To appear in Proceedings of First IEEE Symposium on Large Scale Systems, October 1982 LIDS-P-1224 HIERARCHICAL AGGREGATION OF LINEAR SYSTEMS WITH MULTIPLE TIME. SCALES M. Coderch,* A.S.Willsky,* S.S.Sastry* and D.A. Castanon* Laboratory for Information and Decision Systems Massachusetts Institute of Technology Cambridge, Massachusetts 02139 0. INTRODUCTION Models of large scale systems typically include weak couplings between subsystems. This results in the evolution of different portions of the system at different time scales. Frequently, intuition suggests that subsystems or groups of states evolve slowly in comparison with their internal dynamics. This suggests that the overall system dynamics analysis may be simplified (or so1metimes made tractable) by separately analyzing the dynamics at a specific time scale by assuming constancy of variables evolving at slower time scales and steady state values for variables at. faster time scales. If this intuition were to hold true, it then follows that system behavior may be approximated by means of a hierarchy of increasingly simplified models valid at progressively slower time scales. The motivation for the present work came from mathematical models for interconnected power systems. These models have variations on several time scales - nearly instantaneous adjustment of loadbus angles and voltages, dynamics of the swing equations, voltage regulator and turbine power generation dynamics, generation scheduling (set generation o dynam shavior , 'e point changes) are examples of progressively slower dynamics. With this general philosophy in mind, we study in [8], a very simple class of systems - linear and time invariant (equation 1.1). For this class of systems we model weak coupling by parametric dependence of (1.1) on C, and obtain tight conditions under which a hierarchy of reduced order models valid at different time scales may be constructed. The study of this (deterministic) eauation is also relevant to the hierarchical aggregation of finite state Markov processes with some rare transitions. The details of the application of our results in this context are presented in [3]. 1. PROBLEM FORMULATION We consider here .linear time-invariant systems of the form C C x = AO(C)x x(O)=x 0 (1.1) 0 0 where x C R and the matrix A (E) is an analytic 1 w][4]) Research supported in part by the AFOSR under) grant 82-02,52. The authors are thankful to B. Levy and J.C.Willems for the helpful discussions. function of Et, of normal rank d except at s (A) k A0) k = 0 (1.2) we analyze the asymptotic behavior of x (t) as C 4 0 on the time interval [0,=[. In general Lim sup I X C+0 t>0 (t)-xO(t)II O so that asymptotic behavior at several time scales needs to be considered - x (t) is said to have well defined behavior at time scale t/g(E) (where g(E) o is a monotone increasing C- function on [0,E 3 with g(0)=0) if there exists a bounded continuous function Yk(t), called the evolution at that time scale, lim su EC0 6<t<T li|x (t/g(c))-y (t)|l=0 6>0, T<- (1.3) In this paper we give tight sufficient conditions under which the multiple time scale bee of x (t) can be fully described by its evoh ' lutions at time scales t/E for integers k=0,l1,...,m. (i) (ii) These evolutions are used to: provide a set of reduced order models valid at different time scales. provide an asymptotic approximation to xE(t) valid uniformly on [0,-[. nxn Given A e R , R(A) and N(A) denote the range and null space of A. p(A) denotes the resolvent set of A, i.e. the set of X C C such that the resolvent, denoted R(X,A): = (A-AI) , is well defined If X=0 is an eigenvalue of algebraic multiplicity m, then the Laurent series of R(X,A) at Forrallyv, our results hold even if A (C) is only assumed to have an asymptotic series (see,for eg. of the form (1.2), provided that A (c) has of the form (1.2), provide constant rank d for 0 2 A=0 has the form (see 53) M- 1 (A) m-k-A) R R (XAk ,k k -+P=(A0 - where where 'ko So that it .Akl',1'Ak,+1 only triangularly as shown in Table 1. proceeds of special interest to us in the sequel is (ii) since they ,A1 0 A2 0 the structure of A (2.2) R(k,A)dk requires (i) The computation of Ak+l Remarks: (2.1) where P(A), the eigenprojection; D(A), the eigennilpotent and S(A) are defined by i| k 0 __ -ks(A)(A P(A):= k (ki (ki nd S k ( k - determine