STUDIES IN. AUTOMATION AND CONTROL, 5 MODELS AND SENSITIVITY OF CONTROL SYSTEMS ANDRZEJ WIERZBICKI Institute of Automatic Control Technical University of Warsaw Warsaw, Poland ELSEVIER Amsterdam—Oxford—New York—Tokyo 1984Published in coedition with Wydawnictwa Naukowo-Techniczne, Warszawa Translated by R. Miklaszewski (ch. 2-4) and J. Gandelman (ch. 1) from the Polish Modele i wrazliwosé ukladéw sterowania published by Wydawnictwa Naukowo-Techniczne, 1977 Distribution of this book is handled by the following publishers: for the U.S.A. and Canada ELSEVIER SCIENCE PUBLISHING COMPANY, INC. 52 Vanderbilt Avenue New York, N.Y. 10017 for the East European Countries, China, Northern Korea, Cuba, Vietnam, and Mongolia WYDAWNICTWA NAUKOWO-TECHNICZNE ul. Mazowiecka 2/4 00-048 Warsaw, Poland for all remaining areas ELSEVIER SCIENCE PUBLISHERS B.V. 1, Molenwerf P.O. Box 211, 1000 AE Amsterdam, The Netherlands Library of Congress Cataloging in Publication Data Wierzbicki, A. P. Models and sensitivity of control systems. (Studies in automation and control; v. 5) Translation of: Modele i wratliwosé ukladéw sterowania. Includes bibliographical references and index. 1. Control theory. 2. Mathematical models. I. Title. IL. Series. QA402.3.W5413 1984 629.8312 84-1476 ISBN 0-444-99620-6 (vol. 5) 0-444-41792-3 (series) Copyright © 1984 by WYDAWNICTWA NAUKOWO-TECHNICZNE, WARSAW All rights reserved. No part of this publication may be reproduced, stored in retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the publisher. Printed in PolandContents Introduction vit Chapter dd 1.2 1.3 15 Chapter 24 2.2 23 24 1 Mathematical Models 1 Prerequisites of model building 1 Examples of model building 10 Properties of fundamental types of models 44 1.3.1 Continuous linear models 44 1.3.2. Linear time discrete models 61 1.3.3 Non-linear models 69 1.3.4 Stability, controllability and observability of a model 76 Identification of parameters and model verification 90 1.4.1 Characteristic features of model identification and verification problems 90 1.4.2 Regression analysis and identification of non-linear static models 94 1.4.3. Correlation function identification for linear dynamic models 101 1.4.4 Identification of linear and non-linear dynamic models by the least sum of squares and related methods 111 A general form of the models 123 2 Sensitivity Analysis of Mathematical Models 137 Foundations of sensitivity analysis 137 2.1.1 Mathematical foundations 138 2.1.2. Basic relations in sensitivity analysis 143 2.1.3 Basic examples of application of model sensitivity analysis 156 Structural methods in sensitivity analysis 163 2.2.1 General rules for differentiation of structural models 163 2.2.2 Commutative structural sensitivity models 169 2.2.3 Applications of structural methods of sensitivity analysis 176 Sensitivity of control systems 180 2.3.1 The general problem of sensitivity of a control system 180 2.3.2 Examples of sensitivity analysis of control systems 188 Structures and sensitivity of stabilizing systems 200 2.4.1 Sensitivity of a feedback system 200 2.4.2 System stabilizing structures 204 2.4.3 Parametric sensitivity of stabilizing systems 209 vVI Chapter 3.1 33 Chapter 4d 42 43 CONTENTS 3. Optimization and Optimal Control 217 Theoretical foundations of optimization 217 3.1.1 Problems with resolvable constraints 219 3.1.2 Convex problems 222 3.1.3. Other problems with constraints 226 3.2 Dynamic optimization and optimal control 241 3.24 3.22 3.2.3 3.2.4 Computational methods of optimization 265 3.3.1 Direct gradient methods for problems with no additional constraints 265 3.3.2. Newton’s method in control space 276 3.3.3 Other computational methods 281 Lagrange functionals 241 The optimality principle and synthesis of optimal closed-loop control 251 The maximum principle 257 Examples of determining the optimal control 261 4 Sensitivity Analysis of the Optimal Control Systems 285 Problem formulation and theoretical foundations 285 41.1 4.1.2 The sensitivity of the basic model of an optimization problem 295 4.1.3 Sensitivity of an extended model of an optimal control system 304 4.1.4 Examples of determining the performance losses 311 Structures of optimal control systems 319 42.1 42.2 4.2.3 424 4.25 4.2.6 Survey of particular problems 357 43.1 43.2 43.3 43.4 References Index 393 385 Problem formulation 286 Open- and closed-loop structures 319 Structures of optimal trajectory tracking 325 Structures with optimizing feedback 331 Adaptive structures 336 The problem of structure selection 342 Examples of applications 344 Sensitivity of linear systems with quadratic performance indices 358 Sensitivity of time-optimal systems 368 Variable structure models 375 Survey of sensitivity analysis problems 381Introduction From the very beginning of civilization man has purposely been influencing his environment. This influence can be analysed from various viewpoints. We can consider the philosophical aspects of this influence, characterizing in various ways the purposefulness of that activity. The scientific aspects are related to the development of natural and social sciences which make an intentional influence possible and to that end are, in fact, being developed. The evaluation of the results achieved by this influence can lead to various conclusions; mankind is not always able to predict precisely the results of its activities. Therefore the methodology of a purposeful activity, of decision making and of analysing the possible misjudgements has attracted the interest of many researchers in various branches of science. One of the branches related to those problems is control theory. Control. A purposeful influence of man or of the technical devices constructed by him on the environment, and in particular on other technical devices or processes, is generally called control. The branch of science related to control problems is often referred to as cybernetics. The most mathematical part of that science is called, though less commonly, control theory. The latter has developed its own system of basic concepts and definitions, but relies on the results of modern mathematics. The concept of control consists of such detailed concepts as controlled process, constraints, control goal and control performance. The controlled process or object is that part of the environment which is influenced by the control. The use of the word process emphasizes the fact that control and its outcome should not relate to a given moment and steady state only, and that we are primarily interested in the dynamic development in a given time interval of the results of control. Constraints of control reflect the fact that the process cannot be controlled in an arbitrary way. The goal of VILVII INTRODUCTION control is the postulated result of our control action. If that goal is attainable, then it can usually be reached in many ways. Therefore, it is useful to specify the performance index of control as the measure of the quality of control, making it possible to choose between various control actions leading to the same goal. The choice of the best control action that ensures the realization of the control goal and minimization or maximization of the performance index under the given constraints is called control optimization. The control goal can also be defined indirectly and the definition results only from control optimization, since an optimal control process might be a goal in itself. Mathematical model. The choice of a control action is based on the available knowledge of the controlled process, its constraints, its goal and performance index. This knowledge is often not precisely recognized. It may just be a part of our intuition or experience. However, if control actions are to be performed automatically by technical devices, without continuous interventions of man, then it is necessary to specify the available knowledge in a form which is precise enough to be transferred to technical devices. More generally, if a decision to be made is sufficiently important, a detailed analysis of the possible decisions and their outcomes or even a decision optimization might be desirable. Also in this case it may prove useful to specify the available and pertinent knowledge in the form of a mathematical model. The concept of a model, though often abused in the popular sense, is in fact fundamental for many sciences, has a basic cognitive and methodo- logical meaning, and provides a cardinal tool of research. Models can be divided into cognitive models and purpose-oriented models (though every cognitive model obviously has its purpose); they can be also classified into mathematical, physical, etc., models. For control theory and its applications, mathematical purpose-oriented models are of primary importance since their form and accuracy can be adapted to the particular problem of control or decision choice. The methodology of model building and of analysing various properties and applications of mathematical models has been developed together with control theory, particularly in the course of the last four decades. Some recent books—as an example, the monograph by Kalman, Arbib and Falb [55]—try to summarize this broad knowledge. However, usually this synthesis is limited to more abstract properties of mathematical models. The models are then referredINTRODUCTION Ix to, though not very luckily, as systems or dynamic systems, since the word system has in fact a broader sense. There are many textbooks and monographs dealing with the fundamental types of models in control theory or with the properties of models for given classes of controlled processes. In this book an attempt is made to present the problems related to mathematical models of control processes from a different viewpoint. Starting