tant that the simulator be interactive. An
interactive simulator includes a menu driven
or graphical interface for building and
changing the simulation. Such an interface
allows the user to quickly and easily exam-
ine or change any parameter. Interactive
means that the user can monitor the
progress of a simulation and change param-
eters during the simulation or that the user
can interrupt a simulation, change a param-
eter or configuration, and resume the simu-
lation from where it was interrupted. Inter-
active also means that building or changing
a simulation never requires recompiling or
relinking the simulator. Interactive does not
mean: editing an input file with a screen
editor, searching for and changing a param-
eter, exiting the editor, and then issuing a
command that recompiles and relinks the
simulator and then starts the dynamic simu-
lation. The effort and time such an environ-
ment requires discourages use.
Eliminating the need for programming
greatly speeds the building of a simulation.
In order for a user to avoid programming,
the simulator must have the needed capabil-
ities readily available to solve typical prob-
lems: physical properties for common
chemicals, unit operations, and control al-
gorithms should all be available from data
bases or libraries.
Although not all processes can be simu-
lated using a unit operation library, most
control strategy design problems can be sat-
isfactorily simulated with a library of tradi-
tional chemical engineering unit operations.
For instance, most liquid phase reactors can
be modeled as simple continuous stirred
tank reactors (CSTRs). This type of as-
sumption is possible because a high level of
accuracy is generally not needed for control
strategy design. As discussed previously, less
accuracy will suffice because control strat-
94 Practical Distillation Control
egy design simulations involve relative per-
formance comparisons. However, some pro-
cesses contain very unique unit operations
and no matter how complete the unit opera-
tion library in a ftowsheet simulator, there
will be problems that require custom model-
ling.
6-6 FEATURES OF A PLANTWIDE
PROCESS SIMULATOR FOR
CONTROL STRATEGY DESIGN
There are several tools in a plantwide pro-
cess control simulator that set it apart from
other more general ftowsheet simulators.
A key feature is the capability to perform
both steady-state and dynamic simulation.
Steady-state simulation facilitates the devel-
opment and initial screening of candidate
control strategies by providing a quick way
to calculate several steady-state perfor-
mance measures.
A common first step in control strategy
design is the calculation of steady-state gains
for selected controlled (c) and manipulated
(m) variables. Steady-state gain may be de-
fined as iJc/iJm. To calculate the steady-state
gains, each manipulated variable is per-
turbed one at a time from an initial base
case condition. After each manipulated
variable is perturbed, the simulation is re-
converged to steady state and the changes
in the controlled variables are calculated.
The derivative is then approximated as
ac/ am.
From steady-state gains, the simulator
can then calculate several interaction mea-
sures as described in Chapter 8: relative
gain array (Bristol, 1966; McAvoy, 1983),
singular value decomposition (Forsythe,
Malcolm, and Moler, 1977), Niederlinski's
stability index (Niederlinski, 1971), and
block relative gain (Manousiouthakis,
Savage, and Arkun, 1986).
Also from steady-state simulation, sensi-
tivity plots as illustrated by Luyben (1975)
can be generated. These sensitivity plots
have information similar to the steady-states
gains, but the plots also reveal the nonlin-
earities of the process. Figure 6-6 shows an
example of a sensitivity plot for the first
column of the example process (Figure 6-5).
This plot shows how the bottoms composi-
tion varies for three disturbances while a
proposed control strategy holds the con-
trolled variables constant. The curves are
generated by varying each disturbance vari-
able one at a time and reconverging the
simulation to steady state with each varia-
tion. In this case, the study shows that the
0.0040 --r-------------------,
0.0035
Disturbance Variables
B
A
C
A-Feed Rate
B-Feed Temperature
C-Feed Composition
(Weight Fraction
Component R)
0.0030 +----------.-----------1
10 Percent
Decrease
Base Case
Value
Disturbance Variable
FIGURE 6-6. Steady-state sensitivity analysis.
10 Percent
Increase
composition is relatively insensitive to the
feed temperature whereas it is more sensi-
tive to the feed composition, particularly for
increasing amounts of component R in the
feed. If this amount of variation in the B
composition is acceptable, then the pro-
posed control strategy may merit further
evaluation with dynamic simulation. How-
ever, if this amount of composition variation
exceeds acceptable limits, there is no point
in further considering this control strategy.
After developing a set of candidate con-
trol strategies using steady-state analysis,
dynamic simulation provides a demonstra-
tion of how the control strategies will actu-
ally perform. Often, two strategies that
appear to have comparable steady-state
performance greatly differ dynamically. Fur-
ther, the strategy that appears to have the
best steady-state performance may not per-
form well at all dynamically.
Integrating the steady-state and dynamic
simulation capabilities and the control sys-
tem analysis tools into one plantwide simu-
lation package greatly facilitates control
strategy design. Testing control strategies
Plantwide Process Control Simulation 95
using representative, nonlinear process
models provides good indications of their
actual performance.
References
Bristol, E. H. (1966). On a new measure of
interaction for multivariable process control.
IEEE Trans. Automatic Control AC-ll,
133-134.
Forsythe, G. E., Malcolm, M. A, and Moler,
C. B. (1977). Computer Methods for Mathe-
matical Computations. Englewood Cliffs, NJ:
Prentice-Hall.
Luyben, W. L. (1975). Steady-state energy con-
servation aspects of distillation column con-
trol system design. Ind. Eng. Chem. Fundam.
14,321-325.
McAvoy, T. J. (1983). Interaction Analysis, Prin-
ciples and Applications. Research Triangle
Park, NC: Instrument Society of America.
Manousiouthakis, V. Savage, R. and Arkun, Y.
(1986). Synthesis of decentralized process
control structures using the concept of block
relative gain. AIChE 1. 32, 991-1003.
Niederlinski, A (1971). A heuristic approach to
the design of linear multivariable control sys-
tem. Automatica 7, 691.
7
Identification of Distillation
Systems
R. C. McFarlane
Amoco Corporation
D. E. Rivera
1
Arizona State University
7-1 INTRODUCTION
In this chapter we address the problem of
identification of distillation systems for the
purpose of obtaining models for process
control. Much of the control literature has
focused on controller synthesis procedures
that are derived under the assumption that
suitable models are available. Significantly
less attention has been paid to the specific
problem of defining the requirements of
models for process control purposes and
how to make the best choices of design
variables in identification to obtain them.
Among the objectives of this chapter is to
survey the available literature in this area
and present some ideas and procedures that,
when incorporated into the well established
methodology for system identification, make
the identification more relevant to the needs
of process control.
The identification of distillation systems
has many features that are common to most
chemical engineering systems. Although dis-
tillation processes are inherently nonlinear,
if operated over a sufficiently small region,
lA/so affiliated with the Control Systems Engineer-
ing Laboratory, Computer-Integrated Manufacturing
Systems Research Center, Arizona State University.
96
control systems based on linear input-out-
put models often perform satisfactorily on
them. The literature on linear system iden-
tification is a mature one, and it would be
redundant to provide detailed descriptions
here of methods that are well defined in the
literature. Rather, we present an overview
of the general methodology of linear system
identification, and provide details of specific
methods that are "control-relevant" and
useful for identifying distillation systems.
For distillation systems that are highly non-
linear (for example, high purity columns), a
number of different identification ap-
proaches are possible. We do not review
nonlinear system identification because the
control structures that have proven most
useful for distillation, even in cases where
nonlinearity is pronounced, are based on
linear input-output representations. There-
fore, we suggest adaptations of the basic
linear identification approach (e.g., use of
linearizing transformations) to obtain the
required linear models.
The class of multivariable predictive con-
trollers based on linear step response or
truncated impulse response models has
proven useful in a wide range of industrial
distillation systems (see Chapter 12). Direct
identification using these models poses
problems because they are not parsimo-
nious (i.e., compact) system representations.
Overparametrization and lack of indepen-
dence among the model parameters results
in a poorly conditioned estimation problem
and estimates with high variance result when
conventional least-squares estimators are
used. To compensate for this problem, prac-
titioners are forced to perform long dy-
namic process tests to generate large quan-
tities of data for the estimation. In some
plants, performing very long dynamic tests
may be acceptable. Nevertheless, an overall
goal of system identification is to obtain the
required models with the minimum of dis-
ruption to normal process operation. There-
fore, we present two alternative approaches
for obtaining step or impulse models,
namely, estimation of parsimonious low-
order transfer function models (that are
converted after estimation to the required
step or impulse form) and the use of biased
least-squares estimators (for example, ridge
regression and partial least squares). The
use of biased estimators is one way to pro-
vide more efficient estimation if impulse or
step response models are to be identified
directly from process data.
We assume in this chapter that model
estimation is performed off-line using batch
data collected from a dynamic process test
conducted over a fixed period of time for
the purpose of designing a fixed-structure
controller. We do not consider on-line re-
cursive estimation because the resulting
adaptive controllers, particularly for multi-
variable constrained systems, have not
proven reliable enough to be considered as
practical tools for distillation control at this
time.
Finally, we would like to emphasize that
the daunting programming task previously
associated with getting started in system
identification has been eliminated with the
availability of a number of commercial soft-
ware packages for system identification.
These packages provide a wide range of
data analysis and identification tools, imple-
mented with efficient numerical procedures,
Identification of Distillation Systems 97
allowing a user to concentrate on making
the design decisions and judgments involved
in system identification.
The chapter is organized as follows. Sec-
tion 7-1-1 describes the general class of
transfer function models on which the iden-
tification methods covered in this chapter
are based, whereas Section 7-1-2 presents
the overall iterative methodology of system
identification. More detailed descriptions of
each step in the process are provided in
subsequent sections: Section 7-2 examines
the design of perturbation signals to ob-
tain information for single-input-single-out-
put (SISO) and multi-input-multi-output
(MIMO) systems; Section 7-3 covers topics
associated with identifying model structure
and estimation of model parameters, which
includes a discussion of bias-variance
trade-offs in system identification (Section
7-3-1), nonparametric and parametric iden-
tification (Sections 7-3-2 and 7-3-3), con-
trol-relevant identification (Section 7-3-4),
identification in the closed-loop (Section
7-3-5), and the treatment of nonlinearity in
distillation systems (Section 7-3-6). Model
validation is discussed in Section 7-4, some
practical aspects of conducting dynamic tests
on a process are discussed in Section 7-5,
and an example is provided in Section 7-6.
7-1-1 Discrete Transfer Function
Models for Distillation Systems
We assume that the system to be identified
is linear and time-invariant and can be rep-
resented by a transfer function model of the
form
y(t) = Go(q)u(t) + v(t) (7-1)
where
subscript 0 = the true (unknown) system
yet) = a system output (e.g., a tem-
perature or composition mea-
surement from an on-line an-
alyzer, or transformations of
98 Practical Distillation Control
these measurements) mea-
sured at equally spaced inter-
vals of time (the sampling in-
terval T,)
u(t) = a manipulated system input
(or a measured disturbance
variable)
Go(q) = the transfer function repre-
senting the deterministic pro-
cess dynamics
The disturbance v(t) represents the ef-
fect of all unmeasured disturbances acting
on the system, including measurement noise,
and is modelled as
vet) = Ho(q)e(t) (7-2)
where e(t) = a sequence of random
shocks with a specified prob-
ability density function
(PDF)
Ho(q) = the transfer function of the
disturbance process
Equation (7-2) is used to model serially cor-
related stochastic disturbances (as described
in Box and Jenkins, 1976), as well as piece-
wise deterministic disturbances, as discussed
by Astrom and Wittenmark (1984) and
MacGregor, Harris, and Wright (1984). For
piecewise deterministic disturbances, e(t)
represents a signal that is zero except at
isolated points. A more in-depth discussion
of disturbances that typically affect distilla-
tion systems is made in Section 7-3-3-l.
The object of identification is to find a
parametrized representation of the true sys-
tem in the form
yet) = G(q)u(t) + H(q)e(t) (7-3)
Determining specific structural forms for
G(q) and H(q) is part of the identification
problem, as well as designing perturbations
on u(t) to be applied to the process to
obtain data with sufficient information to
perform the identification. In the case of
distillation, the identification problem in-
cludes determining whether transformations
on certain measured outputs are necessary
to allow use of the linear model Equation
7-3. As pointed out in Chapter 12, even
with transformations, highly nonlinear
columns cannot always be adequately repre-
sented by linear models of this form.
Multivariate systems are represented us-
ing extensions of Equation 7-3. For exam-
ple, for a system with k outputs and m
inputs, a convenient form for the MIMO
transfer function model is
Yj(t) = G
i1
(q)U
t
(t) + G;z(q)u
2
(t) + ...
+ Gjm(q)um(t) + Hj(q)ej(t),
i = 1,2, ... ,k (7-4)
Note that this model assumes that distur-
bances act independently on each output.
This assumption is usually made to simplify
