You are on page 1of 181
Model Order Reduction and Controller Design Techniques Dr. §. Janardhanan Contents 1 Introduction to Large Scale Systems 1.1 What are Large Scale Systems ? 1.2 Hierarchial Structures 1.3 Decentralized Control 1.4. Large Scale System Modeling 2 Large Scale System Model Order Reduction and Control - Modal Analysis Approach 2.1 Introduction 2.2 Davison Technique 221 222 2.23 2.24 2.2.5 Reduced Order Model Using Davison Technique Alternative Method to Obtain Reduced Order Model through Davison Technique Improved Davison Technique Suboptimal Control Using Davison Model Control Law Reduction Approach Using Davison Model 2.3 Chidambara Technique 23.1 23.2 2.3.3 Reduced Order Model Using Chidambara Technique Suboptimal Control Using Chidambara Model Control Law Reduction Approach Using Chidambara Model 2.4 Marshall Technique 2.4.1 Reduced Order model by Marshall Technique 2.5 Choice of Reduced Model Order 25.1 25.2 Model Order Selection Criterion by Mahapatra, Another Criterion for Order Selection and Mode Selection 3 Model Order Reduction and Control - Aggregation Methods 3.1 Aggregation of Control Systems 3.1 312 Properties of Aggregated System Matrix Error in Aggregation 3.2. Determination of Aggregation Matrix 3.3. Modal Aggregation 33.1 33.2 Reduced Order Model Stability of Feedback System 3.4 Aggregation by Continued Fraction aay 13 u u 15 15 16 7 7 18 18 20 23 23 24 26 26 28 28 30 32 4 Large Scale System Model Order Reduction - Frequency Domain Based Methods 37 4.1 Moment Matching 37 4.2. Padé Approximation Methods 40 4.2.1 Padé Approximation Method for SISO Systems 40 422 ModalPadé Method 42 423 Padé Approximation for Multivariable Systems in Frequency Domain 43 4.24 Stable Padé for Multivariable Systems 45 4.2.5 Reduction of non-asymptotically stable systems 46 4.2.6 Time-Domain Padé Approximation for Multivariable Systems a7 42,7 Time-Domain Modal-Padé Method 51 43° Routh Approximation Techniques 53 4.3.1 Routh Approximation Method Using a — 4 Parameters 53 4.3.2 Routh Approximation Technique Using y~ 6 Parameters 54 41.3.3. Aggregated Model of Routh Approximants 56 4.3.4 Optimal Order of Routh Approximant, 59 44 Continued Fraction Method 60 4.4.1 ‘The Three Cauer Forms 60 44.2 A Generalized Routh Algorithm 62 4.4.3 Simplified Models Using Continued Fraction Expansion Forms of 5 Large Scale System Model and Controller Order Reduction - Norm Based Methods, 67 5.1 Introductions 67 5.1.1 Norms of Vectors and Matrices 67 5.12 Singular Value Decomposition 68 5.13 Grammian Matrices and Hankel Singular Values 70 5.1.4 Matrix Inversion Formulae 7 5.2, Model Reduction by Balanced Truncation 73 5.2.1 Balanced Realization 3 5.2.2 Balanced Truncation m4 5.2.3 Steady State Matching 75 5.2.4 Reduction of Unstable Systems by Balanced ‘Truncation 6 5.2.5 Properties of Truncated Systems 76 5.2.6 Frequency-Weighted Balanced Model Reduction al 5.3 Model Reduction by Impulse/Step Error Minimization 82 5.3.1 Impulse Error Minimization 83 5.3.2 Step Error Minimization 88 54 Optimal Model Order Reduction Using Wilson’s Technique 90 54.1 Impulse Error / White Noise Error Minimization 90 6 Pole Placement Techniques 95 6.1 Introduction 95 6.2 What Poles to Choose ? 95 6.2.1 General 95, 6.2.2 Interpretation of Responses from Pole-Zero Locations 96 6.3. Pole Assignment in Single Input Systems 7 6.3.1 State Feedback 97 6.32 Optimal State Feedback ( Brief Introduction to LQR ) 99 6.3.3 Static Output Feedback in Single Input Systems 102 6.3.4 Dynamic Output Feedback ( SISO Case ) 103 64 Pole Assignment and Placement in Multi-Input Systems 104 6.4.1 Concepts of Multivariable Systems 104 642 State Feedback 107 64.3 Design of Optimal Control Systems with Prescribed Eigenvalues 108 64.4 Static Output Feedback 116 64.5 Two Time-Scale Decomposition and State Feedback Design 123 Fast Output Sampling (FOS) 127 7.1. Introduction 127 7.2 Controller Deduetion 127 7.3. Closed Loop Stability 129 7.4 Techniques for Determining Fast Output Sampling Controller Gain 130 TAA ‘Two Time-Scale Approach for Conditioning of State Feedback Gain F130 74.2 Singular Value Decomposition