with the analysis of the concept of a model, special attention is paid to the purpose-oriented form of the model and the methodology of model-building for a given process, thus through an analysis of the properties of various classes of mathematical models and a description of model identification and verification methods we arrive at a general class of abstract mathematical models. This class is slightly larger than that investigated in [55], since it contains both dynamic and static models. The above problems are discussed in detail in Chapter 1 of the present book. Here, we shall review some basic notions related to the use of mathematical models for control problems, since they are necessary to permit an approach to the problems considered in the later chapters. A mathematical model of a control problem is, somewhat simplifying, the sum of the pertinent knowledge related to the controlled process, constraints, control goal and performance index, expressed in the form of mathematical relations. There are many tvpes of such models. A most useful form of a mathematical model is the structural model, which classifies signals or variables in the process to be controlled and indicates types of relations between those variables without specifying the relations precisely. The classification of process variables starts with the classic division into input variables which express the influence of man or of the environment on the process, and output variables which express the influence of the process on the environment. The goal of control is usually expressed in terms of output variables. As regards input variables, it is useful to distinguish (Fig. 1.1) control variables or controls which represent the purposeful influence on the process, disturbing variables or disturbances which represent various other (usually random) influences of the environment on the process and make the achievement of the goal difficult, and parameters or explanatory variables which can be assumed to be known with a given accuracy—at least in a given interval of time in which the model is investigated and assumed to be valid. Similarly, it is useful to classify further the output variables,x INTRODUCTION uy gi Controtu Output variables y yi leaeca Model of @ proce —oO Oy a u,=u, Parameters a | {state variables x De te i flees ze Xe (Q) z z (@) 2 es yy u y= Ue ye oe q A H fe | GemURTT oxy 2, | 4 aff a aay fae & 22 ua=Y te Goan a2 we? Fig. L1 Structural models: (a) simplest one, (6) tandem connection, (c) parallel par: connection, (d) feedback connection. Processes of a veritably dynamic character are distinguished as possessing some type of memory: the output variables depend not only on the current values of the input variables, but also on their past course. In most types of mathematical models it is possible to represent the whole past of the process by introducing a sufficient number of additional internal or endogenous variables, called in this case state variables}; some of these variables can also be interpreted as output variables. More precisely, the state of a mathematical model of a process (briefly, the state of a process) is the smallest set of variables whose values determine the future course of the output variables, if the precise form of the model, its parameters and the future course of other input variables are known. The necessary number of state variables, called the process dimension, is usually greater than the number of output variables that can actually be observed in a real process (though of course, the state variables can always be treated as output variables of the process model). e If the analysed process is complicated enough and it is possible to split it into simpler subprocesses with distinct input and output variables, then ¥ If the concept of state is sufficiently extended, it is possible to provide such a represen- tation of the past for all types of mathematical models. However, the concept of the state is often used only for those models which have a finite dimension, that is, a finite number of state variables. As in the theory of stochastic processes, such models are called Markov-type models.INTRODUCTION XI a structural model comprising the connections between various subprocesses contains important information. There are several basic types of inter- connections between subprocesses; some of them are presented in Fig. I.1. A necessary supplement to the structural model is the information on sets of values of model variables. Depending on the kinds of these sets, we classify the basic types of models. If the sets of values of input and output variables (including state variables) are countable, we speak of a model discrete in level or quantified model. If, furthermore, these sets are finite then we speak of a finite model or a model with a finite number of states}. The simplest models of this type are just logical relations between input and output variables. If the sets of values of input and output variables are uncountable (usually they are compact subsets of real axes), we speak of a model continuous in level. Also vital is the information on the set of time instants in which we analyse the behaviour of the process or its model. If this is a countable set, then we are dealing with a time discrete model (a difference equation is then the typical form of the input/output relation). If, however, the set is uncountable, then we are dealing with a model continuous in time (a differential equation is then the typical form of input/output relation). Similarly important for the sake of mathematical exactness of model formulation are assumptions about function spaces whose elements are the model variables. For example the control of a time discrete model may be a bounded sequence, or a square summarizable sequence; the control of a model continuous in time may be a square-integrable or bounded or piece-wise constant function. All the above types of models may be either static or dynamic. In a static model, the present values of input variables determine the present values of output variables. In such a model there is no need to introduce the concept of state; the model is zero-dimensional. In a dynamic model, the state variables are defined and the model dimension is not zero. Models which are continuous in level (but not necessarily in time) may be linear, i.e. described by linear relations{} between input and output + The finite model should not be confounded with the Markov model. The Markov model has a finite number of indices of state variables, whereas the finite model has additionally a finite number of values of state variables. Technical devices represented by finite models are often called finite automata. tt Strictly speaking, for a model to be linear it is sufficient that the dependence between input, state, and output variables be affine. Addition to the output variable of a component which is independent of input variables does not change the essence of the model.XIE INTRODUCTION state variables. A linear model is usually the simplest form of a model, and its first approximation which is valid for minor deviations from the nominal working conditions. A complete model of a control problem contains not only the structural model and a description of its variables but also the exact form of input/output or input/state/output relations, as well as relations which define constraints, the control goal and performance index. In order to obtain a possibly precise model, in the beginning we usually define only the general form of these relations, which depend on parameters with unknown values. When testing the actual process experimentally, and comparing its behaviour with the model, the values of these parameters may be selected to achieve the possibly best representation of the process by the model. This method of fitting a model to a process is called process identification. There are many identification methods, suitable for different types of models, In some special cases, a model can be ideally fitted to a single run of a process; but it is impossible to build an ideal model which will truly reflect the process behaviour in all possible situations. This results from the fact that a model is just a model and does not cover the wide scope of phenomena occurring in the process and in the surrounding environment. An often useful method of model perfection is the use of a probabilistic model—a model in which relations between output and input variables are not fully determined, but only described by probability distributions or by stochastic processest. But a probabilistic model remains a model; to define it fully we must specify the types of probability distributions and the parameters of these distributions, which again leads to identification problems, etc. In some real applications better results are achieved by applying probabilistic models, in others, deterministic models}. However, + Since model parameters are understood as a special type of input variables, we can assume that without loss of generality, the input/output relations are always non-pro- babilistic, and only the input variables—and, consequently the output variables—may be of a probabilistic nature. A probabilistic model can therefore be understood as one in which at least one input variable is a random variable or a stochastic process. +t More precisely, nonprobabilistic models, because probabilistic models may also be interpreted from a deterministic point of view. The terms determinism and in- determinism are philosophical concepts, expressing a point of view on the character of nature; the probabilistic and nonprobabilistic methods may be used by researchers with either deterministic or indeterministic points of view (even though