the estimation problem. More general mul-
tivariable forms are available in the litera-
ture (e.g., see Gevers and Wertz, 1987;
Ljung, 1987, Appendix 4.A).
7-1-2 Iterative Methodology of System
Identification
System identification has traditionally been
considered as an iterative process, consist-
ing of four main steps:
Step 1. Experimental planning. An engineer
must decide on the type of input [e.g.
pulse, step, relay, or pseudo-random bi-
nary sequence (PRBS)], design parame-
ters associated with each signal type, the
test duration, whether the plant will be
kept in open or closed loop during test-
ing, and, in the case of multivariable sys-
tems, how the test will be conducted (i.e.,
perturbations applied to inputs simulta-
neously or one at a time).
Step 2. Selection of a model structure. Within
the class of linear transfer function mod-
els described by Equation (7-3), a specific
structure must be selected [e.g., AutoRe-
gressive with eXternal input (ARX),
Box-Jenkins (BJ), output error (OE), or
finite impulse response (FIR)] as well as
the orders of the polynomials and time
delays.
Step 3. Parameter estimation. The parame-
ter estimation step involves the selection
of a number of design variables: the nu-
merical procedure to obtain estimates for
the model parameters, the form of the
objective function (e.g., sum of squared
prediction errors), the number of steps
ahead in the prediction-error equation,
and whether to use a prefilter, and if so,
its form. A complicating factor in this
step is the potential for numerical prob-
lems in the parameter estimation algo-
rithm (e.g., ill-conditioning, multiple lo-
cal minima).
Step 4. Model validation. Having obtained a
model, its adequacy must be assessed.
Among the issues to consider are:
What information about unmodelled dy-
namics is resident in the residuals (the
time series resulting from the difference
between the predicted model output and
measured output) and how this informa-
tion assists in the iterative identification
procedure?
Does the step response of the estimated
model G(q) agree with physical intu-
ition?
Although these issues are important, the
ultimate test of model adequacy is
whether it meets the requirements of the
intended application (which in this case
is process control).
In practice, it is desirable to iterate on
Steps 2 to 4 only, because repetition of
dynamic tests is unduly disruptive to process
operation. Because the ability to perform
Steps 2 to 4 efficiently depends to a great
extent on the quality of information avail-
able in the dynamic test data, the design of
a dynamic test to obtain the necessary infor-
mation in one attempt (and often with only
approximate a' priori information about the
system) is perhaps the most challenging as-
pect of process identification. In the follow-
ing sections we discuss each step in more
detail. Much of the discussion is general to
Identification of Distillation Systems 99
the problem of how to perform identifica-
tion when the end-use of a model is for
process control. Topics specific to distilla-
tion are discussed where they arise.
Again, we emphasize that commercial
software packages for system identification
are available today that greatly facilitate
Steps 2 to 4.
72 PERTURBATION SIGNAL
DESIGN
721 Discussion
A primary goal of system identification is to
be able to discriminate among competing
models within a specified model set. This
discrimination (and the related operation of
parameter estimation) will be efficient only
when performed using data obtained from a
process test in which certain identifiability
conditions have been satisfied. These condi-
tions are usually expressed as the require-
ment that a perturbation signal be per-
sistently exciting (e.g., see Ljung, 1987,
Chapter 14). In practice, this requires a
perturbation signal that is sufficiently rich in
frequency content to excite a system up to
some specific frequency, above which the
dynamics are considered unimportant. For
example, for control purposes, we are often
not interested in modelling the fastest modes
of a system (e.g., the dynamics of valve
motion). A number of questions arise:
1. What is the frequency range important
for control?
2. How can one design a perturbation sig-
nal that excites a system over this fre-
quency range without having complete
a priori knowledge of the system fre-
quency response (and controller tuning)?
The first question, as it relates to the
requirements of system identification, is dis-
cussed in Section 7-3-4. The second ques-
tion is addressed in this section. As de-
scribed in Section 7-1-2, the design of a
perturbation signal is part of the iterative
100 Practical Distillation Control
process of identification. In theory, the de-
sign of a control system should also be in-
corporated in this iterative procedure be-
cause controller design and tuning affect
the modelling requirements for process con-
trol and therefore influence signal design,
that is, the frequency range of interest.
However, in practice, including controller
design as an iteration is not necessary; suf-
ficient approximate system information is
usually available to allow informative dy-
namic tests to be designed.
The perturbation signal most often used
to satisfy the preceding identifiability condi-
tions (as well as closed-loop identifiability
conditions; see Section 7-3-5) is the
pseudo-random binary sequence (PRBS)
signal. The design of PRBS signals for
perturbation of single-input-single-out-
put (SISO) and multi-input-multi-output
(MIMO) systems for control-relevant identi-
fication is discussed in the following sec-
tions.
Discrete sampling of a continuous system
leads to inevitable information loss, and
therefore selection of sampling interval is
an important topic in identification for dis-
crete-time control models. Sampling too
slowly can distort the spectrum of a sam-
pled signal in the frequency range of inter-
est as a result of aliasing (or foldover), lead-
ing to faulty identification. The selection of
sampling interval and design of anti-aliasing
filters to avoid this problem is discussed in
most texts dealing with system identification
(e.g., Ljung, 1987, Section 14.5).
7-2-2 Pseudo-random Binary
Sequence Signals
7-2-2-1 PRBS Signals for S1S0 Systems
As described by Davies (1970), a PRBS sig-
nal can be generated using shift registers,
and such signals can be characterized by
two parameters: n, the number of shift reg-
isters used to generate the sequence, and
Tel' the clock period (i.e., the minimum time
between changes in level of the signal, as an
integer multiple of the sampling period Ts).
The sequence repeats itself after T = NTel
units of time, where N = 2
n
- 1. For Tel =
one sampling period, the power spectrum of
a PRBS signal is approximately equal to 1
over the entire frequency spectrum. For val-
ues of Tel > 1, the spectrum drops off at
higher frequency and has a specific band-
width useful for system excitation. The
power spectrum of a PRBS signal generated
using n = 7 and Tel = 3 (with = 1 s) is
shown in Figure 7-1 (plot generated numeri-
cally using the spa function in MA TLAB,
1989).
An analytical expression for the power
spectrum of a PRBS signal generated using
shift registers is given by (Davies, 1970)
_ a
2
( N + 1) Tel [ sine wTel/2) 12
Sew) - N wTel/2
(7-5)
where a = signal amplitude. At low fre-
quencies the power spectrum has the ap-
proximate value a
2
(N + l)T
el
/N. At w =
2.8/T
el
, the spectrum is reduced by half,
defining the upper limit of its useful range.
Due to the periodicity of the autocovariance
function of a PRBS signal, its power spec-
trum is discrete (Eykhoff, 1974) with points
in the spectrum separated by 2'lT /T and the
first point at 2'lT /T. Therefore, the fre-
quency range of a PRBS signal useful for
system excitation is
2'lT 2.8
- s w < - rad/s (7-6)
T - Tel
Selection of n and Tel allows a user to
tailor this frequency range for excitation of
a particular system. If a priori estimates of
the system dominant time constant T dom and
deadtime T d are available, then the follow-
ing can be used as a guideline for designing
a PRBS signal:
1 1
------ s w S ------
fJ( Tdom + Td/2) (Tdom + Td/2)/a
(7-7)
where a is specified to ensure that suffi-
Identification of Distillation Systems 101
SPECfRUM
loo
10-
1
frequency
FIGURE ,.1. Power spectrum of a PRBS signal generated using n = 7 and Tel = 3
(T
s
= 1 s).
ciently high frequency content is available
in the perturbation signal, commensurate
with how much faster the closedloop reo
sponse is expected to be relative to the
open-loop response:
For example, with a = 2, the designer ex-
pects the closed-loop time constant to be
one-half that of the open-loop response (i.e.,
twice as fast). As discussed in more detail in
Section 7-3, the accuracy of the frequency
response of a fitted model, relative to that
of the real process, is greatest in those fre-
quency ranges where the perturbation sig-
nal has most energy. A more aggressively
tuned model-based controller (larger a) will
require an identified model that is accurate
of higher frequencies. Therefore, the per-
turbation signal should be designed (e.g.
using Equation 7-8) to ensure sufficient en-
ergy at higher frequencies.
{3 is specified to tailor how much low-
frequency information will be present in the
perturbation signal (Le., how low in fre-
quency does one wish to model the plant).
Choosing higher values of fJ results in
lower-frequency information: fJ = 3 will
provide information down to a frequency
roughly corresponding to the 95% settling
time of process, fJ = 4 for the 98% settling
time, and {3 = 5 for the 99% settling time.
Comparing Equations (7-6) and (7-7)
gives
and
21T'afJ
N=--
2.8
(79)
Consider the design of a PRBS signal for
a first-order plus dead-time process where
'T = 'Tdom = 100 sand 'Td = 20 s. Choosing
a = 1 (Le., only the open-loop dynamics are
of interest and the controller is not required
to be tightly tuned) gives Tel = 308 s. Thus,
the minimum time between changes in level
of the PRBS signal is roughly 3 time con-
102 Practical Distillation Control
stants, that is, very nearly what can be con-
sidered conventional step tests where the
process is allowed to reach steady state af-
ter each step. In this particular example,
because the control system is expected to be
tuned so that the closed-loop bandwidth is
not significantly wider than the open-loop
one (a situation that arises frequently in
industrial control problems), the perturba-
tion signal need only emphasize low-
frequency excitation. How few step changes
(Le., the duration of the test) will suffice
depends on the presence of noise. In the
absence of noise or corrupting disturbances,
a single step (followed by a step in the
opposite direction to ensure that the system
returns to its original level) in this example
would provide the required information.
Higher-frequency information will be neces-
sary only when the system to be identified is
known to have faster modes (i.e., lower
dominant time constant) or the control sys-
tem is expected to be tuned aggressively so
that frequency response in the higher ranges
is important.
7-2-2-2 PRBS Signals for MIMO Systems
When a dynamic test is performed by per-
turbing only one input at a time, the estima-
tion problem is greatly simplified because
SISO models can be estimated one at a
time (assuming that a noise model is not
estimated). This approach has advantages
and disadvantages from an operational point
of view. By varying only one input, the be-
havior of process outputs is more pre-
dictable, reducing the probability of enter-
ing an undesirable operating region and/or
violating constraints. The probability of op-
erator intervention becomes more likely with
a test specifically designed to perturb all
inputs simultaneously (for example, with in-
dependent PRBS signals applied to each
input), because it is difficult to predict when
certain combinations of input levels might
drive a process into an undesirable operat-
ing region. This is a particular problem for
multivariable systems that exhibit strong
gain directionality, that is, the gain of a
particular output to simultaneous changes
in input levels varies widely with different
combinations of input moves.
As described by Eykhoff (1974) and
Briggs and Godfrey (1966), PRBS signals
applied to multiple inputs simultaneously
must satisfy certain conditions to allow
identification of multivariable systems. Most
important, for a system with p inputs, the
cross-correlation functions cf>uu (7") for each
I }
pair of inputs uiu
j
' i, j = 1,2, ... , p, must
be negligible for 7" less than the maximum
system settling time (this can be taken as
the largest 99% settling time among the
input-output pairs, T 99%,max)' PRBS signals
generated independently for each input
would satisfy this requirement. However, a
more practical approach is to generate mul-
tiple PRBS signals as delayed versions of a
single PRBS. To satisfy the requirement on
the cross-correlation functions, the mini-
mum allowed delay between any two ap-
plied sequences is given by D:
NTcI
D = -- > T99% max (7-10)
p ,
N can be increased as necessary to satisfy
the inequality.
Another approach to conduct a multi-
variable test is to conduct one long test in
which PRBS signals are applied consecu-
tively to the inputs, one at a time. During a
period in which a PRBS signal is applied to
one particular input, the other inputs are
held constant if possible, but allowed to
vary if necessary (e.g., due to automatic or
operator initiated control action). Data for
the entire test are collected as a single batch,
from which multiple-input-single-output
(MISO) models of the form described by
Equation 7-4 are identified for each process
output variable (that are then converted to
SISO models if required for the control
system design). Thus, the effect of one or
more inputs wandering during a period in
which another input is perturbed is ac-
counted for by the estimation procedure.
Design of perturbation signals for multi-
input-multi-output systems that are ill-
conditioned (e.g., high purity distillation
systems with composition to be controlled at
both ends of the tower) is a more complex
undertaking. These processes exhibit strong
gain directionality and the stability of any
control system (regardless of its sophistica-
tion) is very sensitive to identification er-
rors. In these situations, input perturbations
should emphasize directions in which out-
put gain is least, that is, independent PRBS
signals will not suffice. For more informa-
tion on the design of input signals for ill-
conditioned multivariable systems, the
reader is referred to Koung and MacGregor
(1991) and the references therein. Reducing
the effect of nonlinearity in identification of
linear input-output models for high purity
distillation systems is discussed in Section
7-3-6.
7-3 MODEL STRUCTURE
SELECTION AND PARAMETER
ESTIMATION
7-3-1 Bias - Variance Trade-Ofts
in System Identification
One of the most fundamental concepts in
system identification is the trade-off be-
tween bias and variance. Generally speak-
ing, errors in system identification are the