the leading term in the asymptotic expan-XR(X,A)dX D(A):= ' 2iriJ Y S(A)-1 (A):-- 2xx ¥X -R(XA)d ('2.3) * AO ()t . For these matrices, (3.1) can be sion of e simplified considerably. (see Remark (ii) after (2.4) (2.4) the Theorem and Corollary). Theorem withy a positively oriented closed contour enclosing 0 but no other eigenvalue of A. Let AO(E) (1.2) of normal rank d except at E=0. A is said to have semi-simole null structure. = N(A) (SNS)if D(A)=0. In that case, Rn - (A) and P(A) is the projection onto N(A) along R(A). A is said to be semi-stable if it has SNA and all its non-zero eigenvalues are in the open left half plane. At If A is semistable, then lim e =P(A) and At S(A)= -J((e 0O -P(A))dt= (A+P(A)) Rn = R(A00) Q A . Then STATEMENT OF MAIN RESULT below . The proof relies in [8] Z For our development, we need an array of matrices A.ik, i>O, k>O starting from the AOk of k=O (1.2), constructed recursively ((k+l)th row from kthi row) by the following formula m kem k=0 (kl) Z+1 k L k ) kl+ +kp+l (k (k) Ak , Sk ... Ak V p-l (3.2) m ) . (0 N (0 k=0 0 e = (gt) 0 (E:,t) is any one of the four expressions A k eA ° O P 1 - (3.5) ml k t (3.6) iT e k (3k7) t )A (3.3) l c ) V.>l, k.>0 1- . I-Ph. lim sup C+0 t>0 We present here a uniform (in t) asymptotic A (t)t A0E) t .bwhere involving behavior at approximation of e ; k=0,l ...,m. time scales t/E on results in [53 and is given R(Am m, be the prolet Pk' for kO,1 . .... be the projection onto N(Ako0) defined by (2.2) and -1Further Since S(A) is a generalized inverse of A S(A)Ax = x = AS(A)x for all x e R(A)) we denote it 3. A1 10 + rank A0 are semistable with rank A rank A O=d then where N= - P(A). If A 00 further - further be a marrix with SNS of the form e 1- With this theorem in hand, -the entire multiple (1.1) can be read off time-scale structure of as follows: --~~------------------·-~~ ~ c ckt ~'~0 -- `--- - - - 3 Corollarv A20 is the null extension to R Under the conditions of Theorem 1, x (t) of (1.1) has the following multiple-time scale properties: properties: A (i) lim sup IIxE(t/ck)- k e 0 A A mod R(A00)mod R(A)N( 00 02 010001 i.e., A2 0 is given by X0 11=0 C+0 0<6<t<T< of (A -A A A )P P (3.8) (3.8) 10 02 01 00 01 0 1 where P1 is defined as in the theorem Pictorially, for 6>0, T<- and k=0,l,.....,m-l A20 is the null extension to R (ii) lix C(t/em)- X e m sup 0<6<t< lim C+0 (ii uIEtE) Am0t xO l=0 A02-A01A00 01 (3.9) for 0>0 k=PoP .. Pk Rn ~ i0~ A 1 cili 0 ) C10 (A 00 .4. Rn 0 ) N(A where 70=l and P for ,m.k=l. N(A00 Equation (3.8) implies the results of [1] and [2] where the authors analyzed the convergence20 A 0 (E)t/C for fixed 10 (ii) It is important to be able to calculate the AkO for k=0,l,.....,m for the given data A0 0 ,A (1.2), without having to obtain the complete matrix of Table 1. This can be a variety of methods. One approach that cessful is the formal asymptotics of [7] the A.,O to Toeplitz matrices constructed the {A i 1R(Ao0) 20 s and over compact time Remarks: (1) The requirement of semi-stability of the matrices A0 0 ,A ... ,A to obtain well 0 k 00 10 mO defined behavior at time scales t/Ek is a tight sufficient condition. Examples showing the failure of the theorem without these assumptions, are given in Section 4. of R A intervals of the form [O0,T]. A6 2 ,... of A A n A is the null extension to R 10 is t:e null extension to of A01 mod mod of A ) R0(A00)/(A00 , i.e. A!0=PoA01PO, where P0 is defined in the theorem. Pictorially A1 0 is the null 1 l 10 20 i A0 1 A0 A1 0 00 A0 0 01 00 . Connection is made therein with of the A k0from the A i proceeds as follows: 10 and so on. The reader may refer to [7] for complete proof and details as well as the conA .....and the Toeplitz nections between AA matrices between A1 0 1 A 2 0 A ... and the Teplitz 3 0 done by is sucrelating with the Smith McMillan zero of AO(E) at E=O. In par0A00 ticular m is proven to be the order of the Smith McMillan zero of A (C) at C=O. The construction 00(A .... 0 A00 A03 A2 A0 A02 A01 A00 A01 A00 L A0 0 O 1 O A0 0 O -and so on. (iii) The reader should observe using the formulae in remark (ii) above that even if A (E)=A00+A01 t he system can exhibit time scales of order 2 3 t/E_ t/_ and so on. ciated fact. This is not a widely appre- nA n extension to R of A10 obtained as below i A n N (A ,0 Rn R N(A00)P Rn (iv) Reduced order models. It follows from (3.8) and (3.9) that the evolution of xt(t) at time scales t/E , k=O,l ....,m is given by _ P0 Yk.