indeterminism considers probabilistic models as ultimate, cf. the probabilistic theory of elementaryINTRODUCTION XII both these types of model involve a basic inconsistency related to the concept of mathematical model: the greater exactness of the model is required (when the amount of experimental data is limited) the less confident conclusions can be drawn regarding its form, parameters, and extent of applications. The inconsistency between the exactness and utility of a mathematical model is one of the principal paradoxes in the history of science, in the process of man learning about nature. The notion “physical, biological or economical law” has been used in two senses. In one sense, it is understood as an objective property of nature, of our environment. Owing to such objective property, the behaviour of our environment can be scientifically examined and is reproducible provided that the corresponding experiments are conducted in the same conditions. In another sense, a law of nature denotes its model, usually a mathematical model, formulated by man. In this sense we speak of Newton’s laws, the Einstein laws or the probabil- istic model of an atom. And even though such models are of tremendous cognitive significance, it has always been so that their area of application is limited, that more precise models can be constructed, often of a different type, after a suitable amount of experimental data have been gathered and analysed. Thus we are close to the conclusion that it is a methodological error to attach an absolute value to a more-or-less perfect model, which error can only partially be justified by the natural desire of researchers to gain a better knowledge of nature by constructing ever-better models of its objective laws. Sensitivity analysis. In the applications of control theory the choice of a model, when constructing purpose-oriented models for solving a particular control problem, is to a great extent arbitrary, being the result of compromise between the required exactness and the effort required to collect a greater amount of experimental data and determine a more exact model. Since absolutely precise models are not used (in fact do not exist), the effect of model inaccuracy on decisions based on this model acquires basic significance. particles). Nevertheless the term deterministic model is still often used in the everyday meaning of nonprobabilistic model, and it is in this sense that it will be used in the present book.XIV INTRODUCTION This problem is not new and its simpliest variant has been examined in many fields of science. The basic variant may be formulated as follows. A certain decision is considered which is supposed to have a specific effect when applied to a given model. The first question is: how will the effects of this decision change if we apply it to real life and not to a model? This question can only be answered by practical tests which, however, may turn out to be costly and time-consuming. Hence let us pose the next question: how will the results of decision be affected if the model to which this decision applies is changed (usually only slightly)? Notice that in fact we are dealing here with two models; the first, on which the decision has been based, and the second, for which the influence of model changes on the results of the decision is analysed. These models differ in their purpose and character. The first one represents the best purpose-oriented approximation of our knowledge of the problem at hand and serves for decision making. We shall call it the basic model. The other one represents the possible deviations of reality from the first model, more precisely—those deviations which we consider to be crucial and likely to occur; we shall call it the extended model, because it is usually more complicated than the basic one. Consider, by way of example, a problem of tolerance analysis, that is, the examination of the correctness of an engineering design while allowing for possible inaccuracies of the elements. Let the decision, based on the basic model, relate here, for example, to the dimensions of mechanical components; the extended model represents the possible inaccuracies of these dimensions characterized either in a deterministic way by specifying the tolerance limits, or in a statistical way by specifying the random distributions of the deviations or their basic moments. The need for a fully precise distinction between the two models, the basic and the extended one, is not very apparent when only the first variant of the problem is considered. It might seem that the analysis of one model with changing parameters would be sufficient. In the above example it seems to be immaterial whether an extended model which allows for inaccuracies of dimensions (resulting, e.g., from the manufacturing method) is considered, or if it is assumed that deviations of the same character occur in the basic model (resulting, e.g., from inaccuracies in designer’s calculations). Such a principle of relativity of the basic and extended model deviations could be very useful, since it would contribute to a significant simplification of the analysis. However, this principle, though it can beINTRODUCTION Xv applied in many cases, is not universally true. It may give wrong results, for example, when the basic model has extremal properties (i.e., if deviations of any sign in the model result in decision deviations of one sign only). The situation becomes even more complicated when we consider models of control problems. One of the basic methods of decreasing the influence of model inaccuracy on the results of control is the use of feedback or closed-loop control systems. Instead of accepting that control decisions are dependent only on a mathematical model of the problem (the so-called open-loop control system, see Fig. I.2a), we can use the mathematical model to select a certain control law, which makes the control of the real process dependent on the results of measurements of output variables in this process (the closed-loop control system, see Fig. I.2b). (a) (6) Mathematical (Mathematical model model ! : i u uw u y Control : t= Real process += Control law }= Real process © decision ees eee eee eee Fig. 1.2 Basic structures of control systems: (a) open-loop, (b) closed-loop. The problem of the effect of model inaccuracy on the results of control can then be formulated as follows: a certain control law is given, which when applied to the given model brings about some expected results. How would these results change, if the control law were applied to a real process and not to the model? If experimental tests are to be preceded by a more detailed analysis, then how do the postulated results of control change when we change the model of the process to which the law is applied? Of course, the selection of the control law is also a certain decision, so the methodology of analysis does not change fundamentally. Thus for a precise definition of the problem two models are required: the basic one, on which the control law is based, and the extended one, which re- presents the most likely and significant deviations of the basic model from reality. The control law should be applied to the extended model and the influence of model changes on the control results should be examined. However, the methodology of analysis changes since cases in which the basic model and the control law resulting from it have extremal properties,XVI INTRODUCTION are much more frequent. This is the case especially when control optimiza- tion is performed, or the optimal control law is defined. Moreover, optimal control systems can have open-loop or closed-loop structures as shown in Fig. 1.2, as well as many other structures, Examination of the influence of model inaccuracy on the outcome of application of the control law resulting from it is called sensitivity analysis of the control system. These studies have quite a history. For nonoptimal control systems sensitivity, such investigations were started by Bode [12]; Tomovié was the first to devote a monograph to this problem [115]; further, the works of Kokotovié and Rutman [57] as well as the others [18] deserve mention. In all these publications, however, distinction between the concepts of a basic model and an extended one is neither methodologically analysed nor consistently pursued; also, conditions under which this distinction is not necessary are not precisely formulated. This approach resulted from the intuitional assumption of relativity of changes in these two models; on the other hand, this assumption is correct in most cases of nonoptimal control system analysis. The problem of optimal control system sensitivity analysis was defined by Bellman [9] and next analysed by Dorato, Pagurek and Witsenhousen [95, 133, 18]. These publications already distinguish two models without which it would be impossible to analyse the optimal control system sensitivity. However, they provide no thorough analysis of the basic properties of these models and their relation to the formulation of the optimal control problem. Pagurek and Witsenhousen have achieved here a mathematically correct result, which, however, presents serious interpretational difficulties, They have proved that the linear approximations of the performance index changes, caused by changes of parameters in open- and closed-loop systems, do not depend on the use of feedback, and therefore are the same for both these systems. From this fact a paradoxical conclusion was drawn that the perform- ance index sensitivity is the same in both the closed- and open-loop optimal control systems. Further works have either accepted this conclusion or confirmed it under more general assumptions and in a different formula- tion. Kreindler [62], accepting that the performance index sensitivity is identical for closed- and open-loop systems, examined the differences of state trajectory sensitivity in these systems. Kokotovié et al. [58] have gone further: assuming the possibility of an exact evaluation of parameter changes and of an approximation of the control law by a Taylor series expansion depending on these changes, they have proved that also higherINTRODUCTION XVII derivatives of performance index changes can be identical for both the closed- and open-loop systems. This trend has been pursued in a number of other works [18]. The present author undertook studies in another direction [121, 122, 128]. While examining the basic conditions of control optimality and fundamental properties of the basic and extended models, it is possible to prove that the Pagurek—Witsenhousen result is a natural outcome of these conditions and properties, and, moreover, that it is true for any optimal control system, not being limited only to the basic variants of open- and closed-loop systems. One should not conclude, however, that this implies identical sensitivity in open- and closed-loop systems. In view of the basic conditions of control optimality, the sensitivity of different systems can only be compared on the basis of second derivatives of the performance index with respect to parameters (even though the first derivatives may be non-zero, as optimization relates to control and not to parameters). Under realistic assumptions the differences in the performance index sensitivity can be very larget. This trend was followed in many further works [7, 10, 25, 45, 82, 97, 114, 117, 118, 125, 135]. Chapters 2 and 4 of this book dealing with the sensitivity analysis of optimal and nonoptimal control systems constitute a summary and an attempt to synthesize the results of these works. Chapter 2 deals with model sensitivity and control system sensitivity. Model sensitivity analysis is considered to be the examination of changes of solutions, output variables, or state variables of a mathematical model due to changes of its parameters. Results of such an analysis are useful, if the hypothesis of relativity of the extended and basic model parameter changes is valid in the control system sensitivity analysis. Thus concepts and properties of extended and basic models are also discussed in detail in Chapter 2, the relativity principles, that is, conditions under which control system sensitivity analysis is equivalent to examining the sensitivity of only one model, being examined. Though a part of the detailed results in Chapter 2 has been derived from earlier works, their presentation and + Seemingly, this statement contradicts the results obtained by Kokotovié et al. [58] that also higher derivatives of the performance index changes can be identical. However, the results of the quoted authors are based on the assumption that a precise measurement of the parameter changes is possible and that these measurements are used for control Purposes. Such an assumption might be subjected to a methodical criticism: had it been acceptable, then an expanded model could always be made equal to the basic one and entire sensitivity analysis would be pointless.XVIIL ’ INTRODUCTION full interpretation is based on the generalization of the concept of a model and on an analysis of the properties of extended and basic models, Chapter 3 constitutes a preparation to Chapter 4. It deals with optimiza- tion problems and methods for the determination of optimal control. This chapter is a synthetic review of an extraordinarily wide range of research results in this field, also taking into account some results of the present author and his collaborators [42, 68, 73, 126, 129]. The systematics of the material is taken from the present author’s work [39]. Chapter 4 is dedicated to the sensitivity analysis of optimal control systems. It is a synthesis of research results obtained by the author and his collaborators in this field. It contains a generalized formulation and a detailed discussion of optimal control system sensitivity analysis, the distinction between the extended and basic model analysis, an interpretation of the Pagurek-Witsenhousen paradox, and several examples of differences in performance index sensitivity. Different structures of optimal control systems, conclusions on their applicability, as well as some more detailed problems are discussed in this Chapter. The present book is basically a monograph addressed to researchers in the field. However, the more fundamental parts of it are formulated in text-book form, to facilitate its use. It may serve as supplementary reading for senior or graduate students of control sciences, computer sciences, operational research, or systems sciences, and also for students of mathe- matics interested in mathematical control theory. For its full use, a certain mathematical knowledge in the fields of algebra, differential and difference equations, operations calculus, the Laplace, Fourier, and Laurent transforms, fundamentals of the calculus of variation and functional analysis, mathematical statistics and stochastic process theory is required. Although the book has a fairly theoretical approach, it gives examples of the construction of specific models for industrial processes as well as of the identification and utilization of these models for optimal control purposes, together with sensitivity analysis. It may therefore also be of interest to engineers designing complex systems of industrial control. Many people have contributed to the creation of this book. The whole staff of the Institute of Automatic Control of the Technical University of Warsaw has influenced it, especially professor W. Findeisen, professor A. Gosiewski, Dr. J. Putaczewski, and Dr. J. Szymanowski. The following workers participated directly in the research work on sensitivity analysis of control systems: Dr. A. Dontchev, Dr. B. Frelek, Dr. M. Machura,INTRODUCTION XIX Dr. P. Szablowski, Dr. A. Urbag, A. Bauer, D. Zardecki, W. Bernacki, A. Urbanik, and many others. In research on the fundamentals of optimiza- tion theory and its applications participated: Dr. A. Lewandowski, Dr. Z. Fortuna, Dr. St. Kurcyusz, Dr. I. Lasiecka. Also, discussions with professor S. Rolewicz of the Institute of Mathematics of the Polish Academy of Sciences encouraged the author to write this book. To all these colleagues the author wishes to convey his cordial thanks. Special thanks are due to the author’s family for the tolerance with which the writing of this book was encouraged.Cuaprer 1 Mathematical Models This chapter presents the general principles of mathematical model building for control purposes. The definition of a model is given and discussed, particular attention being payed to the concept of a purpose-oriented model, selected for a particular problem for which it is to be used, as distinguished from a cognitive model. The typology of purpose-oriented models for control purposes results from a multi-layer classification of control problems. Next, in Section 1.2 several examples of model building for control processes are presented. These examples justify establishing several basic types of mathematical models whose main properties are discussed in Section 1.3. Section 1.4 reviews model identification and verification methods, while Section 1.5 summarizes the present chapter, discussing a general abstract form of the model of a control process. 1.1 Prerequisites of model building A large part of research or technical activity of man is related to mathematical model building. Every physical or chemical law, every industrial regulation or standard can be considered as a model of reality. However, the concept of a model has crystallized only in the last forty years, mainly due to the development of cybernetics, control theory, and computer science as well as to a widespread use of mathematical models for simulation and modelling of various processes in operational research and system analysis, In 1945, Norbert Wiener defined a model as a representation of a process or a system (existing in reality or planned for implementation) which reflects the essential properties of the process or system in a purpose-oriented form. The above definition contains a number of points which require some comment.vi MATHEMATICAL MODELS CH, 1 First of all, the concept process or system is treated as a primary concept. There are many definitions of the concept of a system; here, we shall understand a system as a set of elements of arbitrary nature together with the relations or linkages between these elements, considered from the point of view of their common properties or purposes; the particular properties or purposes defining a system and its elements should be pre- determined. For example, if we consider an automatic control system of a distillation column in an oil refinery, we are concerned with all the facilities related to the control of the processes occurring in the distilla- tion column and with the relations or linkages between those facilities. However, those facilities and relations differ depending on the technological system design. Hence, a list of system elements and relations between them might be considered as an essential part of the definition of a system. On the other hand, a more holistic understanding of the concept of a system might be preferable in the stage of analysis and design: the elements of a system and the relations between them are then less important, while the purpose and scope of analysis are paramount. For example, the distillation column itself may be considered as an element of the system or as an outside element only related to the system, the latter depending on the properties of the