result of bias and variance:
error = bias + variance
Bias is often considered in the context of
systematic error in parameter estimates or
model predictions. However, as discussed in
more detail later, examining bias in the fre-
quency domain (i.e., the error between the
frequency response of the real system and
that of an estimated model) provides signif-
icant insight when identifying models for
control. Bias effects are present in identifi-
cation even if the data are free of noise or
the data length is arbitrarily long. Among
the factors contributing to bias are:
Choice of model structure. An incorrect
choice of model structure (or a low-order
Identification of Distillation Systems 103
structure that is an approximation of the
real system) for G(q) will lead to bias.
Frequency-domain analysis also shows
that the disturbance model H(q) affects
bias in the estimated G(q). Different
forms of H(q) (i.e., autoregressive, mov-
ing average, and the orders of the poly-
nomials) affect the frequency domain bias
of G(q) in different ways.
System excitation. The spectrum of the input
signal can significantly affect bias in the
estimated model. Frequency-domain bias
is reduced in those frequency ranges
where spectral power is strong.
Mode of operation. Whether the estimation
is carried out on data generated from
open-loop or closed-loop identification is
important. The presence of a control sys-
tem can either amplify or attenuate the
frequency content of the input signal in
certain frequency ranges, thus affecting
frequency-domain bias (that may be un-
desirable) in those ranges.
These factors are explored in more detail in
subsequent sections.
Variance in the parameter estimates and
predicted outputs, on the other hand, is
affected by such phenomena as noise in the
data, the number of model parameters, and
the duration of the experiment. Variance
effects exist in identification even when un-
biased estimation is possible.
The phenomenon of bias-variance trade-
off occurs regardless of the type of parame-
ter estimator. Analysis in the frequency
domain is an effective means for under-
standing bias-variance effects. This does not
mean that the identification has to be car-
ried out in the frequency domain, but rather
that frequency-domain arguments provide
very useful insight into the identification
problem when the end use of the model is
for process control. The information for
analysis is resident in the power spectrum of
the prediction errors, as described in suc-
ceeding text.
Consider the true system represented by
Equations 7-1 and 7-2 that is to be identi-
fied using the model in Equation (7-3). Pre-
104 Practical Distillation Control
filtering of the input and output data, the power spectrum is equivalent to the
squared magnitude
YF(t) = L(q)y(t) (7-11)
uF(t) = L(q)u(t) (7-12)
where L(q) is the specified filter, may be
performed to remove nonstationarities and
to influence the goodness-of-fit (in the fre-
quency domain), as will be described. In this
case the power spectrum <l>eF(w) of the pre-
filtered prediction error [eF(t) = L(q)ep(t)],
where the prediction error e pet) is given by
ep(t) = yet) - Y(t) (7-13)
and y is the one-step-ahead output predic-
tion
Y(t) = H-1(q)G(q)u(t)
+(1-H-
1
(q)y(t) (7-14)
is given by Ljung (1987):
<PeF(W) = (IGo(e
jW
) - G(e
jw
)1
2
<1>uC w)
1 L( e
jW
) 12
+<1>,,( w) . 2 (7-15)
IH( e
JW
) I
<I>/W) and <I>,,(w) are the power spectral
densities of the input signal u and the dis-
turbance v, respectively, computed from the
expression
<l>xCw) = L cx(k)e-jwTk (7-16)
k= -00
c x(k) is the covariance of a time series x at
lag k, evaluated using
1 N-k
cx(k) = - L x(i)x(i + k),
N i=1
k = 0, N - 1 (7-17)
When dealing with a deterministic signal,
(7-18)
It can be shown (Ljung, 1987) that the pa-
rameter estimation criterion based on mini-
mizing the sum of prediction errors (see
Equation 7-35) can be equivalently ex-
pressed in the frequency domain using
Parseval's theorem, that is,
N 1 7T
lim L e;'( i) = - f <l>ei w) dw (7-19)
N-oo i=1 -7T
From Equations 7-15 and 7-19 we can dis-
cern four factors that significantly affect fre-
quency-domain bias in the estimated G(q)
in the parameter estimation problem:
The input signal power spectral density fPJ w).
Because the input spectrum directly
weights the term IGo(e
jW
) - G(e
jw
)1
2
(i.e., the frequency-domain bias), error
between the true system and the esti-
mated model is reduced at those fre-
quencies at which the input power
spectrum is highest. Therefore, it is im-
perative to define the frequency range of
importance to the control problem and
ensure that the input signal is rich in
frequency content in this range.
Prefilter L(q). The prefiltering also acts as a
frequency dependent weight, providing
another way to selectively emphasize the
frequency range of importance for con-
trol.
The structure of G(q) and H(q). The choice
of model structure plays a significant role
in frequency-domain bias in the parame-
ter estimation problem. Note that the
noise model H(q) weights the objective
function in a manner similar to the pre-
filter, and therefore has a direct effect on
frequency-domain bias in the estimated
G(q).
The disturbance power spectral density 4>,l w).
The parameter estimation problem, when
examined in the frequency domain, is a
multiobjective minimization whenever a
disturbance model H{q) is present [be-
cause both G{q) and H{q) contain pa-
rameters to be estimated]. In this case
there is trade-off between reducing bias
between Go{q) and G{q), and fitting the
spectrum of the disturbance v{t). This
problem becomes particularly pro-
nounced when autoregressive terms that
are common to both G{q) and H{q) are
included in the parameter estimation
problem.
Variance issues can be examined in the
frequency domain using an expression of
the form (Ljung, 1987)
. (n <1>,,( w) )
vanance - - --
N <l>u( w)
(7-20)
where n = the number of parameters in
the model
N = the length of the data set
Equation 7-20 clearly shows that reducing
the number of model parameters, increasing
the length of the data set, and increasing
the magnitude of the input signal all con-
tribute to variance reduction in system iden-
tification.
7-3-2 Nonparametrlc Methods
We present nonparametric techniques be-
cause they can serve as useful precursors to
parametric estimation, providing approxi-
mate information about the plant dynamics
that can be useful later in the identification
process.
7-3-2-1 Correlation Analysis
Correlation analysis is often used to obtain
an initial guess of the process impulse re-
sponse. The kth impulse parameter Uk is
Identification of Distillation Systems 105
given by (Box and Jenkins, 1976)
where
1 N-k
cuy(k) = N E u'(i)y'(i + k),
i=1
k = 0, N-1
N is the number of observations, whereas
the prime denotes that detrending and
prewhitening operations were performed on
the time series; under these conditions,
cuy{k), the estimated cross-covariance func-
tion, can be used to obtain an estimate of
the kth impulse response of a system. Other
uses of correlation analysis include discern-
ing the deadtime of a system, the 95% set-
tling time, and also, most importantly,
determining if an input signal u has a sig-
nificant effect in describing the dynamics of
the process. This latter use involves evaluat-
ing the estimated impulse response coeffi-
cients against the hypothesis that the data
reflect only white noise. The standard devi-
ation of Uk is given by
where CT
u
and CT
y
are the standard devia-
tions of the input and output sequences,
respectively (Box and Jenkins, 1976). Two
standard deviations constitute 95% confi-
dence bounds for the estimated impulse pa-
rameters.
7-3-2-2 Spectral Analysis
The frequency response of a system is the
Fourier transform of its impulse response,
and provides information about the domi-
nant time constant and the order of the
106 Practical Distillation Control
transfer function to be estimated. The esti-
mated frequency response, G(jw) can be
obtained directly from plant data using
spectral analysis (Jenkins and Watts, 1969)
with the equation
(7-23)
where <l>UY and <l>u denote the "smoothed"
cross- and autospectra between u and y,
described by the formulas
M
<l>UY(w) = L cuy(k)w(k)e-
illlk
(7-24)
k=-M
M
<l>u(w) = L cu(k)w(k)e-
illlk
(7-25)
k=-M
w(k) is the lag "window" that accomplishes
smoothing. A common window choice is the
Tukey (Hamming) lag window
w( k) = (!(1 + cos( 'lTk/M ,
0,
lui :s; M
lul>M
(7-26)
where M is the truncation parameter for
the window.
Spectral analysis can also be applied to
the residual time series to obtain norm-
bound uncertainty descriptions that are use-
ful for robust control. A more detailed anal-
ysis of this is explained in Section 7-4.
7-33 Parametric Models
7-3-3-1 Prediction-Error Model (PEMJ
Structures
For prediction-error methods, the general
model structure Equation 7-3 is represented
using polynomials in the backward shift op-
erator q-l [defined as q-kX(t) = x(t - k):
B(q) C(q)
A(q)y(t) = F(q) u(t - nk) + D(q) e(t)
(7-27)
where A, B, C, D, and F are polynomials
in q-l:
A(q) = 1 + a q-l + .,. +a q-n.
I n.
B(q) = b
l
+ b
2
q-1 + ... +bnbq-nb+l
C(q) = 1 + c q-l + ... +C q-nc
I nc
D(q) = 1 + d1q-1 + ... +dndq-n
d
F(q) = 1 + flq-l + ... +fn,q-n,
The A polynomial is an autoregressive (AR)
term; B corresponds to the external (X)
input u, whereas C is a moving average
(MA) term. D is an autoregressive term
applied exclusively to the disturbance model.
n
k
is the system dead-time. Comparing
Equations 7-27 and 7-3 gives the transfer
functions
B(q)
G(q) = A(q)F(q) q-n
k (7-28)
C(q)
H(q) = A(q)D(q)
(7-29)
Note that the A polynomial contains poles
that are common to both the disturbance
and process transfer functions.
Equation 7-27 is written for multi-input
systems by adding terms on the right hand
side
A number of factors must be considered
in the decision to identify a disturbance
model [i.e., the polynomials C(q) and D(q),
and possibly A(q) as part of the overall
system model to be used for control of a
distillation system. For processes that are
affected by unmeasured disturbances that
can be modelled adequately by Equation
(7-29), and where the disturbance structure
is fixed (i.e., time-invariant), then inclusion
of the disturbance model into a fixed-struc-
ture controller design can be expected to
deliver superior performance over a con-
troller lacking a disturbance model. As
mentioned previously, Equation 7-29 is ca-
pable of modelling stochastic autoregressive
integrated moving average (ARIMA) distur-
bances [as described by Box and Jenkins
(1976)] or randomly occurring deterministic
disturbances. However, when performing
off-line batch identification and off-line con-
troller design, it is assumed that the distur-
bance process does not change with time
(Le., its structure is fixed). This assumption
is very often not valid in typical processing
systems. Disturbances affecting a distillation
column, as well as most other process units,
are typically random in structure. Process
units within a plant are highly intercon-
nected and integrated through process and
utility streams. A particular process unit
may be disturbed at any instant from an
upset or change in operating condition from
any number of other process units that in-
teract with it. These upsets arise from a
multitude of different sources; for example,
equipment failure, scheduled or unsched-
uled maintenance operations, abrupt
changes in ambient conditions due to pas-
sage of fronts, changes in raw material type
or composition, operator control moves mo-
tivated by any number of causes (e.g., shift
changes, optimization moves, response to
disturbances), and so on. In addition, dis-
turbances from different sources propagate
through a unit in different ways, again sug-
gesting that each will have a different struc-
ture if modelled with a linear disturbance
model in the form of Equation 7-29. Con-
trol structures that employ an identified
model for G(q) only, and ignore distur-
bance dynamics, have been found in prac-
tice to deliver acceptable control perfor-
mance.
Again, we do not consider model-based
adaptive schemes (particularly for multivari-
able constrained systems) to be viable at
this time. Thus the identification problem as
discussed in this chapter is to obtain the
best possible estimate for G(q) from the
given batch of data. Estimators that are
Identification of Distillation Systems 107
capable of giving unbiased estimates for the
parameters in G(q), even when structural
inadequacies are present in H(q), including
the case where no disturbance model is
present [Le. H(q) = 1], are particularly use-
ful in this situation. The prediction error
estimates implemented by Ljung (1988) fall
into this category.
Another consideration is frequency do-
main bias in the estimated G(q) that is
introduced by including a disturbance model
in the estimation problem. From Equations
7-15 and 7-19 it is clear that H(q) acts as a
frequency domain weight on the estimation
of G(q). Including the term D(q) [and
A(q)] emphasizes high frequencies in the
estimation of G(q); the model will more
accurately predict fast dynamics and hinder
the low-frequency and steady-state predic-
tive ability of the model. Including C(q) has
the opposite effect; high-frequency fit is de-
emphasized. These effects are a current area
of research in the literature.
Various substructures arising from Equa-
tion 7-27 and the resulting parameter esti-
mation problems are now described.
ARMAX ARMAX identification uses
the structure
A(q)y(t) = B(q)u(t - n
k
) + C(q)e(t)
(7-30)
This structure has been popular in the con-
trol literature, particularly for designing
adaptive controllers.
ARX The ARX structure is
A(q)y(t) = B(q)u(t - n
k
) + e(t) (7-31)
Using this structure, the problem of deter-
mining the right order for A and B is often
circumvented by overparametrization, that
is, the orders of the A and B polynomials
(n = na = n
b
) are selected to be high. The
theoretical justification for this choice of
variables is that for an infinite number of
observations, u white noise, and e a zero-
mean stationary sequence, the estimate of
the first n impulse response coefficients of
108 Practical Distillation Control
G(q) will be unbiased. A high-order ARX
model is therefore capable of approximating
any linear system arbitrarily well.
Although high-order ARX modelling
possesses a number of attractive theoretical
properties, the conditions under which they
apply are seldom seen in practice. Further-
more the presence of the autoregressive
term A(q) results in an emphasis on the
high-frequency fit, which is not necessarily
good for control system design. The bias
problem with autoregressive terms becomes
worse if low-order ARX models are used.
FIR The FIR (finite impulse response)
model structure
yet) = B(q)u(t - n
k
) + e(t) (7-32)
is convenient because some predictive con-
trollers are based on this model structure
[and it is easily converted to step-response
form for others that require it, e.g., dynamic
matrix control (DMC)]. Although the ap-
propriate order of the finite impulse re-
sponse depends on the selected sampling
time and the settling time of the process,