of Measurement Matrix C 132 74.3 Approach for Multi-Plant Systems 132 7.5 An LMI Formulation of the design problem 133 7.6 A Modified Approach for Fast Output Sampling Feedback 133 Periodic Output Feedback (POF) 135 81 Review 135 8.2. Periodic Output Feedback Controller Deduction 135 8.3 Multimodel Synthesis, 137 Robust Control of Systems with Parametric Uncertainty 139 9.1 Concepts Related to Uncertain Systems 139 9.1.1 Interval Arithmetic 139 9.1.2 Hermite-Bicler Theorem 140 9.1.3 Kharitonov Theorem 142 9.14 Gerschgorin Theorem 145 9.1.5 Simultaneous Stabilization of Interval Plant Family Based on Kharitonov ‘Theorem 146 9.2 Bhattacharyya’s Method 7 9.3 Jayakumar's Method 154 9.3.1 Routh Table Based Kharitonov Algorithm in Controller Design 154 9.32 An Alternative Proof for Existence of a Simultancously Stabilizing State Feedback for Interval Systems 156 9.4 Smagina & Brewer's Method 157 9.4.1 The Problem Statement, 187 94.2 Main Results 158 9.4.3 The Algorithm 9.5 State Feedback for Uncertain Systems Based on Gerschgorin Theorem Numerical Problems in Large Scale Systems Al Davison, Chidambara and Marshall Techniques A2 Routh and Padé Approximations AB. State and Output Feedback Design Ad Periodic Output Feedback and Fast Output Sampling A.5 Uncertain Systems 161 163 a7 171 172 173 174 175 Chapter 1 Introduction to Large Scale Systems 1.1 What are Large Scale Systems ? A great number of problems are brought about by the present day technology and societal and environmental processes which are highly complex and "large in dimension and stochastic by nature’. The notion of ‘large scale’ is highly subjective one in that one may ask : How large is large?”. There has been no accepted definition for what constitutes a large scale system. Many viewpoints have been presented on this issue. One viewpoint has been that a system is considered large scale if it can be decoupled or partitioned into a number of interconnected subsystems or small scale systems for either computational or practical reasons. Another viewpoint is that a system is large scale when its dimensions are so large that conventional techniques of modeling, analysis, control, design and computation fail to give reasonable solutions with reasonable computational efforts. In other words a system is large when it requires more than one controller Since the early 1950s, when classical control theory was being established, engineers have devised several procedures, both within the classical and modem control contexts, which analyze or design a given system, These procedures can be summarized as follows. 1, Modeling procedures which consist of differential equations, input-output transfer fune- tions and state space formulations. 2, Behavioral procedures of systems such as controllability, observability and stability tests and application of such criteria as Routh - Hurwitz, Nyquist, Lyapunov’s second method ete., 3. Control procedures such as series compensation, pole placement, optimal control ete., ‘The underlying assum; trality i.e., all the calculations based upon system information and the information itself are localized at a given center, very often a geographical position. ion for all such control and system procedures has been cen- A notable characteristic of most large scale systems is that centrality fails to hold up due to either the lack of centralized computing capability or centralized information. Need- less to say, many real problems considered are large scale by nature and not by choice. 1 2 Large Scale Systems ‘The important points regarding large scale systems are that their hicrarchial (nmultilevel) and decentralized structures depict systems dealing with society, business, management, the economy, the environment, energy, data networks, power networks, space structures, trans- portation, aerospace, water resources, ecology and flexible manufacturing networks to name a few. ‘These systems are often separated geographically, and their treatment requires con- sideration of not only economic costs as is common in centralized systems but also such important issues as reliability of communication links, value of information ete.,.. It is for the decentralized and hierarchial control properties and potential applications that many researchers throughout the world have devoted a great deal of effort to large scale systems in recent years. 