system which are considered at a given moment. When considering the cost of the control system we have in mind the cost of the automatic control equipment which may amount to, say, 30% of the cost of the whole technological equipment. If we consider the dynamic properties of the control system, then the analysis must include the properties of the distillation column as a part of the system, since the properties of the control equipment cannot be considered separately. By the term process we understand all what occurs in a given system in relation to the purpose of the system. For example, consider the process of crude oil distillation that takes place in a distillation column. It is a fairly complex process and a number of physical and chemical laws must be included in the process model. The processes that occur in a control system for this column are even more complicated. When modelling them, not only some essential parts of the distillation column model should be considered, but also models of complex control facilities must be taken into account. It should be stressed here that the terms system or process are used in the literature [55] often in a somewhat different, more abstract sense, corresponding to the concept of a dynamic model of a system or a process;PREREQUISITES OF MODEL BUILDING 3 this will be discussed in Section 1.5. However, in the present book we shall understand those terms in the above given, more colloquial sense, often making a clear distinction between the concept of a real process and the concept of its model. It was stressed in the definition of a model that the latter should represent the essential features of the modelled process. Thus, a model is a reflection of reality, like every scientific law. However, a purpose-oriented model is not simply the sum of knowledge on a particular process, a set of all physical, chemical, etc. laws, pertaining to the phenomena occuring in this process. A model should represent only that part of knowledge on a process which is significant from the point of view of the purpose of the model. For example, if we build a model of a chemical process in order to control its behaviour, we do not need to consider the complete theory of the process catalysts; it is sufficient to know how does the reaction rate depend on the catalyst concentration. Thus, only those features of the process are significant which must be considered when trying to achieve the goal for which the model has been built and not those which could be included in view of our knowledge about the processes occurring in the model. The simpler model we build, the better, as long as it represents all the essential features of the process under consideration. Of course the judgement whether a model includes all the essential features and the compromise between the model simplicity and its correct reflection of the process features are often a matter of intuition of the model builder, and only in quite particular cases can be evaluated by a strict mathematical analysis; the sensitivity analysis is a tool facilitating such an estimation. The requirement of model simplicity and veritability refers clearly to purpose-oriented models only, meant for specific applications. Cognitive models are another class of models whose purpose is to represent the entire knowledge about a process in consideration and to help to analyse this process under many aspects, part of which may prove, as a result of this analysis, unimportant. Most scientific laws may be considered as cognitive models, Like any model, cognitive models may have different areas of application. Newton’s Laws of Motion are undoubtedly important cognitive models and also good purpose-oriented models for many applications. The laws of relativistic mechanics are, no doubt, more precise cognitive models, but the range of their application as a purpose-oriented model is not too wide yet, for we seldom deal with velocities in the light velocity range. However, even the relativistic laws of mechanics are not the best4 MATHEMATICAL MODELS CH. 1 possible model of reality, and therefore even from the cognitive point of view they should not be ranked as absolute}. A cognitive model is in a sense also a purpose oriented model; it is built to allow better under- standing of reality. Thus, may be, the best definition of the notion of a model is the following, rather abstract one: a model is a reflection of a part of reality constructed to serve a given purpose. Last but not least, any definition of a model stresses explicitely or implicitly the utility of its form with regard to its purpose. Various classi- fications of the form of a model are possible. A model can be of conceptual, physical or mathematical nature [31]. The conceptual model represents a certain logical proposal of ordering the significant features of the process (structural models discussed in the Introduction are examples of conceptual models); interpretation of those features by the physical model is consistent with the physical nature of the process; and finally the mathematical model represents the model features in the form of mathematical relations. However, the terms: physical model and mathematical model are also used in a different sense. A physical model (also called hardware model) is a copy, in a reduced scale, of the system or process in question; since a change of scale brings about various deformations, some elements of the system might be replaced by other elements representing the main features of the former ones. The important point is that all elements of the system be represented by actual physical devices. A mathematical model (software model) is a set of mathematical relations describing the system. However, at the present stage of development of computers and simulation techniques, this distinction is rather vague. While using an analog or digital computer to simulate a given system, it is difficult to judge what is more important: the definition of mathematical relations and the programming of a model (software) or the fact that those relations are implemented by specific physical facilities (hardware). Similarly, at the present stage of computer technology it is not of utmost importance whether a mathemat- ical model is solved on an analog computer (analog model) or a digital + The famous discussion of Niels Bohr with Albert Einstein is here a classical example. A group of physical scientists indeterminists, headed by Bohr, built a probabilistic model of an atom providing it with an interpretation that the universe is indeterministic in its very nature. The position assumed by Einstein is best represented by his statement which in brief was as follows: No doubt, you have built an elegant theory. However, to assume that the universe is indeterministic means to foreclose generic concepts of those researchers who will come after us and build still better theories.PREREQUISITES OF MODEL BUILDING 5 computer (digital model) even though both these types of model implemen- tation have their advantages and disadvantages. Many other forms of model classification could be considered. A model can be of a structural, disaggregated nature, if the models of system elements and relations between them are distinguished or it can be of an aggregated nature, if it pertains to the system as a whole. A model can be dynamic or static, linear or non-linear, deterministic or stochastic. Regardless of the detailed classification, the important point is to suit the model form to the goal which it has to serve. For cognitive models it is of course difficult to discern any preferable form. For purpose-oriented models their specific purpose often determines the useful form of the model. The above comments on model definition imply that there are no general rules for model building, as there is no general rule for the development of science. Model building is an art and cannot be formalized. However, in the case of purpose-oriented models with a specific goal, it is possible to give some methodical advice. The models which are considered in this book are meant for analysis and synthesis of control systems. By a control system we shall understand both the control object (plant), that is a system or process, subject to control, as well as the control equipment, that is, all control devices, sensors, actuators, etc. that help to implement the control and to influence the processes in the control object so as to achieve their desired course. By control system synthesis we shall understand the choice of control devices and the determination of their functions and algorithms (also called control laws), account being taken of various properties of the object and the purpose of the system or the control goal. Analysis of a control system means simply that we investigate its properties. In order to perform the synthesis or the analysis of a control system it is necessary to define a model of the control object first; undoubtedly that will be a purpose-oriented model. A preferred type of such a model is a set of mathematical relations; however, more specific requirements regarding the model form depend on the system purpose and the type of the control goal. The hierarchy of control system purposes, given in [37], is referred to as multi-layer structure of the system. We distinguish the following typical control system purposes: 1, stabilization, 2. optimization.6 MATHEMATICAL MODELS CH. 1 3. adaptation, 4. planning, coordination and management. The control system purposes denoted by smaller numbers are of a more basic significance and of higher implementation priority, e.g. they are difficult to optimize without stabilizing the processes in the control object. However, the purposes denoted by smaller numbers correspond to lower layers of the control system (Fig. 1.1) because models used in the synthesis of higher layers should take into account the modifications of properties of the control object which result from the implementation of the lower layers. 