the result is usually high (30 to 100). When
u is persistently exciting, e is stationary, and
u and e are uncorrelated, the estimated
impulse response coefficients will be unbi-
ased. Again, as with ARX, these conditions
are seldom fully observed in practice. As
mentioned in the introduction, direct esti-
mation of the impulse response model is
inefficient, resulting in parameter estimates
with high variance. Direct identification of
the FIR model requires special test and
estimation procedures, as described in sub-
sequent text.
Box-Jenkins Box and Jenkins (1976)
proposed a model of the form
B(q) C(q)
yet) = F(q) u(t - nk) + D(q) e(t)
it separately parametrizes the input and dis-
turbances, avoiding transfer functions that
have common poles. The estimation prob-
lem is nonlinear, but reliable and fast esti-
mation methods are available. Nonstation-
ary disturbances are handled by forcing
D(q) to have one or more roots on the unit
circle [Le., write as (1 - q-l)dD(q), where
d is typically equal to 1].
Output E"or The output error model is
a simplified form of the Box-Jenkins model
structure
B(q)
y(t) = F(q) u(t - nk) + e(t) (7-34)
The output error structure retains the ad-
vantage of separate parametrizations for the
input and disturbance, and does not require
a choice of the structure for the disturbance
model.
7-3-3-2 Prediction-Error Parameter
Estimation
The objective minimized in prediction-error
methods is the sum of the squared predic-
tion errors
1 N 1
V= - E -(yet) - .9(t)2
N t=1 2
1 N 1
= - E -eJ,(t)
N t=1 2
(7-35)
In the case of FIR and ARX models, the
predictor structure is
.9(t) =B(q)u(t) + (1-A(qy(t)
(7-36)
which can be written in regression form as
(7-37)
(7-33) lP is the regression vector
One advantage of this structure over the
ARMAX structure in Equation 7-30 is that
cp=[-y(t-l)
u( t - 1)
-yet - na)
u(t-n
b
)]
and () is the vector of parameters to be
estimated
Because the prediction equation is linear in
the parameters, estimates are obtained in
the usual manner for linear least-squares
problems:
[
iN ]-1 1 N
8 = - 1: (fJ(t)(fJT(t) - 1: (fJ(t)y(t)
N t =l N t =l
(7-38)
For the ARMAX structure, the predictor
structure becomes
9(t) =B(q)u(t) + (l-A(qy(t)
+(C(q) - l)(y(t) - 9(t
(7-39)
In this case the regression vector is
(fJ=[-y(t-1)
u( t - 1)
ep(t - 1)
and the parameter vector is
-y(t - na)
u(t - n
b
)
ep(t - n
c
)]
For ARMAX estimation, the regression
problem is no longer linear with respect to
the parameters and a simple least-squares
solution is no longer applicable. Again, non-
linear least-squares algorithms are readily
available for solution of this estimation
problem.
Output error and Box-Jenkins estima-
tion are also nonlinear least-squares prob-
lems. A summary of explicit search tech-
niques (e .g., Newton - R aphson,
Levenberg-Marquardt, Gauss-Newton) is
found in Ljung (1987). Although these tech-
niques have favorable theoretical properties
(i.e., separate parametrization of input vs.
Identification of Distillation Systems 109
disturbance effects), numerically they repre-
sent more challenging problems than their
linear counterparts. The possibility of lack
of convergence or multiple local minima
must be considered when using these meth-
ods.
Estimation algorithms for all of the pre-
ceding model structures, including the case
where there are multiple inputs, are avail-
able in commercial identification packages.
73-4 Identification for Control System
Design
From the previous discussion it is clear that
a myriad of design variables exist in system
identification. Consequently, it is important
to determine how these design variables can
be manipulated for the purpose of identifi-
cation for control design. In this section,
our objective is to present some recent tech-
niques that have been described in the liter-
ature that consider this problem.
7-3-4-1 Use of Auxiliary Information
in Identification
The trade-off between performance and ro-
bustness of a control system in the presence
of model uncertainty can be analyzed in the
frequency domain using the Nyquist stabil-
ity criterion (Morari and Zafirou, 1989). For
robust stability, each member of the family
of Nyquist curves resulting from a charac-
terization of the model uncertainty (see
Chapter 2 of Morari and Zafiriou for a
more detailed discussion) must satisfy the
Nyquist criterion. For a given level of model
uncertainty, a controller must be detuned so
that the Nyquist curve corresponding to the
worse case model does not encircle the point
( -1,0). Thus higher model uncertainty
leads to poorer performance. Conversely,
less uncertainty in the fitted model, particu-
larly around the frequency where the
Nyquist curve most closely approaches the
point ( -1, 0), allows tighter controller tun-
ing and thus a greater closed-loop band-
width and better control performance.
Therefore, reducing the bias (and uncer-
110 Practical Distillation Control
tainty) of the fitted model around this fre-
quency is an important objective when the
model is to be used for control system de-
sign.
From previous discussions, one way to
improve the accuracy of the fitted process
model G(q) in a particular frequency range
is to ensure that the perturbation signal is
rich in these frequencies. Designing a per-
turbation signal to emphasize system excita-
tion around the frequency where the Nyquist
curve most closely approaches the point
( -1,0) poses a problem because a priori
knowledge of the closed-loop frequency re-
sponse for a given controller tuning would
be required. Although an iterative design
procedure could be adopted, an alternative
is the use of a simple relay test (Astrom and
Hagglund, 1984) that determines the point
of intersection of the Nyquist curve with the
negative real axis of the open-loop system
Go(q) (i.e., the critical point, or the ultimate
gain and frequency). In an industrial setting,
a typically tuned PID controller (or other
controller design) provides only a small
amount of phase lead, and the frequency at
which the Nyquist curve most closely ap-
proaches the point (-1,0) (as well as the
crossover frequency of the compensated sys-
tem, of equal importance) is not often not
far from the region of the critical point of
the open-loop process. In these situations,
knowledge of the open-loop critical point is
of practical value in system identification
for control purposes.
Astrom and Hagglund used the relay test
to provide information for PID controller
tuning. The use of such auxiliary. informa-
tion (i.e., knowledge of one or more points
on the system Nyquist curve, obtained from
experiment or from other process knowl-
edge) to improve the quality of models iden-
tified for control purposes has been de-
scribed by Wei, Eskinat, and Luyben (1991)
and Eskinat, Johnson, and Luyben (1991).
The basic idea is as follows. Data are col-
lected from the process to be identified
using a conventional perturbation proce-
dure (e.g., step or PRBS tests). A model of
specified structure is fit to these data, sub-
ject to the constraint that the Nyquist curve
of the fitted model passes exactly through
the known point(s) on the Nyquist diagram,
thus reducing the bias in the fitted model at
this frequency to zero. As shown in the
preceding references, the constraint can be
expressed as a set of linear equations, re-
sulting in a quadratic optimization problem
when the objective function for fitting the
model is quadratic (as is the case with pre-
diction-error methods). For certain model
structures, analytical solutions to the
quadratic optimization are available; other-
wise a numerical procedure must be used.
To perform a relay test to obtain auxil-
iary information for fitting a discrete time
model, a relay is connected in a sampled
feedback loop, as shown in Figure 7-2. The
sampling frequency for the relay test must
be the same as the sampling period used to
collect the data from the perturbation test.
If the process has a phase lag of at least - 7T'
radians (which will always be the case in
practice because all real processes exhibit
some deadtime), under relay feedback it
will enter a sustained oscillation with period
P
u
(the critical period), giving the ultimate
frequency Wu = 27T'/P
u
The ultimate gain is
given approximately by Ku = 4h/(a7T'),
where h is the height of relay and a is the
amplitude of the principal harmonic of the
output. The relay test is easier and safer to
apply than the procedure suggested by
Ziegler and Nichols because it does not
require adjustment of controller gain in or-
der to generate sustained oscillation. It only
requires approximate knowledge of the pro-
+
SeIpaInt
FIGURE 72. Configuration for a relay test to esti-
mate crossover frequency for use as auxiliary informa-
tion in identifying a discrete-time model.
cess gain (ideally at the critical frequency,
but because this is not likely available,
steady-state gain can be used) in order to
specify an appropriate relay amplitude.
Because the relay is a nonlinear device,
when applied to a linear process the mea-
sured critical point is an approximation to
the real one. However, as shown by Wei,
Eskinat, and Luyben (1991), the informa-
tion is sufficiently accurate to be of value in
system identification for control purposes.
The presence of noise may require the
relay test to be implemented with a dead
band within which the response variable can
move without causing a jump in level of the
input. In a very noisy or highly disturbed
process, it may be difficult to obtain useful
results from a relay test.
7-3-4-2 Control-Relevant Prejiitering
The previous section described a procedure
for reducing the impact of modelling bias
on control design by emphasizing the fit
near one frequency, namely, the crossover
frequency. Prefiltering, on the other hand,
allows the engineer to emphasize the good-
ness of fit over a range of frequencies, and
thus is a useful technique for identifying
models that are not only to be used for
feedback control but for other control struc-
tures, such as combined feedback-feedfor-
ward control, decentralized control, and
cascade control (Rivera, 1991a).
As described in Section 7-3-1, the pre-
filter L(q) acts as a frequency-dependent
weight in the parameter estimation prob-
lem, and as such can be used to emphasize
those frequency regions that are most im-
portant for control system design. Assuming
that a sufficiently informative dynamic test
has been performed, a prefilter can be used
to arbitrarily influence the goodness-of-fit in
the frequency domain without having to
modify the model structure or repeat the
dynamic test. For example, prefiltering can
be used to obtain a good steady-state fit
without requiring a step test input (provided
that sufficient low-frequency information is
available from the test data).
Identification of Distillation Systems 111
The literature remains vague about pre-
filter design; questions regarding the sys-
tematic choice of structure (low pass, high
pass, bandpass) and the choice of filter pa-
rameters remain unanswered issues. How-
ever, recent work in this area has begun to
provide some answers. In Rivera et aI., 1990
and Rivera, Pollard, and Garcia, 1990, the
derivation of a control-relevant prefilter for
SISO feedback control is described that ex-
plicitly incorporates the model structure, the
desired closed-loop response, and the set-
point-disturbance characteristics of the
control problem into the statement of the
prefilter:
L(q) = H(q)G-
1
(q)(1 - ij(q
Xij(q)(r(q) - d(q (7-40)
ij = GK(1 + GK)-l is the complementary
sensitivity function that describes the de-
sired closed-loop response. Two algorithms
for implementation of the prefilter, one it-
erative and the other "single-pass", are de-
scribed by Rivera et al. (1990) and Rivera,
Pollard, and Garcia (1990). The single-pass
implementation is described here.
The statement of the single-pass prefilter
is obtained from Equation 7-40 following
some assumptions and simplifications. This
prefilter assumes that the essential dynam-
ics of the system can be reasonably approxi-
mated by a first-order with deadtime model
where a = e-T,/Tdom and Tdom is a user-pro-
vided estimate of the dominant time con-
stant of the system. An estimate of the
steady-state gain is not necessary because
the gain simply appears as a constant in
Equation 7-40.
Regarding the structure for ij, one must
recognize that it is usually dictated by the
control system design procedure. In
quadratic dynamic matrix control (QDMC),
for example, the effect of move suppression
112 Practical Distillation Control
results in fi(q) that is an order greater than
the plant (Balhoff and Lau, 1985). In in-
ternal model control (IMC) (Morari and
Zafiriou, 1989), fi is specified directly via
the factorization of nonminimum phase ele-
ments and the choice of a filter. In accor-
dance with this philosophy for the model in
Equation 7-41 we can use the structure
f(q) is a low-pass filter that either corre-
sponds to the filter used in the IMC design
procedure or approximates the effect of
move suppression in QDMC. For model
structures such as output error (DE) and
finite impulse response (FIR), H(q) = 1,
which leads to the definition of the prefilter
as
L(q) = (q - a)(l - q-nkf(q))
xq-lf(q)(r(q) - d(q (7-43)
For autoregressive with external input
(ARX) models, we can approximate jJ e with
the same dominant time constant guess
made for jJ:
q
H(q) "" ( ) (7-44)
q-a
which leads to
L(q) = (1 - q-nkf(qf(q)
x(r(q) - d(q (7-45)
Specific forms for these prefilters are
shown via an example. Consider the first-
order plant model according to
k
G(q) = -- (7-46)
q-a
which will be controlled using QDMC. The
resulting structure for ii is second order, so
we define f(q) as second order:
(7-47)
where l) = e - 1.555T, / 'T cI
Tel = the desired closed-loop time
constant, specified by the engi-
neer
Assuming step setpoint and disturbance
changes for the control system, the resulting
prefilter for FIR and DE estimation is fourth
order,
(7-48)
and (for the ARX structure)
(7-49)
Note how Equations 7-48 and 7-49 meet the
requirements for a practical prefilter. Hav-
ing defined a model structure and the na-
ture of the controller design problem, the
choice of the prefilter is reduced to simply
providing estimates for the closed-loop
speed-of-response (Tel) and the open-loop
time constant (Tdom)' This information can
be readily obtained in most situations faced
by process control engineers.
Restrictions regarding the benefits of
prefiltering can be gathered through an un-
derstanding of bias-variance trade-offs, as
noted in Rivera (1990a). In output error
estimation, control-relevant prefiltering was
shown to substantially improve the estima-
tion for low-order models. Prefiltering in
the ARX estimation problem, however,
worsens the quality of the plant estimate if
the ratio <I>,,(w)/<I>u(w) is high over the
bandwidth of the prefilter because this
causes the estimator to favor the fit of A(q)
to noise. For the case of FIR, a significant
reduction in the low-frequency uncertainty
is obtained using prefiltering; variance ef-
fects, however, become more pronounced,
and the resulting step responses may look