1.2 Hierarchial Structures One of the earlier attempts in dealing with large scale systems was to decompose a given system into a number of subsystems for computational efficiency and design simplification. ‘The idea of decomposition was first treated theoretically in mathematical programming by Dantzig and Wolfe [1] by treating large linear programming problems possessing special structures. The coefficient matrices of such large linear programs often have relatively few nonzero elements, i., they are sparse matrices. There are two bas ss for dealing with such problems, ‘coupled’ and ‘decoupled’. The coupled approa he problem's structure intact and takes advantage of the structure to perform efficient computations. The ‘compact basis triangularization’ and ‘generalized upper bounding’ are two such efficient methods. The decoupled approach divides the original system into a number of sub: involving certain values of parameters. Each subsystem is solved independently for a fixed value of so-called decoupling parameter, whose value is subsequently adjusted by a coordina- tor in an appropriate fashion so that the subsystems resolve their problems and the solution to the original system is obtained. Recently the decoupled approach has been termed as ‘multilevel’ or ‘hierarchial’ approach. Consider a two level system shown in Fig (1.1). At the first level, \V subsystems of the original large scale system are shown. At the second level ‘a coordinator receives the local solutions of the NV subsystems, $,,i = 1,2,---, produces a new set of interaction parameters a,,i = 1,2,---,.N . The goal of the coordinator is to arrange the activities of the subsystems to provide a feasible solution to the overall system, and then 1.3 Decentralized Control Most large seale systems are characterized by a great multiplicity of measured outputs and inputs. For example, an electrie power system has several control substations, each being responsible for the operation of a part of the overall system. This situation arising in a control system design is often referred to as decentralization. The designer for such systems determines a structure for control which assigns system inputs to a given set of local con- trollers(stations), which observe only local system outputs. In other words, this approach, called decentralized control, attempts to avoid difficulties in data gathering storage require- ments, computer program debuggings and geographic separation of system components, Coordinator system Figure 1.1: Hierarehial Control 4 Large Scale Systems g E Tes Ty Figure 1.2: Two Controller Decentralized Large Scale System Fig. (72) shows a two controller decentralized system. ‘The basic characteristic of any decentralized system is that the transfer of information from one group of sensors or actuators to others is quite restricted. For example, in the system of Fig. (??), only the output y and the external input 1 are used to find the control vy and likewise the control v is obtained through only the output yz and external input uz. The determination of control signals and v2 based on the output signals y: and yp respectively is nothing but two independent output feedback problems which can be used for stabilization or pole placement purposes. It is therefore clear that the decentralized control scheme is of feedback form, indicating that this method is very useful for large scale linear systems. ‘This is a clear distinction from the hierarchial control scheme, which was mainly intended to be an open loop structure. In the previous part the concept of a large scale system and two basic hierarchial and decentralized control structures were briefly introduced, Although there is no universal definition of a large scale system, it is commonly accepted that such systems possess the following characteristics 1, Large scale systems are often controlled by more than one controller or decision maker involving “decentralized’ computations 2. The controllers have different but correlated ‘information’ available to them, possibly at different times 3. Lange scale systems can also be controlled by local controllers at one level whose control actions are being coordinated at another level in a hictarchial( multilevel) structure. 