4.Planning coordination and management layer 3.Adaptation layer | 2. Optimization layer 7, Stabilization layer Control object (controlled process) Fig. 1.1. Multilayer structure of control goals. Moreover, the basic types of control purposes have several variants related to the basic types of processes occurring in the controlled objects. We distinguish the following types of processes: 1. stationary processes of long duration (sometimes called continuous), 2, cyclic processes, and 3. individual, non-repeatable processes. This classification reflects the variability of a control goal and of the disturbance affecting the control system. If the processes occurring within a control object are of a stationary nature, as in many modern continuous technological processes, e.g. in distillation columns of oil refining industry, or in continuous diffusers in beet sugar factories, then we are dealing with a Jong duration (or continuous) process. Any control processes, also those of long duration type, are influenced by disturbances which vary in time; however, we assume that the stochastic nature of these disturbancesPREREQUISITES OF MODEL BUILDING 7 does not undergo major changes in time for processes of long duration. If the disturbances, the control goals or the control processes themselves are characterized by a cyclic variability, then we classify them as cyclic processes. Cyclic disturbances are typical, for example, in the problem of supply of natural gas to large cities; cyclic control goals occur, for example, in the process of drawing of mono-crystals; finally, the process itself can be of a cyclic nature, for example, the melting of a high grade steel in an arc furnace. However, there are also processes where it is not sufficient to take into account the stationary or cyclic character of disturbances, control goals or of the process itself, and individual events have also to be considered; such processes are classified as individual ones. For example, the landing of an aircraft is a typical process, but must be conducted safely at various loads of the craft, various air densities and weather conditions; it should be considered therefore as an individual process. Set value of output variable Yo Controller i £fRegulator|— Control Controlled i Gt (measured) 4X Control object variable output variable ee Fig. 1.2. Structure of a typical closed-loop control system. Despite the diversity of the types of control objects and of the types of variability of the processes, the basic control goals are of a fairly general nature. The goal of stabilizing a process of long duration is the most natural one: the variations of basic variables in the process should be kept as low as possible (e.g. the temperatures at different points of a continuous beet-sugar diffuser should not change too much with time) despite the disturbances occurring. Such a goal is called a set-value control goal; typically, this goal is attained in a feedback system called the closed-loop system (Fig. 1.2). However, the stabilization goals for the cyclic or individual processes are attained by similar means. In a cyclic process the goal of stabilization can be attained by maintaining the desired program of changes of the basic variables in the process (e.g. the program of changes of the current in a steel melting arc furnace) in spite of disturbances; we are8 MATHEMATICAL MODELS CH. 1 dealing then with a program-following control system. In an individual process the basic variables often must follow an unrepeatable program; for instance, if a pilot chooses a certain landing route, he will try to follow that route despite incidental gusts of wind. The stabilization of individual processes is often related to the requirement of following a program which has not been set in advance; that is why we use here the term follow-up control. The goal of control optimization is somewhat more dependent on the nature of variability of the processes. For processes of long duration, optimization consists usually in selecting optimal, constant values of the basic variables in the process; these variables become the goal for the lower layer of the set-value stabilization control (Fig. 1.3). We observe Optimization layer {Yor } Yon ; * Stabilization layer Control process Fig. 1.3 Typical structure of a set-value optimization system. here a strong influence of the type of goal on the form of the process model useful for attaining this goal. A model of process dynamics is essential for the set-value stabilization control (as otherwise the design of a closed- loop control system might be false); however, a linear dynamic model is often sufficient since the variability range of control variables is usually small. However, for optimization of the set values we require a model of completely different type. A static model related only to the steady state of the process can be used, but it must take into account the non-linear character of the dependence of the performance index on the basic variables in the process that has a larger variability range. In the literature on the theory of optimal control [5, 71, 85] much attention is often given to the goal of dynamic optimization in set value stabilization control, that is, to the problem of an optimal functioning of the process of disturbances control. This problem plays an importantPREREQUISITES OF MODEL BUILDING 9 cognitive role, however, in practice the static optimization of the set values is much more important. In cyclic processes the optimization problem is usually related to the determination of the optimal time-varying program for basic variables of the process. The process models used when solving this problem are, obviously, of a dynamic character, but also in this case they usually differ from the models used in the problem of program control, be it j of incorporating in the optimization model the changes in the process dynamics resulting from the fact of applying a stabilizing control system. Similar remarks are pertinent to the optimization of individual processes. The goal of adaptation is, generally speaking, to fit the model and the control system to the variable properties of the controlled process. This goal can be attained either through measurements of some disturbances, or through model identification during the process under the assumption that the changes of process properties are sufficiently slow. Adaptation of dynamic models constructed for stabilization purposes is significant for individual processes, somewhat less so for cyclic processes; as regards processes of long-duration, there is usually no need (with some exceptions) to adapt the dynamic model. However, adaptation of static models for optimization of set values in processes of long-duration, as well as of other optimization models, is of great importance. The goals of planning, co-ordination, and management are usually related to more fundamental changes in control methods, control goals, etc. When solving a problem of this type, it is impossible to use complete models of the process describing both the dynamics of the stabilization systems and the variability of the optimization models. Other, more synthetic models have to be constructed here in view of the specific nature of the planning or co-ordination goals. Thus, the purpose-oriented models of controlled processes, constructed with a view to apply them in the synthesis of a control system, are closely fitted to a particular type of control purpose. For the purpose of stabilization, linear dynamic models are usually used, describing the control object satisfactorily for small deviations from the normal course of the process; they must, however, represent fairly exactly the dynamics of the controlled process. For the purpose of optimization, stress is put on significant non-linearities of the model, the dynamic properties being neglected or simplified. In the case of adaptation, we are interested only in that part of the model which contains parameters subject to identification or which change under the10 MATHEMATICAL MODELS CH. 1 influence of external factors. In the case of planning, co-ordination and management the models must be much more synthetic. The above considerations on building mathematical models will be best illustrated by specific examples. 