increasingly more jagged. One way to over-
come this variance problem is to estimate
the FIR parameters using a biased least-
squares estimator, as discussed in Section
7-3-4-3.
Work is currently being carried out
(Rivera, 1991b) that unifies the control-rele-
vant prefiltering approach with the crossover
frequency approach described in Section
7-3-4-1. A fundamental analysis of mod-
elling requirements for control shows that
there are four factors that determine the
adequacy of a model for feedback control
purposes:
1. The choice of plant and noise model
structure.
2. The desired closed-loop speed-of-
response.
3. The shape of the closed-loop response
(Le., underdamped or overdamped).
4. The nature of the setpoints and distur-
bances that will be encountered by the
closed-loop system.
Under certain circumstances (the presence
of an integrator in the model and control
requirements demanding very fast, under-
damped responses, for example), the power
spectrum of the prefilter shows a very strong
amplification of the crossover frequency, in
which case both approaches yield equivalent
results. However, when considering the case
of a first-order system with substantial
deadtime (ratio of time delay to time con-
stant equal to 1), the requirement of an
overdamped closed-loop response with a
correspondingly smooth manipulated vari-
able response results in modelling require-
ments that emphasize the intermediate and
final segments of the plant's step response
without demanding a close fit of the plant's
ultimate frequency. This occurs because for
the specified control requirements, the fre-
quency responses of the plant and the loop
transfer function Go{s)K(s) differ substan-
tially at the crossover point; the assumption
that the controller does not introduce sub-
stantial phase shift (a key assumption for
Identification of Distillation Systems 113
the method described in Section 7-3-4-1) is
violated in this case.
7343 Biased LeastSqUllTeS Estimators
Direct estimation of a truncated impulse
response model results in an estimation in
which the bias-variance trade-off is shifted
toward reduction in bias at the expense of
variance, due to the large number of param-
eters (and their lack of independence).
Therefore, an estimated impulse model
might give an excellent fit to the data that
were used in the estimation, but its use as a
prediction tool on a different set of data
may be extremely limited because of the
resulting high variance of the parameter
estimates (and thus predictions). There are
a number of ways to reduce the variance of
estimates to improve the predictive ability
of models:
1. Estimate a parsimonious transfer func-
tion model (low number of parameters)
and convert the nominal estimated model
to impulse or step form (with no loss of
information).
2. Perform a longer dynamic test to in-
crease the size of the data set, (which, as
mentioned previously, may be quite ac-
ceptable for some processes).
3. Use a biased estimator that shifts
bias-variance trade-off toward allowing
some bias in order to reduce variance.
Biased estimators, as the name implies,
produce estimates that are biased and not
least squares, but are more desirable from a
prediction point of view because the reduc-
tion in prediction mean square error (MSE)
they achieve more than compensates for the
minimal bias that is introduced. We briefly
describe a few approaches that have been
used in the literature for biased estimation
of truncated impulse response models and
we provide references for further informa-
tion.
Ridge Regression Consider the finite im-
pulse response (FIR) model Equation 7-32.
The prediction equation for this model
114 Practical Distillation Control
structure is given by Equation 7-36 [with
A(q) = 1], and the least-squares parameter
estimates are given by Equation 7-38, which
can be written in matrix notation as
8 = [X
T
Xr
1
X
T
y (7-50)
With a typical order of 30 to 100 for
B(q), the parameters will not be indepen-
dent, resulting in poor conditioning of XTX
and an inversion that is sensitive to small
variations in the data. In the simplest form
of ridge regression, Equation 7-50 is re-
placed by
where I = an identity matrix of appropri-
ate dimension and
c = a specified positive number
For more information, the reader is re-
ferred to Draper and Smith (1966) and
Hoerl and Kennard (1970).
Partial Least-Squares Estimation An-
other biased estimator that has been used
to estimate FIR models (Ricker, 1988) is
partial least squares (PLS), a general method
for solving poorly conditioned least-squares
problems. A brief explanation of the method
follows. For more information, the reader is
referred to Ricker (1988), Lorber, Wangen,
and Kowalski (1987), Hoskuldsson (1988),
Geladi (1988), and a tutorial paper by Geladi
and Kowalski (1986).
A data matrix X (dim m x n), can al-
ways be decomposed into two matrices, t
(m x r) (the "scores") and P (n x r) (the
"loadings"):
X
- - t pT + t pT + ... + t pT
-.1., -1122 rr
(7-52)
where r = rank(X)
t., p. = are the ith columns of t and
I I ....
P, respectively
Because of the correlation that exists be-
tween the parameters in an FIR model, not
all of the ti and Pi are required to explain
the total variation that exists in the data
matrix X. For the same reason, not all the
ti and Pi are required to develop a satisfac-
tory predictive relationship between X and
the response vector y. Therefore, in PLS, X
is written as
X= t
1
P[ + t
2
pr + ... + E
= TpT +E
where q is referred to as the number of
latent variables, and is usually very much
less than r. E is a residual matrix that
contains the random errors in X as well as
the discarded components of X that have
no significant predictable effect on y. In
PLS, the regression is performed between y
and the ti in Equation 7-53 and is an inte-
gral part of the numerical procedure that
performs the decomposition (in the general
PLS problem where there are multiple re-
sponse variables that may be correlated, the
numerical procedure includes decomposi-
tion of the response matrix Y; see for exam-
ple, Kresta, MacGregor, and Marlin, 1991).
PLS requires the user to select the num-
ber of latent variables used in the decompo-
sition. The object in PLS is to find the
minimum number of latent variables that
results in a satisfactory predictor of the re-
sponse variable. This involves evaluating the
predictive ability of fitted models with in-
creasing number of latent variables, usually
using cross-validation (the model is fit on
one segment of data, and evaluated for pre-
dictive ability on another segment). The
number of latent variables is increased until
no further significant decrease in the sum of
squares of residuals between the predictions
from the fitted model and the validation
data set is observed. The number of latent
variables will be far less than the number of
FIR parameters in a typical FIR model.
With fewer parameters to estimate, a more
acceptable balance between bias and vari-
ance in the estimated model is obtained. As
pointed out by Rivera et aI., 1990 and
Rivera, Pollard, and Garcia (1990), pre-
filtering prior to PLS estimation gives the
engineer a further handle on the bias-vari-
ance trade-off in a frequency-dependent
sense.
7-3-5 Identifiability Conditions for
Closed-Loop Systems
When data are collected with a process in
closed-loop operation, certain identifiability
conditions must be satisfied to ensure that
proper identification can be performed.
Historical plant data collected under closed-
loop conditions do not generally satisfy these
conditions and cannot be used for dynamic
model identification. The pitfalls associated
with ignoring these requirements have been
well documented in the literature (e.g., Box
and MacGregor, 1974). Two issues are in-
volved:
Identifiability. Identifiability refers to
whether a suitable model describing the
plant dynamics can be extracted from the
available closed-loop data. When specific
identifiability conditions are satisfied (for
example, see Ljung, 1987, Chapter 8,
specifically Theorem 8.3) prediction-
error estimators can be applied to
closed-loop data as though they were ob-
tained under open-loop conditions, and
no other special consideration is required
(except that the model to be identified be
sufficiently parametrized that it is capa-
ble of representing the true system dy-
namics).
Closed-loop frequency-domain bias. Even if
identifiable, the quality of the estimated
model obtained from closed-loop data is
an issue (discussed in Section 7-3-5-2).
7-3-5-1 Identifiability
Soderstrom, Ljung, and Gustavsson (1976)
and Gustavsson, Ljung, and Soderstrom
(1977) provide detailed analyses of the
closed-loop identifiability problem. Depend-
ing on the conditions imposed during the
Identification of Distillation Systems 115
collection of closed-loop data, two levels of
identifiability can be achieved:
System identifiability (SI). Given that certain
strict conditions are met with regard to
the model structure and the order of the
compensator, model parameter estimates
that correctly represent the plant can po-
tentially be obtained from the data.
Strong system identifiability (SSI). This means
that no special restrictions need be im-
posed on the model structure or the
order of the compensator in order to
obtain reasonable estimates. Frequency-
domain bias remains as an issue, how-
ever.
It is clearly desirable to obtain experimental
conditions that are SSI because regression
techniques can be applied in a direct fash-
ion with no other special considerations.
Conditions that lead to SSI are
Closed-loop data generated by shifting be-
tween different feedback controllers.
Closed-loop data generated by adding an
external signal to the feedback loop that is
persistently exciting (typically a PRBS sig-
nal). Possible injection points for this sig-
nal include the controlled variable set-
point (r) and the manipulated variable
(u).
From a process operations viewpoint, apply-
ing an external signal to the feedback loop
is preferable to switching between regula-
tors.
One important result of the work of
Soderstrom, Ljung, and Gustavsson and
Gustavsson, Ljung, and Soderstrom is that
data are collected from a closed-loop sys-
tem responding strictly to disturbances is SI
but not SSI. Only if certain restrictive con-
ditions are met with regard to the plant and
the compensator will useful models be ob-
tained from data. These conditions are
unattainable from a practical standpoint,
however, because they depend upon a priori
knowledge of the true plant structure; hence
it is not possible to determine a posteriori if
116 Practical Distillation Control
the system was really identifiable. One is
better off carrying out closed-loop identifi-
cation under circumstances that generate
SSI, which implies that a carefully designed
external signal should be used.
It should be noted that nonparametric
identification techniques (such as correla-
tion analysis and spectral analysis) cannot
be applied directly to closed-loop data from
normal plant operation. This has been pre-
viously pointed out by Box and MacGregor
(1974), among others. The standard applica-
tion of these techniques results in model
estimates that are the inverse of the con-
troller.
7-3-5-2 Closed-Loop Frequency Domain Bias
An expression similar to Equation 7-15 can
be derived for closed-loop systems to ana-
lyze frequency-domain bias when a regula-
tor is present. Such an analysis is more
complicated than its open-loop counterpart.
We will thus focus on the effect of the
feedback controller C on the power spec-
trum of the input signal, in order to provide
some insight into the problem of
frequency-domain bias generated from
closed-loop dynamic tests. If an external
signal U
d
is added onto the manipulated
variable U, the power spectrum of the input
is given by
where 13 = (1 + GOK)-l is the sensitivity
function of the closed-loop system that acts
as a weight to the parameter estimation
problem in a manner similar to prefiltering.
The effect of the controller is to attenuate
low-frequency information, as shown in Ex-
ample 7-1. For controllers with integral ac-
tion, the sensitivity function is 0 at w = 0,
which means that the controller attenuates
the low-frequency portion of the external
signal. Significant detuning of K, the feed-
back controller, may be required in order to
obtain an appropriate steady-state fit.
It is possible to avoid detuning by intro-
ducing the signal at an alternate point in
the loop, the controlled variable setpoint.
The expression for the input power spec-
trum in this case is
where 11 = G
o
K(1 + GOK)-l is the com-
plementary sensitivity function. Note that
the effect of an external setpoint change on
the estimation of G(q) is weighted by
IG
o
l
1112, as opposed to 1131
2
, which does not
attenuate the low frequencies when the
controller is tightly tuned. Furthermore, a
tightly tuned controller may also lead to
amplification of the higher frequencies, as
Example 7-1 shows.
Example 7-1
Consider the first-order plant given by:
k 0.096
G (q) - -- - (755)
o - q _ a - q - 0.904 -
This plant corresponds to a first-order
transfer function with an open-loop time
constant of 10 min. Assume that the desired
closed-loop response is also a first-order
transfer function, represented by the ex-
pression
where
(1 - 8)
11(q) = 1 - e(q) = -q---8-
(7-56)
1'. is the sampling time (and control inter-
val) and 'Tel is the closed-loop time constant.
A PI controller tuned using the rules de-
scribed by Prett and Garcia (1988) would
result in such a closed-loop system.
In this example we set r. = 1.0 min and
examine the amplitude ratios of G
0
1
1/ and
8 for the three cases:
Tel = 1 min (closed-loop speed faster than
open-loop).
Tel = 10 min (closed-loop speed equivalent
to open-loop).
Tel = 100 min (closed-loop speed slower than
open-loop).
Figures 7-3 and 7-4 compare the normalized
amplitude ratios of Go 11/ and 8. 8 resem-
bles a high-pass filter with bandwidth (the
frequency at which the amplitude ratio first
reaches 1/ {i) defined by l/Tel' G
0
1
1/ is a
low-pass filter for slow closed-loop speed-
of-response; however, its ability to attenuate
the high-frequency range decreases with in-
Identification of Distillation Systems 117
creasing 'Tel' and when 'Tel < 'Tal' the high
frequencies are amplified. In this example,
G
0
1
1/ is an allpass when Tel = 'Tal' which
represents an ideal situation. We thus rec-
ommend that when injecting an external
signal at the controlled variable setpoint,
the controller should be tuned such that the
closed-loop speed of response equals the
open-loop speed of response.
7-3-6 Treatment of Nonlinearity
Distillation columns are inherently nonlin-
ear. Severe nonlinearity is most often seen
in high purity columns, but a number of
other factors can contribute as well, for
example, non ideal vapor-liquid equilib-
rium. Although columns operated above ap-
proximately 98% purity often exhibit severe
nonlinearity, behavior varies from column
to column and it is difficult to define a
specific purity above which severe nonlin-
10
0
______ ""' __ :::c _____________________________________ _
.g