(t) = e 10 ni R (here, i kOx=O, ,. .. m Thus, x (t) may be represented asymptotically by R (AOO) N(AO) the followinc expression unifcr-mly valid m is inclusion) Y x (t) = k=O k t) (1- for t>0. ; 7k) + o(1) k=O 4 From the direct sum decomposition (3.2) of the m not implY the semistability of {Ak)k=l theorem, it is clear that a basis T e Rn chosen such that Counterexample 2 (Semistability of A in I A Cmt m n can be mt T seristability of Ako e Let Alt A () t - (1) O e .0 0 T +A (E) eAoCt l At0 e I_ where A.....,A li (3.10) are full rank square matrices re- of A0 (C). the new basis. (3.10) shows that the system (1.1) decouples asymptotically into a set of subsystems evolving at different time scales governed by the reduced order dynamics of {A}k=o. A1A0 The significance of each row of matrices in Table I is discussed in [81. 0 0 11 1 -2 0+C1o -1 0 0 -1/2 0 so that A10 is nilpotent. -2 0 are semi-stability In this example A li- e does not exist showing no well!l e does not exist showing no wellES0 2 defined behavior at time scale t/E in spite of a real eigenvalue of order 4. 0 It may further be verified that 0 (v) Two time scale systems have been the focus of considerable effort by Kokotovic and coworkers (see [6], for example). It is easy to check in our framework that the assumptions in their systems guarantee precisely two time scales. O 0 Note that the eigenvalues of A (E) 0, -2+o(1), -S2+o(E2 ) , showing the presenting the non zero portions of A00 ..... AmO in (vi) ) . TIGHTNESS OF THE SEMI-STABILITY CONDITION The requirement of semi-stability for the matrices AOO'A10 . ' 5. CONCLUSIONS ,Am0 for the system (1.1) to Our theorem in section 3 provides a uniform aproxiiation over te entire real line [ ) to [O0, to be line approximation over tne entire realcan tight.Counterexam~les have well defined behavior at time scales /Cm , t ..... t/E, t, be can Us is tight.Counterexamples t, t/C,...,t/Cm. found for the nonexistence of well defined be-e havior at different time scales if x0 iS not x0O semi-stable for some k: Counterexamole 1 (A00 does not have SNS) LetlA 1] 0 A 1 evolution of he system (1.1), thereb extending the results of [1], which are valid only for inervals of the form [OT/. Furthermore, intervals of the form [0,T/CE. Furthermore, the hierarchy of models which result from the Corollary is an extension to multiple time scales, of tne aggregation results in [6]. The application of these approximations to problems of estimation K ~1 | (S )~= be reported in later publications. 1 o REFERENCES [13 Then, cosVs t -vS sin¢v t' 1 0sinvF .oe.-t L E sjin¢ t cos- t _ S.L. Campbell and N.J.Rose, "Singular Perturbations of Autonomous Linear Systems", SIAM. J. Math. Anal., Vol. 1C0,No. 3, p.542, 1979. eA (Si t e [23 Vol. 29, p. 362, 1978. A (E)t/E e does not exist for any t 1Note that l ([3 Finite State Markov Processes," LIDS P-1198, April 1982, to appear in Stochastics. Recall that AkO is semistable if it has SNS a-.ndall non zero eicenvalues have negative real parts. The necessity of the second condition is obvious and we will not furnish a counterexample to illustrate it. We would like to note that the semistability of A0(E) for EE[0,E] does ------------ M. Coderch, A. Willsky, S.S.Sastrv and D.A. Castanon, "Hierarchical Aggregation of . showing lack of well defined behavior at time Remark: S.L. Campbell, "Sincular Perturbations of Autonomous Linear Systems II," J. Diff. Ec., - - - - - - - - - ~ [41 W. Eckhaus, "Asymatotic Analysis of Singular , No A l or (Cerators," ~o Springer Verlag, Berlin, 1966. [6) P.V. Kokotovic, R.E. O'Malley, Jr. and P. Sannuti, "Singular Perturbations and Order Reduction in Control Theory-An Overview," Automatica, Vol. 12, p. 123, 1976. [7] S.S. Sastry and C.A. Desoer, "Asymptotic Unbounded Root Loci Formulae and Computation," College Eng.,. U.Berkeley, Memo UCB/ERL M 81/6 Dec. 1980, to appear in IEEE Trans. (AC), April 1983. [8) M. Coderch, A. Willsky, S.S. Sastry and D.A. Castanon, "Hierarchical Aggregation of Linear Systems with Multiple Time Scales," to appear in IEEE Trans. on Automatic Control. A00 A01 . Ao A10 A11 ........... A1,1 Al 1,0 A£-,11 Ai,0 Table 1: The array A k,Z is grown triangularly.