1.2 Examples of model building Beet-sugar diffuser. Consider the problem of control of a continuous beet-sugar diffuser shown in Fig. 1.4. The technological process in the diffuser consists in sweetening off the beet slices by water extraction when the sugar passes into the diffusion juice. The slices are fed to the diffuser by a belt conveyor, and transported in the diffusor upwards by a worm conveyor countercurrent to the water fed to the upper part of the diffuser. beet pulp Water inlet Beet slices f ao Worm yeyor convey! ‘a pipfusion jute? tl jacket oul ‘ ¥ \ to heating * steam supply * Fig. 1.4 Diagram of a continuous beet-sugar diffuser. Water comes first into contact with the sweetened off slices, flows down the diffuser and leaves its lower part as the diffusion juice. The purpose of control here is to maintain the optimum conditions of sweetening off the slices and obtain a juice of possibly good properties. If the beets were completely squashed or overboiled (and then further squashed by the worm inside the diffuser) or very large amounts of water were added to the diffuser, complete sweetening off of slices would be achieved. However, were the beets completely disintegrated or overheated, excessive amounts of cellulose and other organic pollutants would pass to the juice, and the necessary purification of the juice would cause considerable sugar loss. On the other hand, excessive water supply to the diffuser would causeEXAMPLES OF MODEL BUILDING i increased power consumption for heating the diffuser as well as energy loss in the stage of juice evaporation. The process in the continuous diffuser is a typical long duration process. The disturbances in that process are incidental being mainly due to the varying properties of the beet slices and changes of temperature of the heating steam. The amount of slices fed into the diffuser (the diffuser capacity) varies only in accordance with the co-ordination decisions, but is constant under normal operating conditions. However, maintaining normal operating conditions of the diffuser requires stabilization of several basic variables: the slice volume to input water volume ratio, the rotations of the worm conveyor, the level of juice in the lower part of the diffuser (stabilizing the flow rate of the juice through the diffuser) and the distribution of temperature in the diffuser. Consider, for example, the problem of stabilizing the temperature distribution. The latter can be influenced by controlling the heating steam supply (shutting or opening the steam supply valves) to several chambers constituting the diffuser heating jacket. The heat is next transferred through the diffuser coat to the mixture of slices and juice where it is dispersed mainly due to the transport of slices and juice by the worm conveyor. Designing a possibly accurate model of the diffuser heat dynamics would require determining the partial differential equations (with respect to time and diffuser length and, more precisely, with respect to the radial distance from the diffuser surface also) representing the heat balance in a unit volume of the mixture in the diffuser and, additionally, an equation of heat conductance through the diffuser jacket. However, to determine these equations, also the partial differential equations of slices and water transport through the diffuser are required. The latter equations are very difficult to determine, since the resistance of the slices to the worm conveyor rotation and flow of water is not linearly dependent on the temperature of the slices, and that dependence has not been explained as yet. As a result, the problem of determining a model of the heat dynamics of the diffuser by an analytic method is practically unresolvable. Still, synthesis of the temperature stabilization system requires some model of the object dynamics, be it even a very simplified one. The following simplifications are permissible here: consider a linear model only, for small deviations from the nominal values; consider a model with parameters constant in time only (for example, constant speed of slices and water transport in the diffuser, since it is known that changes of the rotary speed12 MATHEMATICAL MODELS CH. 1 of the worm produce changes in the heat dynamics); it is also possible to take advantage of the fact that temperature measurements are conducted at several points along the diffuser only, so as control is conducted by adjusting several control valves only. Thus it is permissible to assume a very general structural scheme of a model (Fig. 1.5), where only the controlling inputs, measurement outputs and operators characterizing the relations Steam supply | uy Temperature to Ist chamber | in fst chamber ! | | 1 | | Steam supply | up| .—> to 2nd chamber) Tomperature in 2nd chamber Ssals) SzafS) | | | | ! | Steam suoaly | ua to 3rd chareber| Temperature Gal) in 3rd chamber [ee |-/ | i i | Gaal} I Steam supply | ue ye [Temperature z, Bea} He mera, aie ? to 4th chamber’ ico 4th chambe! Fig. 1.5 Structural diagram of a diffuser heat dynamics model. between those inputs and outputs are distinguished. According to the simplifying assumptions, these are linear, stationary operators, so they can be represented by transfer functions G;,(s). There are, however, few premises as to the specific form of these transfer functions; nevertheless it is known [7] that for transport processes the transfer functions of an + See Section 1.3,EXAMPLES OF MODEL BUILDING 13 element with delay e~*7* are typical, whereas for mixing processes inertial element transfer functions 1/(7s+ 1), are representative. Thus, a hypothesis can be made that Gij(s) = kije~*74/(Tijs+1) where: kj is the gain coefficient, T,;—the delay time constant, T;;—the inertia time constant, are parameters of a dynamic model describing the influence of the j-th input on the i-th output. This hypothesis proves to be quite acceptable in practice. When introducing a step change in the position of the j-th control valve and recording the temperature responses (called step-responses) at the successive outputs, we obtain the functions shown in Fig. 1.6. This diagram shows also step-responses of elements with transfer functions k,je7'Tou/(Tjjs-+1) where the parameters are chosen to fit the original step-responses; the manner of selection of these parameters is also indicated. U2 0 30 60 : 90 min Fike arom Dynamic effects negligible Ie eee 0 30 60 90 mn Fig. 1.6 Temperature responses in the successive parts of the diffuser to step changes in steam supply to the second chamber of the heating jacket.14 MATHEMATICAL MODELS CH. 1 The model of heat dynamics of the diffuser obtained in this way, with its parameters estimated, is not very exact of course. However, models of this type are sufficient for the synthesis of stabilizing control systems; the reason for this will be explained in Section 2.3. In a similar manner models for controlling the level of juice in the lower part of the diffuser or the supply of water to its upper part can be constructed. The purpose of diffuser control encompasses not only stabilization but also optimization of its operation. As in many optimization problems, the first question is to choose a performance index of diffuser operation and to determine the model of dependence of that index on the basic variables in the process, A performance index with economic interpretation can be accepted to account for the loss associated with the percentage of sugar content P, in the sweetened off pulp, removed from the upper part of the diffuser, the cost of water evaporation from the diffuser juice with water percentage J,, and the cost of purifying the juice with a percentage of non-sugar components J, (the greater the non-sugar components content J,, the larger is the sugar loss during juice purification) Q = (4, P, +425, +43 5,)T ay where: Q is the performance index, T—the diffuser throughput in tons of beet slices per hour, a1, a2, a;—the technical-cconomic coefficients. The determination of coefficients a,, 4,, 43, is already a fairly difficult task. However, the basic difficulties are connected with the determination of the dependences of the P, and J, values on the main variables of the process, even under strong simplifying assumptions. The process in the diffuser is a typical process of long duration and its variables vary incidentally about their average values. Therefore the average values, e.g. hourly average values, of the performance index Q should be optimized with respect to the average values (or, what is equivalent in correctly operating stabilization systems, with respect to the set values) of the variables of the process that have the greatest influence on the performance index. These variables include: the diffuser throughput T, temperatures ¥ = (X,,X2,X3, X4) along the diffuser, and the slice } The structural diagram in Fig, 1.5 is already simplified, since not only the effect of steam supply to a given jacket chamber on the temperature of the neighbouring chambers should be considered, but also that on the temperature of more distant chambers. A different structure of the model can also be assumed, where not only the effect of steam supply on temperature is considered but also the heat transmission effects occurring between the successive chambers.EXAMPLES OF MODEL BUILDING 15 volume to input water volume ratio Y. Thus it is essential to determine the static dependence of the average values of P,, J,, and J, on T, Y, and X. Analytically these dependences cannot be determined with sufficient accuracy. There are only certain experimental data that reveal the nature of these dependences. For example, the experimentally found dependence of P; (pulp sugar content) on the temperature X, in the first part of the diffuser, while assuming average values of the other factors, is shown in Fig. 1.7. In the neighbourhood of the extremum this dependence can be approximated by a polynomial. Thus a hypothesis can be made that the whole dependence of P,(T, X, Y) may be approximated by a polynomial PAT, ¥,X) = bo thy, T+bo,T? +b3,T8 +... + tbiy¥+bay¥?2 + bsyY24 0 + Dyer Xp tbo Xi tbs XGH o. + +Dpy TY +851 YX +». a2) 4 SE eee eee ea see Frege 60 70 80 90 °C Fig. 1.7 Averaged dependence of the pulp sugar content on the temperature in the first part of the diffuser. If a sufficient amount of reliable experimental data are available, then the determination of coefficients by, by,, ... and estimation which of them may be neglected constitutes a typical problem of regression analysis (Section 1.3), However, the basic technical difficulty is to gather reliable data on the values of sugar loss in the pulp, P,. Similar arguments hold when building a model for the contents of solid non-sugar components J, in the diffuser juice and the experimental difficulties encountered are similar. So far P, and J, can be measured only in a laboratory and not with16 MATHEMATICAL MODELS CH. 1 automatic measuring equipment. Samples of juice and pulp