4, Large scale systems are usually represented by imprecise aggregate models, 14 Large Scale System Modeling 5 5. Controllers may operate in a group as a ‘team’ or in a conflicting manner with singlo- or multiple-objective or even conflicting-objective functions 6, Large scale systems may be satisfactorily optimized by means of suboptimal or near optimum controls, sometimes termed as a satisfactory strategy. An attempt is made here to consider primarily modeling and control of large scale systems Most of the discussions are focussed on large scale linear, continuous time, stationary and deterministic systems. 1.4 Large Scale System Modeling Scientists and engineers are often confronted with the analysis, design and synthesis of real- life problems. ‘The first step in such studies is the development of a ‘mathematical model which can be a substitute for the real problem. In any modeling task, two often conflicting factors prevail - simplicity and accuracy. On one hand, if a system model is oversimplified, presumably for computational effectiveness, incorrect conclusions may be drawn from it in representing an actual system. On the other hand, a highly detailed model would lend to a great deal of unnecessary complications and should a feasible solution be attainable, the extent of resulting details may become so vast that further investigations on the system behavior would become impossible with questionable practical values. Clearly a mechanism by which a compromise can be made between a complex, more accurate model and a simple, less accurate model is needled. Such a mechanism is not a simple undertaking. ‘The key to a valid modeling philosophy is to set forth the following outline. 1, The purpose of the model must be clearly defined, no single model ean be appropriate for all purposes. 2. The system's boun parating the system and the outside world must be defined 3. A structural relationship among different system components which would best repre- sent desired or observed effects must be defined, 4. Based on the physical structure of the model, a set of system variables of interest must he defined. If a quantity of important significance cannot be labelled, step (3) must be modified accordingly. 5. Mathematical descriptions of each system component, sometimes called elemental equations, should be written down 6. After the mathematical description of each system component is complete, they are related through a set of physical laws of conservation (or continuity) and compatibility, such as Newton's, Kirchoff's or D’ Alembert's. Elemental, continuity and compatibility equations should be manipulated and the mathematical format of the model should be finalized, Large Scale Systems 8, The last step to a sucessful modeling is the analysis of the model and its comparison h real situations Chapter 2 Large Scale System Model Order Reduction and Control - Modal Analysis Approach 2.1 Introduction It is usually possible to describe the dynamics of physical systems by a number of simulta- neous linear differential equations with constant coefficients. = Ant bu But for many processes ( like chemical plants and nuclear reactors) the order of the matrix A may be quite large. It would be difficult to work with these complex systems in their original form, In such cases, it is common to study the process by approximating it to a simpler model. For instance, the response of an airplane is quite commonly approximated by a second order transfer function, These mathematical models correspond to approximating a system by its dominant pole-zeros in the complex plane, They generally require empirical determination of the system parameters. Many different methods have been developed to