..
"0
.e
:=
Q.

10-1
Frequency (Radians/Minute)
FIGURE 7-3. Amplitude ratio for G
O
l
1/. -: 7'el = 1 min; - - -: Tel = 10 min;
-'-: 7'e1 = 100 min.
118 Practical Distillation Control
Frequency (RadiaDI/MiDute)
FIGURE 7-4. Amplitude ratio for E. -: Tel = 1 min; - - -: Tel = 10 min; -'-:
Tel = 100 min.
earity can be expected. As with other non-
linear systems, if operated over a sufficiently
small region, linear models often provide an
adequate system description. Although large
setpoint changes are uncommon in distilla-
tion systems, disturbances often drive re-
sponse variables far enough from steady-
state conditions that the performance of
linear model-based control systems can de-
teriorate significantly. This section presents
a number of tools for adapting the basic
linear identification approach when signifi-
cant nonlinearity is present in distillation
systems.
The most widely applied tool for identify-
ing nonlinear distillation systems is the use
of linearizing transformations on composi-
tion measurements (Alsop and Edgar, 1987;
Koung and Harris, 1987). The Koung and
Harris transformation is based on an analy-
sis of the fundamental performance equa-
tions described by Eduljee (1975), which are
particularly accurate for columns operated
near minimum reflux. The transformed
composition measurements are given by
_ 1 - X
D
X
D
= In t (7-57)