should be collected so that they represent average hourly values, and not the instantaneous values. Furthermore, when building static models P,(T, Y, X) and J,(T, Y,X), the dynamics of the object cannot be fully disregarded; the sugar contents in the pulp depends, after all, on the temperature in the first part of the diffuser when the beets are in it, and the time of beet transport through the diffuser is of the order of one hour. To determine the dependence shown in Fig. 1.7, a delay of one hour between X, and P, was assumed; to determine the parameters of the whole models P,(T, Y,X) and J,(T, Y, X), it is necessary to find first what delay is to be assumed for the particular quantities. The above factors make difficult the problem of building a diffuser static model for the purpose of optimiza- tion, and it has not been solved in practice with satisfactory exactness so far. Had the parameters of the models P,(T,Y,X), J,(T, ¥,X) and J,(T, Y, X) been determined, then the calculation of optimal values would be fairly simple, and the optimization of the diffusion process possible, The optimizing control system would then simply consist in automatic collection of data, up-dating the models based on thesc data, determination of the optimum values of 7, Y, X and their application as set values for stabilization systems. This would be an optimization system with necessary adaptation of the model to the slow changes of properties of the processed sugar beets. Another variant of the system, where provisions are made for co-ordination of the whole sugar plant, reserves the decisions regarding the throughput 7 for the co-ordination layer, while the optimum values of Y¥, X = (X,, X2,X5, X4) are selected at a given value T. ‘Steel-melting arc furnace. Consider another example of a controlled process: melting of high grade steel in an arc furnace. A diagram of the furnace with a power supply transformer and electrode servomechanism systems is shown in Fig. 1.8. The cycle of furnace operation is as follows: after tapping steel, when the electrodes are raised and the roof is removed, the furnace is charged with scrap iron using a basket transported by a crane. Next, the scrap is melted in the electric arc between three electrodes, fed from a three phase transformer. Contemporary furnaces have a capacity of 10 to 100 tons of scrap, power supply of 10 to 100 megawatts, and arc current of the order of 10 to 100 kiloamperes; the melting timeis 1-3 hours. After the scrap is melted, oxidizing agents are added, and the bath is slightlyEXAMPLES OF MODEL BUILDING 17 heated to burn out the carbon contained in the scrap. Next slag is removed, alloy components are added and steel is refined with slight heating. After the oxidation and refining steps, the electrodes are raised again and the steel is tapped by tilting the furnace. Pi in sw & Hh & bs & & Fig. 1.8 Diagram of a steel melting arc furnace: 7r—transformer, PW—primary winding, SW—secondary winding (with spurs), HA—high ampere line (flexible cables to electrodes), M—motor and electrode drive transmission, ES—electrode servomechanism, E;, E,, E3;—electrodes, R—roof. Control of the steel melting process in arc furnaces may have different targets. First, the aim of control is to assure a stable supply of energy at a specified rate or, in other words, to stabilize the working conditions of the electric arc in spite of the melting and sinking of scrap during melting and in spite of the varying resistance of the arc in the further steps. Secondly, the optimization of the melting step, i.e. selection of such a program of power supply that will allow the least energy consumption for melting the scrap in a possibly short time, may be the aim of control. In the oxidation and refining steps the power supply program results from the technological requirements; optimization of these steps may be connected with the expected chemical composition of steel and specification of oxidizing and alloy additives necessary for achieving the greatest possible proportion of good melts, i.e. such that conform to the requirements regarding the composition of the steel. Co-ordination and planning of work in an electric steel mill including several arc furnaces can be considered as another control purpose. The melting cycles of the particular furnaces have to be suitably shifted in time to assure the best operation efficiency of the overhead cranes in the scrap preparation yard and in the pouring bay, a proper rythm of work and elimination stoppages of furnaces, due to waiting for charging or tapping, as well as the even spread of mains18 MATHEMATICAL MODELS CH. 1 power consumption by appropriate distribution in time of the melting steps requiring greatest power consumption. Various types of mathematical models correspond to those varied control purposes. Consider first a model of the dynamics of the electrode servomechanism necessary for the stabilization of power supply to the furnace. A diagram of such a system (for one electrode) [86] is shown in Fig. 1.9a with the provision that the electrode has an electric drive (in modern designs hydraulic drives are also often used; their dynamic models being slightly simpler). The electrode E is driven by the motor M via a transmission PR. The motor is controlled by a rotary amplifier A (amplidyne). This amplifier has two feedbacks: the proportional P and the differential D ones. The input winding of the amplidyne is controlled by the magnetic amplifier ay yt of the amplidyne-motor system without feedbacks (b).EXAMPLES OF MODEL BUILDING 19 MA whose inputs, in turn, are: the controlled variable signal proportional to the are current J, the set value signal proportional to the voltage of furnace transformer U, and a corrective signal from a tachometric generator 7. The basic static dependence in this system is the variation of the arc current J with the position of electrode EZ; however, for small deviations from the rated current this dependence can be approximated by a linear one, specifying only the ratio of the current increase to the change of electrode position. The dynamic properties of the system result mainly from the inertia of the motor, transmission and electrode, and from the dynamics of the amplidyne. An equivalent diagram of the amplidyne-motor system is shown in Fig. 1.9b, where the additional feedbacks are not shown for the sake of simplicity. The linearized equations of the amplidyne dynamics are relatively simple I, qT, Shy =—U, U,=khh Tea dt dl, 1 L, on Tr tha Re Usalah T= Re where /, is the difference of currents controlling the amplidyne, R;, L;—the resistance and inductance of the amplidyne control circuit, J,—the current in the transverse winding of the amplidyne, R,, L,—the resistance and inductance of the transverse winding, k,,k,—the gain coefficients, U;— the output voltage of the amplidyne. The non-linearities deforming these equations result mainly from the non-linearity of magnetic circuits in the amplidyne and can be approximated in first approximation} by the relations Uz =f,(/;) and U3 = fo(h) instead of the linear relations U, = k,J, and U; = k2J,. However, the resultant relation U; = f,(f,(J)) can be approximated by a straight line; the insensitivity zone resulting from this characteristic is small and the saturation of this characteristic determines only the maximum voltage U3 and maximum rotational speed wm of the motor (see Figs. 1.10a, 1.10b and 1.10c). The moment of friction in the motor under load has a more significant effect on the characteristics; the fundamental equations of the motor dynamics have the form + To achieve a more exact model, the non-linear dependence of magnetic induction on current should be fully accounted for, which also means accounting for the dependence of the inductances L; and L, on the currents J, and J,, what would have a somewhat stronger effect on the form of Eq. (1.3).20 MATHEMATICAL MODELS CH. 1 do I + My = nls (1.4) di, 1 Tag th = RE (skew) Ta = Z where: J; is the current of the rotor, #—the rotational speed, J—the moment of inertia of the motor load (including the load due to the transmission and the electrode), M,—the sum of friction moments, Lm, Rm—the inductance and resistance of the rotor circuit, k,,—the coefficient of the moment generated by the motor, k.—the coefficient of the electromotive force induced in the motor. \Eiact Z <_2nd_agpraximation I Fig. 1.10 Static characteristics of the amplidyne-motor system: (a) characteristic of the amplidyne, (b) characteristic of the friction moment in the motor under load, (6) resultant static characteristic of the system, The dependence of the moment of friction My on the rotational speed of the motor and its simple approximation are shown in Fig. 1.10b. When analysing relationships (1.4) for specified rotational speeds (under the assumption that dw/dt = 0 and di//dt = 0) it is easy to prove that, in first approximation, the moment of friction produces only an increase in the insensitivity zone of the amplidyne-motor system (cf. Fig. 1.10c). A fairly detailed model of the amplidyne-motor system dynamics is shown in Fig. 1.11a, where relationships (1.3), (1.4) and the non-linearity of the magnetic characteristics of the two circuits of the amplidyne and the non-linear characteristic of friction are taken into account. However, such a model is too complicated for the purpose of synthesis of a motor control system, even though it may be used for analysing its properties by simulation on a computer. In order to simplify this model it can be assumed that all non-linearities of the model can be approximated by one non-linear static characteristic at the system input}; this characteristic can be determined by analysing the steady-state properties of the model.