accomplish the purpose by estimating the ‘dominant’ part of the large system and finding a simpler (or reduced order) system representation that has its behavior akin to the original system 2.2 Davison Technique A structured approach to the model reduction problem was given by E. J. Davison in [2 The method suggests that a large (n x n) system can be reduced to a simpler (! x 1) model ( Re(A2) > --- > Re(A,), then the transformation Z = P'XxX (2.3) P= [um wm } uo = [a1 my ana)” would transform the system into Zr AZ~ Bu where, PAP, P'B and A, would be either in the diagonal or the Jordan canonical form, Truncation of non-dominant eigenvalues is simpler in this case, In this case, the state response of the system for an input Bu = B,, can be shown to be X= [Perr Bae (2.4) PO? = [ehimtngadn which would give ae xp PY abi + dub t+ babe) (25) If 1 states that need to be considered for the reduced order model are represented as a, £e2,**+,ta then the I" order reduced model can be derived as 22 Davison Technique 9 Ag+ APPS? (26) PIP B| (27) fa y — |#e (28) fat where, a Ga Bae ag = | tae Maa ct es) Qa aaa Beret r a3 aa aa aaa Geet + Aeterna Gerelet + Achetst at + Ahern (2.10) Qaast 0 Meher Gert ett aercte ett Gan Gain pr, = | 728 222 ad faa ai ta fat ma Ha ty ay aaa acai Fag Fatal p= | eeu ta Tavis (212) dar Fae Teas Bde Fdei2 Teter na End Enh 10 Large Scale Systems PB) = first U rows of PB (2.13) Derivation : ‘The principle involved in reducing the matrix is to negle the system. ‘The following equation is obtained if the first alone are retained from Eqn. (24), ie., considering only the first | terms of Eqn. (2.5), and taking into account only the I states considered (i.., equating the other states to zero), the following is obtained. ‘igher-order time constants of y-y6] 7" (2.14) = ES Coins + Giaba +--+ dintn) and therefore & & P2Y (2.15) & then Zant & tant & (216) a1 & Substituting Eqn. (2.15) in Eqn. (2.16) 22 Davison Technique iy n = PPSY (ean) Considering the equations for es, ée2,++,éa alone from Equ. (2.1), the following equa- tion can be obtained y= ayer] | 4 ea rBw (2.18) P= [ho Substituting Eqn. (2.17) in Eqn. (2.18), V = AG + APPPS) + PP Blu (2.19) 2.2.2 Alternative Method to Obtain Reduced Order Model through Davison Technique ‘The Davison Model can also be computed thus Initia the system states are rearranged in such a manner that the eigenvectors corre- sponding to the states to be retained from Eqn. (2.1) are placed first. If P is represented P an Pa [fe z| (2.20) then, if the state vector Xi (consisting of the states to be 1 artitioned into dominant and non-dominant parts as X; ained) and Xz (consisting of the states to be discarded), 12 Large Scale Systems and the state and input matrices are also partitioned in appropriate fashion, then X’ can be represented as x An An} {Xi By 2 x lS u 2.21) [3] ‘An | xl] (221) -(e)-[e eZ] e where, Z; and Z, are the states of the decoupled system representation, Thus, X= Puli + Pale, (2.23) Xp = Pad, + Puér (2.24) If the decoupled representation of the system is -[&a)_[%o]fa)yny, 2-[2]-[@ w)LZ)+[2] (2.25) where, Ay 0 1 6) [? wl POMP, (2.26) MN] pa 9 [ri] = PB. (2.27) ‘The modes in Z, are non-dominant and therefore can be ignored (according to Davison 2)). Thus, setting Z, to zero, and substituting in Eqns. (2.23 and 2.24) the following are obtained. X= Pads (2.28) X= Pak (2.29) Xp = PaPalX (2.30) and from Eqn. (2.21) AuX) + AzX2 + Bu (2.31) Substituting Eqn. (2.29) in Eqn. (2.31), X= AnXi + ApPnZ + Bu X= (An+ AvPnPi') Xi + Bu (2.32) Eqn. (2.82) gives the reduced order model for the system computed through the alternate method. Both models in Eqn. (2.2) and Eqn. (2.32) represent the same system dynamics. 