- XB
X
B
= In ---;et (7-58)
X
B
where x = composition measurement
set = setpoint
Georgiou, Georgakis, and Luyben (1988)
applied dynamic matrix control (DMC)
based on transformed composition mea-
surements to three simulated distillation
systems: a moderate purity methanol-water
system (99% purity), the same system at
higher purity (99.9%), and a very high pu-
rity system (10 ppm impurity). As purity
increased, significant performance improve-
ments were obtained by using nonlinear
DMC (i.e., variables transformed) over lin-
ear DMC and a conventional LV control
strategy using multiple PI loops. However,
with the very high purity column, the per-
formance of the conventional PI-based con-
trol system was superior to that of the non-
linear DMC, suggesting that the simple
Koung and Harris transformation was in-
adequate at this purity level. The more
complex Alsop and Edgar transformations
were not investigated in the Georgiou,
Georgakis, and Luyben study.
A number of different identification
approaches were used in the Georgiou,
Georgakis, and Luyben study. Linear trans-
fer function models were readily obtained
for the moderate purity column using con-
ventional pulse testing techniques (Luyben,
1990). However, at higher purities, nonlin-
earities make it difficult to obtain a suitable
linear transfer function model. Estimated
transfer function parameters can vary
widely, depending on the size and direction
of the perturbations to the inputs. Luyben
(1987) proposed an identification procedure
to overcome this difficulty in high purity
distillation systems. In the proposed proce-
dure, steady-state gains are obtained from
detailed and highly accurate steady-state
simulation (rating) programs. Relay tests (as
described in Section 7-3-4-2) are then ap-
plied to the actual system to obtain the
critical gains and frequencies for the diago-
nal elements of the plant transfer function
matrix. Individual transfer functions are ob-
tained by finding the best transfer function
fit to the zero frequency and ultimate fre-
quency data. The utility of this identifica-
tion procedure for very high purity columns
when combined with linearizing transforma-
tions was demonstrated by Georgiou,
Georgakis, and Luyben (1988).
As discussed in Chapter 12, it is difficult
to predict in advance whether a linearizing
transformation will be helpful for a given
column. A useful procedure for testing the
value of a linearizing transformation for a
particular column is to put competing mod-
els on line and compare the variance and
autocorrelation structure of their prediction
errors over an extended period.
Identification of Distillation Systems 119
Further information on the use of lin-
earizing transformations can be found in
Shinskey (1988).
74 MODEL VALIDATION
7-4-1 Classical Techniques
Having performed parameter estimation,
the remaining problem in identification is to
assess the validity of the model. As noted
previously, the issue of importance here is
whether the model meets the intended pur-
pose, which is control system design. We
briefly describe some of the commonly used
validation techniques.
Simulation Simulation of the measured
(y) versus predicted (9) output is the most
common validation tool in identification.
Adequately determining how "close" the
predicted output is to the actual output,
however, is a more challenging task than it
appears.
Impulse or Step Responses Examining
the model's response to a standard input
such as a step or a ramp is a commonly used
technique that allows an engineer to com-
pare the results of identification versus per-
sonal process understanding. A model with
wrong gain sign or one with dynamic behav-
ior that contradicts the engineer's own intu-
itive understanding of the process will result
in repeating some step(s) of the identifica-
tion procedure.
Cross- Validation Generally speaking, if
a model is validated using the same data
that were used for parameter estimation,
the fit can be expected to improve with
increasing number of parameters in the
model. To judge the quality of a fitted model
in this circumstance requires a measure that
accounts for both the decrease in the fit's
loss function and the potential loss in pre-
dictive power of a model with increasing
number of parameters. One commonly used
measure is the Akaike information theoretic
criterion (AIC) (see Ljung, 1987, Section
16.4). An alternate approach is to perform
cross-validation, in which a portion of the
120 Practical Distillation Control
experimental data set is retained for valida-
tion purposes and not used in parameter
estimation. Simulating the model output
over the cross-validation data set allows an
engineer to assess the true predictive ability
of the model.
Residual Analysis Residual analysis con-
sists of testing the residual time series
against the hypotheses that the noise series
is Gaussian and uncorrelated with the input
series. If a nonunity disturbance model is
chosen, that is, H(q) * 1, then residual
analysis is performed on the calculated pre-
diction errors. Residual analysis is discussed
in detail by Box and Jenkins (1976). The
measures that are used in residual analysis
are the calculated autocorrelation
(
k) = cep(k)
Pep 2 (7-59)
(Fep
and for the cross-correlation between resid-
ual and the plant input
If the auto- or cross-correlation functions
exhibit a recognizable pattern (i.e., autocor-
relation at certain lags fall outside their
estimated 95% confidence intervals), model
inadequacy is suggested. If examination of
the cross-correlation function(s) suggests no
inadequacy of the process transfer function
G(q) (i.e., no cross-correlations outside their
estimated 95% confidence limits), then the
inadequacy is probably in the disturbance
model. An overestimated deadtime (nk too
large) shows up as significant cross-correla-
tion between residuals and the input as lags
corresponding the missing b k terms in the
polynomial B(q).
7-4-2 Control-Relevant Techniques
A fundamental problem with the model val-
idation techniques described in the previous
section is that they are open-loop measures
for the goodness-of-fit. An increased under-
standing of modelling for control design has
led to the conclusion that a model that may
appear to be a good fit in the open-loop can
result in very bad closed-loop performance
and vice versa (see ;\'str6m and Witten-
mark, 1989, Chapter 2). Clearly, there is a
need for control-relevant validation mea-
sures that determine the adequacy of the
model for control purposes. To this end,
current research draws from robust control
theory and the structured singular value
paradigm to develop these validation mea-
sures (see Rivera, Webb, and Morari, 1987
and Smith and Doyle, 1989). The key chal-
lenge here, however, is obtaining appropri-
ate estimates of the plant uncertainty from
the residual time series.
The most conventional means for obtain-
ing model uncertainty is to use 100(1 - a)%
confidence intervals generated from the so-
lution of the linear least-squares problem
Equation 7-38. These confidence intervals
can be computed from the identification
data set and an arbitrary plant input a (such
as a step or an impulse) as follows:
[
IN ]-1
aT -E<p(t)<pT(t) a
N 1=1
(7-61)
t(N - n, 1 - a/2) is the t statistic obtained
from Student's distribution. These confi-
dence intervals must be viewed with caution
because they imply that only Gaussian noise
affects the process. Neither the effects of
undermodelling nor the problems associ-
ated with autocorrelated residuals are con-
sidered.
An alternative is to compute frequency-
domain uncertainty descriptions from the
residual time series, as shown by Kosut
(1987). Kosut uses spectral analysis to ob-
tain an estimate of unmodelled dynamics.
This result is applied by Rivera et al. (1990)
and Rivera, Pollard, and Garcia (1990) to
generate uncertainty bounds that capture
both bias and variance effects in the identi-
fication. A brief summary follows.
Consider plant residuals represented by
the symbol v, which reflect both unmod-
elled dynamics and the effect of noise 1I(t):
v(t) = (Go(q) - G(q)u(t) + lI(t)
= a(q)u(t) + lI(t) (7-62)
Estimating uncertainty, then, requires ob-
taining an estimate of the unmodelled dy-
namics Mq) and realizing that the quality
of this estimate is affected by the presence
of noise and length of the data set. The
additive uncertainty bound I a is obtained
from
(7-63)
where the first term is a bias term whereas
the second is a variance term. To obtain the
bias term, one uses spectral analysis to ob-
tain .1:
where <l>vu and <l>u denote the "smoothed"
estimates of the cross- and autospectra
between v and u. A result presented by
Jenkins and Watts (1969) is then used to
obtain the variance term
Identification of Distillation Systems 121
where <1>" = an estimate of the distur-
bance power spectral density
<1>,,( w) = <l>v( w) - 1!I2<1>i w)
<l>v(w) = computed power spectrum
for v
111 = degrees of freedom of the
spectral estimator (which is
proportional to the length of
the data set)
12117-2 = the two-way Fisher statistic
, for a user-specified confi-
dence level.
These results are used by Rivera et al.
and Rivera, Pollard, and Garcia (1990) to
analyze the effects of different design vari-
ables on diverse prediction-error model
structures.
The usefulness of uncertainty modelling
from residuals for robust control purposes
must be clarified. To begin with, both the
confidence interval and frequency domain
approaches assume that the true plant is
linear. Hence their usefulness when applied
to nonlinear plants must be qualified, as is
done by Webb, Budman, and Morari (1989).
Because of operating restrictions in identi-
fication testing, however, many data sets
gathered from process plants fall under this
assumption. Thus the appropriate question
to ask in evaluating these approaches is,
even if the estimated uncertainty does not
cover all true plant uncertainty, is this infor-
mation still useful for controller design?
The answer here is yes. Performing control
over varying operating regions is a task that
requires a careful combination of robust
control design, process monitoring tools, and
adaptation mechanisms. The usefulness of
uncertainty modelling from residuals must
be viewed within this context.
7-5 PRACTICAL
CONSIDERATIONS
In the previous sections we discussed selec-
tion of design parameters for experimental
122 Practical Distillation Control
design, model structure selection, and pa-
rameter estimation that are pertinent when
the end use of the identified models is for
process control. Ideally, one would consider
the model requirements and trade-offs pre-
sented by different choices of these design
parameters and then proceed with the dy-
namic process test and identification based
on the particular design parameters se-
lected. In practice, however, constraints
(process as well as management) sometimes
make it impossible to conduct the desired
process test. Prejudice for selection of a
certain model structure also affects the pro-
cess test, as described in the following text.
The presence of disturbances and mea-
surement noise directly affects the variance
of estimated parameters. In general, the
size of perturbations applied to the inputs
should be as large as permitted (to maxi-
mize the signal-to-noise ratio) in order to
reduce the required duration of the experi-
ment. In situations where measurement
noise is low and a process operates rela-
tively smoothly between disturbances that
occur only infrequently, a shorter duration
experiment can be designed. In these situa-
tions, a small number of step changes (per-
haps combined with a relay test) will pro-
vide sufficient information to obtain an
acceptable control model. However, this is
only true if (low-order) transfer functions
are to be identified. If an a priori decision is
made to statistically estimate truncated im-
pulse (or step response) models, then this
dictates that a long duration experiment be
performed, to improve the variance of the
resulting estimates. As mentioned previ-
ously, a biased least-squares estimator might
also be performed with this model structure
to obtain a more favorable bias-variance
trade-off.
Other practical considerations may also
influence the decision to identify truncated
impulse models directly over transfer func-
tion forms: dimensionality (number of in-
puts and outputs) and the ability to conduct
tests on a multivariable process one input at
a time. If the process is sufficiently well
behaved to allow one input to be perturbed
while holding the others constant, then this
greatly reduces the complexity of the esti-
mation problem, because SISO models can
be estimated one at a time (assuming the
multivariable model Equation 7-4 is deemed
adequate and no disturbance model is to be
identified). Determining the structure,
deadtime, and polynomial orders for a
transfer function model Equation 7-27 can
be performed readily. The resulting transfer
function forms can be converted numeri-
cally to truncated impulse or step response
form without any loss of information.
Regardless of whether inputs are per-
turbed one at a time or simultaneously, it
may become necessary to allow an operator
to make moves in other inputs in order to
keep a process within an acceptable operat-
ing region. In either case the structural and
parameter estimation problem becomes
more complex if a multi-input model rather
than a single-input model must be identi-
fied, particularly if the number of inputs
becomes large. The trade-offs that occur
between duration of experiment, signal-to-
noise ratio (input amplitudes), model struc-
ture (impulse vs. low-order transfer func-
tion) must now also include the experience
level of the person performing the identifi-
cation. As dimensionality increases, the
problem of satisfactorily determining dead-
times, structures, and polynomial orders for
transfer function models becomes more
complex (particularly if the system exhibits
dynamic behavior that is more complex than
simple first or second order with deadtime,
as very typically happens). In these situa-
tions the probability of obtaining an incor-
rect structure becomes higher. Depending
on the training and experience of the user,
direct identification of impulse or step mod-
els (that require only one parameter to be
specified-the process settling time) may be
the best alternative. However, it must be
recognized that as dimensionality increases,
experiments of increasing duration are re-
quired. At some point, even the most ac-
commodating operations superintendent
may not permit such a long test, and a more
efficient approach will be required.
The following are some suggestions for
increasing the probability of performing a
successful dynamic test. The first suggestion
is to perform a "pre-test" on a system some
time before the actual dynamic test. The
pre-test allows one to identify potential
problems that may hamper (or postpone)
the actual test. Typical things to look for are
equipment problems (valves, transmitters,
analyzers, etc.), inadequate tuning of regu-
latory controllers, equipment in "nonnor-
mal" operating mode or range (e.g., bypass
valves), and functioning of the data-logging
program(s), to name a few. The pre-test
also allows one to explain to the operations
staff what the dynamic test needs to accom-
plish (and to agree on how it should be
performed to ensure that quality and oper-
ating constraints are not violated). In addi-
tion to information obtained from process
instrument diagrams, the pre-test can help
to identify points that should be logged dur-
ing the test. In addition to the points for
which models will be required, any point
that may potentially be used to verify nor-
mal operation (or conversely to later explain
an unusual operating occurrence) should be
logged. Points on the unit under test as well
as points on upstream (or interconnected)
units that have a direct effect on the test
unit sh01;lld be logged. Redundant or use-
less data can always be discarded, but miss-
ing data can, in the worst scenario, invali-
date a test. Data from the pre-test can also
provide information on the system dynam-
ics, which allows one to design the pertur-
bation signals for the subsequent dynamic
test.
During the actual test it is important to
have an operations engineer monitor the
system to record any events (e.g., opening of
a bypass valve, upset or shutdown of an
upstream unit, flushing of a heat exchanger,
tuning of a controller, putting a controller
in manual, unscheduled maintenance oper-
ation, etc.) that might invalidate a segment
of test data. The data logger should record
Identification of Distillation Systems 123
points with a time stamp so that segments
of data corresponding to periods of upsets
or unscheduled events can be identified
later.
76 EXAMPLE
As an example, an identification is per-
formed on a moderate purity refinery de-
propanizer. Feed to the column is the dis-
tillate flow from a debutanizer, which is
manipulated by the debutanizer pressure
controller. Flow rate of this stream is mea-
sured, but its composition is not. Analyzers
on the overhead and bottom streams of the
depropanizer provide composition measure-
ments every 3 min. A multivariable control
design resulted in three inputs: feed flow
rate (a feedforward variable), reflux, and
reboiler steam flow rates (manipulated vari-
ables) and two controlled outputs: overhead
and bottoms compositions. Discrete impulse
models were required for the control pack-
age, one for each input-output pair.
All data analysis, model structure deter-
mination, parameter estimation, and model
validation was performed using the System
Identification Toolbox (Ljung, 1988) in
MATLAB (1989).
The data set for the identification con-
sisted of 3 min samples taken over approxi-
mately 60 h. Discrete impulse response
models relating reflux flow, reboiler steam
flow, and feed flow to overhead and bot-
toms composition were obtained as follows.
The entire data set was used initially to
obtain least-squares estimates of truncated
impulse models relating each input -output
pair, which were converted to step response
models for examination. From plots of the
step response models, approximate dead-
times and model orders were determined. A
smaller portion of the data (the portion
considered to be most reliable) was then
used to estimate low-order transfer function
models [of MISO output error (OE) struc-
ture], using trial and error search of values
of deadtime and model polynomial orders
around those determined from the esti-
124 Practical Distillation Control
mated impulse models. The best model was
selected as the one that minimized the
Akaike information criterion (AlC) and that
satisfied model adequacy criteria based on
cross-correlation analysis (between the
residuals of the fitted OE model and the
corresponding inputs), cross-validation us-
ing a segment of the original data set that
was held back from the estimation, and
examining the poles and zeros of the fitted
model (if the 95% confidence region of a
pole of the transfer function for a particular
input -output pair overlaps with a zero of
the same transfer function, it is over-
parametrized). Disturbance models were not
identified because the control package used
assumed disturbance models (randomly oc-
curring deterministic steps), and it was not
considered necessary to modify it for this
particular application. Nevertheless, simul-
taneous identification of process and distur-
bance transfer function models can be read-
ily accomplished using the available fitting
routines in the toolbox.
All data were normalized by subtracting
off sample means (see function detrend in
the toolbox). Variance normalization (to unit
standard deviation) can also be performed
(for an example, see Section 5.4 of the
MA TLAB manual).
Only the identification of MISO models
for bottoms composition is shown here. Sim-
ilar procedures were followed to obtain
overhead composition models.
Normalized test data are shown in Figure
7-5 (a: bottoms composition (% C3); b:
reflux flow rate; c: reboiler steam flow rate;
d: feed flow rate).
Early in the test it was realized that the
column was overrefluxed (and reboiled).
Over roughly one shift (sample number 125
to sample number 275-approximately 7.5
h), both reflux and reboiler steam flow rates
were gradually reduced to bring the column
back to a more normal operating region.
These data were retained for preliminary
estimation of the impulse models (to obtain
rough estimates of dead times and model
orders) and for cross-validation (with a cer-
tain level of skepticism due to the abnormal
operating condition).
Due to exceptional operating problems
and disturbances in the debutanizer, feed
flow to the depropanizer experienced
greater fluctuation than normal during the
dynamic test period. This was not entirely
undesirable because it excited the column
for identification of models relating feed
flow rate and the response variables. It
sometimes can be difficult to arrange inten-
tional perturbation of feedforward variables
(because this usually means upsetting an
upstream unit). Because of the wide swings
in feed flow rate during the test, operators
did occasionally have to manipulate reflux
and reboiler steam rates to maintain opera-
tion within quality constraints as much as
possible. As described earlier in the chap-
ter, when data for identification are ob-
tained when feedback is present (inad-
vertently, as may have occurred in this case,
or due to a controller that is on-line), it is
recommended that external perturbations be
applied somewhere in the feedback loop.
During the depropanizer test, random per-
turbations were applied to reflux and reboil
steam flow rates to as great a degree as
possible, considering the unusual operating
conditions. However, these conditions did
preclude applying preprogrammed PRBS
signals to the inputs. As well, it was impos-
sible to perform a dynamic test in which
only one input at a time was perturbed
(while holding the others constant), which
would have simplified the model estimation
stage of the identification.
Several additional operating problems
arose during the dynamic test that were not
apparent in the pre-test, which had been
conducted several weeks earlier. As shown
in Figure 7-5b, a problem in the reflux con-
trol loop caused high-frequency oscillations
in the reflux flow. Tuning of the PID flow
controller did not resolve the problem.
Troubleshooting for mechanical problems
in the loop hardware also did not immedi-
ately solve the problem. Because the oscilla-
tions were high frequency, and would sim-
6
5
4
3 f\
h