22 Davison Technique 13 2.2.3. Improved Davison Technique It was noted that the above said methods were able to reproduce the transient behavior of the large scale system well. But, as pointed out in [3-5] and rectified in [6], there is an error in the steady state value of the system response. An improvement in the method was proposed by Davison in [6]. In the improved Davison Technique, the state equation in Eqn (2.2) is replaced by Y = DADDY + DBw (2.33) where, ay a D (2.34) a may HIB AON, Loy (2.38) # [AB], <0 where [A*-'B"], is the j"* clement of the ! vector A*-"B* and [A~*B],. is the element of the n vector A~'B which corresponds to the j" state retained in the simplified system, The new simplified system is equivalent to the following system. Xv = AX) Btu (2.36) Y = Dx" (237) and so it can be seen that the response of the new system will have correct steady- state values for a step-function input (provided that [A*-'B"], # 0), and will still maintain satisfactory dynamic behavior.. It should be noted that if [A*-?B*), = 0 for some j, then the steady-state values of the variable y; may be in error. Variables to be retained in the reduced order model should, therefore, always be chosen so that [A"-"B'], £ 0. For the case of multi-input systems, the corresponding model would be X= ANT + Btu, Loar Yo yo DX? (2.38) where, B= [Bi By Br | (2.39) and D,,i = 1,2,--+,r is determined from Eqns. (2.34 - 2.35), using By in place of B where B=(B B B, | (2.40) u Large Scale Systems 2.2.4 Suboptimal Control Using Davison Model For the large scale system represented in Eqn. .(2.1), an optimal controller may be deduced by minimizing the cost function J [eax +R) dt (241) where, Q and R are the weights of the states and inputs of the system respectively. Partitioning Q and using Eqn. (2.30), XPQX = XT QuXi+ 2XPQiX2+ XFQaXr (2.42) = XT (Qn +2QnPa Pi! + (Pil)! PRQmPaPi (243) Thus, I= Iu= f (XP QuX: +u? Ru) dt (2.44) fs Qn = (Qu + 2QaPaPit + (Pa')” PEQnPa Pat) (248) If the reduced order system is represented as. X, = FX, + Gu (2.46) then, the suboptimal controller is u=-R'G7™2 (2.47) where, 2 is the solution of the Riccati equation, QF + FTA OGR*GTN + Quy (2.48) 2.2.5 Control Law Reduction Approach Using Davison Model A large scale system (as in Eqn. (2.1)) with a performance exiterion described by Eqn. (2.41) can have its optimal control law deseribed as ul = Ke (2.49) = [mh 2) [2] (2.50) Using the Davison model in Eqn. (2.30), ul = (Ki + KPa Pi") X (2.51) 2.3 Chidambara Technique 15 2.3. Chidambara Technique Noting that the basic model of Davison in (2] does not give accurate steady-state response, Chidambara, in his correspondence with Davison ( (3-5]) had suggested an approach for model order reduction. In this model only the transient response of the left out igmored (i.c. X»), the steady-state contribution of these states are taken into account in order to nullify the steady-state error seen in the basic Davison technique. 2.3.1 Reduced Order Model Using Chidambara Technique For the Chidambara model, the system is represented as in Eqn. (2.21). The transformation P is computed as earlier and represented as in Eqn. (2.20) Consider the differential equation for Z, as in Eqn. (2.25). Dy = MyZn + Te (2.52) Performing Laplace transform on Bqn. (2.52), (sI — Aa) Za(s) = F2U(s) (2.53) Since only the steady-state contribution of Zz is to be considered, setting s = 0, MoZals) = TrU(s) Z, = ~My Tau (2.54) Now substituting the value of Z: in the relation in Eqn. (2.24), Xa = Pu By — Puy Tu (2.55) Solving for Z; in Eqn. (2.23) and substituting it in Eqn. (2.55), Xp = Py Pat Xi + (Pa Pia’ Pro ~ Pa) Ag" Tow (2.56) = LX; +Hw (257) Substituting Eqn. (2.57) in Egn. (231), the reduced order model of the large scale system is obtained using Chidambara technique as X1 = (An + An PuPs!) Xi + (Av (Pa Pa'Pae~ Pa) Ap'Ta + Bs) (2.58) = FX; +Gu (2.59) 16 Large Scale Systems 2.3.2 Suboptimal Control Using Chidambara Model ‘The method of deducing a suboptimal control law for the reduced order Chidambara model of a system was proposed by Rao and Lamba in [7]. ‘The modified performance criterion of the reduced order system reflects the performance criterion desired from the original system. If the cost function is defined as in Eqn. (2.41) and the original system as in Eqn. (2.21), then as in the case of the Davison model, the value of X™QX can be deduced using Eqn. (2.57) as, XTQX = X7QuXi + 2X7 Qua(LX1 + Hu) + (XPLT + wT H)Qaa(LXi + Hu) ‘The cost function can therefore be represented as Ine -/ (XP (Qu + 2QiL + L’QuL)Xi + 2X7 (Qual + AT QnL)u + (R+ HT Qn H)u) at ° (2.60) Defining Q = Qn+2Qul+L7Qal, (2.61) Ry = R+H'QuH, (2.62) S = Qu +H7Qnl, (2.63) and tsu+RISTR, (2.64) the simplified model represented in Equ. (2.59) is equivalent to X, = (F-GR,'ST) X1+ Gi (2.65) and the performance criterion in Eqn. (2.60) is equivalent to Ju -/ (XT (Qi — SiRy'ST)X, + aT Ryai) dt (2.66) If the matrices Ri and Quy = (Q1 — S:Rj*S?) are positive definite and positive semi- definite, respectively, then an optimal solution of the problem represented be Eqns. (2.65 and 2.66) is given as a = -R'GTOX, (267) where £1 is the solution of the Riecati equation O(F — GRIST) + (F — GRIST)" O-OGRAGTA+Q.— SiR; SP =0 (2.68) Thus 2.4 Marshall Technique 7 a — Ry1STX, = -R1(GTQ+ ST) X, (2.69) This optimal controller to the simplified system could also serve as a suboptimal controller to the original system, 2.3.3 Control Law Reduction Approach Using Chidambara Model Ifa control law w has been designed to control the original system then it can be applied on the reduced order model thus. Assuming the derived control law is of the form u = KX ~ ef X [x wl[X = KX 4+ 1X Using the relation between Xz and X; described in Eqn. (2.57), the control law can be represented as us (Ky + Keb) X, + KyHu (2.70) = (I~ KyH)u= (Ki + Kol) Xy If (I ~ KzH) is invertible then the control law can be represented for the reduced order system as (I — KH) "\(Ky + KaL)X, (2.71) 2.4 Marshall Technique S. A. Marshall had proposed an alternate way to compute the reduced order model in [8 This technique is quite similar to the Chidambara technique since it too takes into account the steady-state values of the X2 states, The difference exists in the manner in which the reduced order state equation is obtained 2.4.1 Reduced Order model by Marshall Technique Ifa large scale system is represented as in Eqn, (2.21) and the transformation that produces a decoupled system be as in Eqn, (2.22), then its reduced order model may be deduced thus. If the transformation relating the X states to the Z states be given as Q then, [Z]-z-rex-ox=[Se Se] [35] om 18 Large Scale Systems Observing Eqn. (2.25), it can be scon that the submatrix A, is associated with larger time constants of the system, whereas the response of any element in Zp settles very fast. ‘Thus, it may be approximated as an instantaneous step change. ‘This is the essence of the technique. Mathematically, this approximation is equivalent to putting ty=0 (2.73) Substituting Eqn. (2.73) and P 1 = Q in Eqn. (2.25), A = M%+Tw (2.74) 0 = AZ +Tyu and using Eqn, (2.72) in Eqn. (2.74), Zp = QuXi+QuXr Xp = —QQnX ~ Qa. Tau (2.75) Substituting Eqn, (2.75) in Eqn. (2.31) and using the relationships between P,, and Q.;, one obtains the reduced order model be Marshall technique as X= Pads PX + (Br — AnQy Ay 'T2 (2.76) 2.5 Choice of Reduced Model Order ‘The methods presented above give a reduced order model that is an approximation of the original system. However, it is not yet clear how small the approximate model can be and. yet accurately represent the process. It is seen that the smallness of the approximate model can be decided in terms of the largest eigenvalue neglected, the size of the original plant, and the reduced plant. This criterion is judged by comparing the time responses of various low order systems. 2.5.1 Model Order Selection Criterion by Mahapatra A criterion for selecting the model order for a reduced order Davison model was proposed by Mahapatra [9]. For a large scale system model in Eqn. (2.1), if the transformation matrix P is partitioned as Pu Pa P [z relstr Pr | (2.7) Mahapatra had made the assumption that the eigenvalues are real, negative and distinct. ‘The initial conditions and the inputs are considered as zero and unit step respectively. Then the solution of the system would be 2.5 Choice of Reduced Model Order 19 X(t) =P; fe — TAP +P fe — T] ADT (2.78) ‘The approximate solution provided by the reduced order model is Xe) =P fe — apr, Hence, the error involved in ignoring the higher order modes A,41, ++, An in state equation Eqn. (2.78) is given by E(t) = Pie —1Ay'T: (2.79) |E(Q|| S| Pall [je — 2) Aa" [Pall (2.80) mye pe lio (Ast) — 1] << vR=T a [= (3) | my sD WPall < [PIs (Pall < IP} = PBI (2.82) From Eqns, (2.80 - 2.82), \£()|| < PII. Tal]. [—-,0 < t < 00 (2.83) Dual For the given system in Eqn, (2.1), let [Pll [IPel = (2.84) ‘Therefore, IE) < Ta (2.85) Eqn. (2.85) shows that for a given system, the error in states due to neglecting the higher modes depends upon ‘Therefore, the error can be made small when Di vat —o Duval Mahapatra had extended his approach for Marshall’s model [10] and obtained the rela- tionship vnal+1 E()|| =k (2.86) Using the above equations, it is suggested that a suitable value of J can be selected such ‘that the error is within the tolerable limits. This criterion has following drawbacks 1. Ibis based on the assumption that the eigenvalues selected are dominant. 20 Large Scale Systems 2. ‘The assumption which are made carlicr are very crude. 3, is monotonically decreasing with increasing | and hence yields the trivial Daal result. ! = nas the best choice of Rao ef al have suggested an alternative approach to overcome the third drawback, IE@I s KU, Duel Ura Tr Upper bound on error ||£(4)| is proportional to U; and hence U; gives an idea about the actual error while Vj represents a measure of improvement achieved by increasing the model onder from U1 to L However all these criteria are based on the assumption that the eigenvalues retained are dominant. 2.5.2 Another Criterion for Order Selection and Mode Selection Lot a SISO system be given by X = AX+Bu (287) y = CX In this case scalar y is the output for evaluation and one is interested in constructing an 1 order system whose output is close to that of the system for a given input. Let Ar,Az,--+,An be distinct and negative real eigenvalues and of these \,,---, Ar are retained in the reduced model. yo hifi pace =CMZ = ew] 2.88) y Us ] (2.88) where, (M is the modal matrix Z is the transformed matrix (X = MZ) ‘fh, are the elements of the row vector h = CM ‘fare the elements of the column vector f = M7!B Output of Marshall's reduced order model can be obtained as 2.5 Choice of Reduced Model Order a Using y = CX = CMZ, G, output of the reduced model is obtained as (2.89) A An ISE (Integral Square Error) criterion as a measure of evaluation can be derived as 5 u hihi fi E= edt a (2.90) I, Ly (sr Ot mn) - For a given (** order reduced model, the procedure is to select a combination of | eigen- values and determine the ISE. This is repeated for all possible combinations and the one, which gives the least ISE is selected. Example 1 Let us consider the example of a third order system ool o 0 -10 -17 [1 0 ojx SOLUTION: Modal matrix M is given by M = Mo = 22 ‘Thus the result. Order | Eigenvalues Retamed [SE T 195 T =z 327 5 105 T 2 C0z7s 5 694 2,—5 313 Large Scale Systems Chapter 3 Model Order Reduction and Control - Aggregation Methods 3.1 Aggregation of Control Systems One of the most important techniques for modeling of large scale systems is aggregation. The concopt of aggregation was first introduced not in control systems, but in economics literature to address the appropriateness of the analogy between microeconomic and macroeconomic relationships... Aggregation of control systems was proposed by Aoki [11]. This aggregation approach is discussed in the following, Consider a large seale continuous dynamic system described by the state space equations X = AX+Bu (3.1) y = ex where X eR is u € RM is the control input vector R? is the output vector ‘The matrices A,B and Care consta: (A, B,C) is completely controllable and fhe state vector with appropriate dimensions and the triplet vable. We wish to replace the large model description by a satisfactory aggregated model given by FZ+Gu (32) w= He wher Z RB’ is the aggregated state vector w € RP is the aggregated output vector ‘The aggregated model description is considered satisfactory if for a given class of inputs {u}, the aggregated outputs w are good approximations of the original outputs y of the large model. Intuitively, the aggregated model has an order r such that m

You might also like