2
r
I
\
0
)

-1,
l(I
II
-2
V
V\
V
3
V

V
-40
200 400 600 800
sample number
(a)
800
400
200
200
-400
-600
0
200 400 600 800
sample number
(b)
FIGURE 75. Depropanizer test data. (a) bottoms composition; (b) reflux flow rate;
(c) reboiler steam flow rate; (d) feed flow rate.
\j
V
1000 1200
1000 1200


g
1>0

.=


1


3
2
1
0
-1
-2
0
1.5
1
0.5
1 0
-0.5
5
5
10 15 20 2S 30 35 40 45 SO
. (a)
10 15 20 2S 30 35 40 4S so
(b)
DGURE 77. Estimated impulse models. (a) bottoms composition/reflux; (b) bot-
toms composition/reboiler; (c) bottoms composition/feed.
ply be filtered by the process, and it was still
possible to make setpoint changes to the
reflux flow controller, it was decided to pro-
ceed with the test while troubleshooting
continued. A mechanical problem was fi-
nally resolved near the end of the test
(around sample number 975).
Power spectra were generated for all in-
puts using the spa function in the toolbox to
determine the useful frequency range of the
input perturbations. The spectrum for re-
flux flow rate is shown in Figure 7-6. Exami-
nation of the frequency functions (Le. Bode
plots) obtained from spectral analysis (gen-
erated by the spa function) and the Bode
plots of the final estimated models indicated
that the input perturbations did not contain
sufficiently high frequency information. The
resulting control models would be consid-
ered adequate only if relatively sluggish
controller tuning proved to be adequate.
Power spectra for the other inputs were
similar.
Low-frequency drift in the data should
also be removed by prefiltering (for exam-
ple, using the filter function idfilt in the
toolbox as a high-pass filter). Because any
controller with integral action will be able
to regulate effectively against very slow dis-
SPECI'RUM
frequency
FIGURE 7-(,. Power spectrum of reflux flow rate.
Identification of Distillation Systems 127
turbances, models identified for use in pro-
cess control need not be accurate at low
frequencies (Le., near steady state). The
need to obtain a high quality estimate of
steady-state gain for control models is
overemphasized in the identification folk-
lore. Differencing data will also remove
nonstationarity (low-frequency drift), but
will result in fitted models in which bias at
higher frequencies is reduced (at the ex-
pense of accuracy at lower frequencies that
are of interest for process control) because
taking the derivative of a signal distorts its
spectrum (Le., amplifies power) at higher
frequencies.
For systems that do not exhibit strong
gain directionality (as is the case with this
moderate purity tower), input perturbations
should be independent (noncorrelated).
Cross-correlation functions for the input se-
quences were generated and plotted (not
shown) using the routine xcov in the Signal
Processing Toolbox in MATLAB. No sig-
nificant cross-correlation between the in-
puts was found.
Preliminary least-squares estimates of the
impulse responses for each input-output
pair were obtained by fitting a three-
input-one-output ARX model [Equation
I .S ,--------r-----.-------r----- .------------,------,
200 400 600
I8mple number
( c)
800 1000 1200

o
-500
-1000
200 400 600 800 1000 1200
sample number
(d)
FlGURE 7-5. Continued.
Identification of Distillation Systems 129
1.S .----.----.----.----.----.-----,.------',.-----,.-----,.-----,
O.S I
..
..
"
0
'"
..
..
..
..
co
...

0

H
( c)
F1GURE 7-7. Continued.
7-31, with na = 0, i.e., A(q) = 1 and nbi =
50, the order of Bj(q) for input i, i = 1,2,3]
using the toolbox routine arx. Alternately
the toolbox routine era, which obtains im-
pulse models by a cross-correlation tech-
nique, could have been used. The estimated
impulse models are shown in Figure 7-7.
The toolbox routine idsim was used to ob-
tain the corresponding step response mod-
els (by simulating the response of the esti-
mated three-input-one-output impulse
model to a unit step change applied to each
input, one at a time).
The unit step response models are shown
in Figure 7-8. From these plots it was deter-
mined that the responses of bottoms com-
position to reflux, reboiler, and feed flow
rate each has a deadtime of approximately
five sampling periods (15 min). Each of the
responses appears to be low order [the or-
der of F(q) for a transfer function model
will be low]. The response of bottoms com-
position to reboiler steam flow rate rate
exhibits inverse behavior, suggesting that the
order of B/q), i = 2 (reboil), for a transfer
function model should be at least 2.
Low-order transfer function models were
identified using sample numbers 630 to 1193
(approximately 28 h). Because no distur-
bance models were to be estimated, an out-
put error (OE) structure was selected for
identification. An ARX structure could also
have been used (Equation 7-31), but as
shown by Equation 7-29 this structure im-
plicitly assumes a disturbance model of the
form H(q) = 1/A(q). This is undesirable
mainly because H(q) = 1/A(q) results in a
fit in the frequency domain that emphasizes
reducing bias at high frequencies, as shown
by Equations 7-15 and 7-19 [with H(q) =
I/A(q), the quantity 1/IH(e
jwT
)1
2
becomes
larger at higher frequencies, thereby result-
Xlo-'
8
7
6
5

..
<=
4
0
...
..
:l
...

..
3
'"
..
....
;
2
0
-1
0
5 10 15 20 2S 30 35 40 45 SO
(a)
1
0
-1
-2
..
..
<=
0
-3 ...
II
:
...

..
-4
'"
..
....
g
-5
-6
-7
-8
0
5 10 15 20 2S 30 3S 40 45 SO
(b)
FIGURE 7-8. Unit step responses obtained from estimated impulse models. (a)
bottoms composition/reflux; (b) bottoms composition/reboiler; (c) bottoms composi-
tion/feed.
nO"'
8
7
6
5
u
..
"
0
4
'"
..
:!
'"
..
..
3
'"
..
....
:
2
1
0
-1
0
5 10 15 20
FIGURE 7-8. Continued.
ing in greater weight being placed on reduc-
ing bias in the fitted model at higher fre-
quencies].
The toolbox routine oe for estimating
output error models allows only one input
and so the general routine for prediction-
error models, pem, was used instead. With
three inputs and the polynomial orders for
A(q), C(q), and D(q) set to zero, Equation
7-27 becomes
Bt(q)
y(t) = Ft(q) ut(t - nkt)
B2(q)
+ F
2
( q) u2(t - nk2)
B3(q)
+ F3(q) u3(t - nk3 ) + e(t) (7-66)
(c)
Identification of Distillation Systems 131
2S 30 35 40 45
where y = bottoms composition
u 1 = reflux flow rate
u
2
= reboiler steam flow rate
u
3
= feed flow rate
50
The following estimated DE model mini-
mized the Ale (and satisfied all other vali-
dation criteria, as shown below). Parameter
estimates (denoted by carets) are given with
their 95% confidence intervals in parenthe-
ses.
Bottoms composition/reflux flow rate:
nbt = 1 b
u
= 0.57 X 10-
3
( 0.11 X 10-
3
)
nfl = 1 111 = - 0.92 (0.014)
132 Practical Distillation Control
Bottoms composition/reboil steam flow
rate:
n
k2
= 4
nb2 = 2 b
21
= 1.03 (0.64)
b
22
= - 1.34 (0.74)
n/
2
= 2 121 = - 1.59 (0.15)
122 = 0.634 (0.14)
Bottoms composition/feed flow rate:
nk3 = 5
nb3 = 1 b
31
= 0.98 X 10-
3
(0.07 X 10-
3
)
nf3 = 1 131 = - 0.86 (0.015)
None of the confidence intervals encom-
passes zero, indicating that all the estimates
are statistically significant.
The poles and zeroes of the transfer
function for each input -output pair were
1.5
1
0.5
'iii

'tI
0
...
III
II
...
-0.5
-1
-1.5
0 100 200
calculated using the toolbox routine th2zp
and plotted (along with their 95% confi-
dence regions, which are not shown) using
the toolbox routine zpplot. Possible redun-
dancy in poles and zeros was only a concern
for the transfer function relating bottoms
composition and reboil steam rate because
this input-output pair had two poles. The
pole-zero plot obtained from zpp/ot for
this input-output pair showed no overlap-
ping of the 95% confidence regions for the
poles and the zero, and therefore over-
parametrization was not suspected.
A plot of the residuals for the fitted
three-input-one output OE model is shown
in Figure 7-9. The autocorrelation function
for these residuals (Figure 7-10a) shows sig-
nificant values (Le., outside the 95% confi-
dence limits given by the dotted lines), which
is expected because a disturbance model
was not estimated. At this point a distur-
bance model could be identified by adding
the polynomials C(q) and D(q) (Equation
7-27) in the toolbox routine pem and deter-
300
'00 500 600
sample number
FIGURE 79. Residual plot for estimated OE models.
Identification of Distillation Systems 133
Correlation function of residuals. Out ut * 1

0.5
o

o 10
lac
(a)
15 20 25
Cross corr. function between input 1 and residuals from out ut 1
0.4r= ...... = ....... = ...... =."" .. ...... = ..... "= ....... = ... ""= ...... r.""=" ..... = ....... = .... = ...... ...... = ...... = ....... = ....... = ...... .. "= ..... = ..... "= ....... =." ... 7 .... ".= ...... = .... "= ...... = ...... ...... =_ .. = .. " .. = ....... =." .... .. _=."_.= .. = .. =" ... .. __ = .. .. _= ..... ... "_- .... ,
0.2
o

-0.2
-0.4'"'' .......................... .
-25 -20 -15 -10 -5 o 5 10 15 20 25
lac
(b)
Cross corr. function between input 2 and residuals from out ut 1
.. .. ... .. ...... ...... = .... .... .... .... ..... .... .. ...... ... ... ....... ......
0.2
o
-0.2
-0.4 ...... = ....... = ...... "" .... = ....... =. =""== .. '=.' = ...... "" .. = ...... = ....... = ....... = ...... "" .... = ....... = .... "= ....... = .. _.""' .... = .. =""""= .. = ....... = ...... = ..... "" ... = .... _= ... = ... _=."", .. = .... = ...... =_."'-' .. .
-25 -20 -15 -10 -5 0 5 10 15 20 25
lag
( c)
Cross corr. function between input 3 and residuals from out ut 1
0.4 ...... = ....... = ....... = ..... :r.: ..... =. ....... = ..... = ..... =. ...... =.;::: ... =.:: ....... = ...... =_.::: ...... :r.:.= ....... =.:: ....... = ..... :=.!.: ..... :r. .. :::.: ...... = ....... = ...... =_ .. :.:. _= .... = ...... = .... _=.::r.:::: ... = .. _=_.::::::;:. =.._=_ ... = ... .. !:::% .. =_= ..... = ..... :::::-, ... .
0.2
----
----

o
-0.2
-0.4 L...::;;===".;;;;; .. ;;;;; .... ;;;;; ..... "" ..... ;;;;; ...... ;;;;; ..... ;;;;; ...... ;;;;; ...... ::;L . ;;;;; =="".;;;;; ...... ;;;;; ...;;;;; ...... ;;;;; ... '.::;L.' ;;;;; ;;;;; . ;;;;; ;;;;; ;;;:l .... "" ..... ;:: ...... ;:: ...... "" ..... "" ..... "" ;;;;; ..... ;;;;;; . ;;;;;; ._.;:: ...;;;0 "" ."" ...""_ ."" ...... ;;;;; .. ""-;:: ... =;.;.;.....0
-25 -20 -15 -10 -5 0
lac
(d)
5 10 15 20
FIGURE 7-10. Auto- and cross-correlation functions for estimated OE models. (a)
Residual autocorrelation function; cross-correlation functions between residuals and
inputs; (b) input 1 (reflux); (c) input 2 (reboiler); (d) input 3 (feed).
25
134 Practical Distillation Control
mining an adequate structure as described
by Box and Jenkins (1976).
The cross-correlation functions betw