ALGORITHMS FOR LARGE SCALE SYSTEM CONTROL DESIGN L. Bakule Institute of Information Theory and Automation, Czechoslovak Academy of Sciences, Prague, Czechoslovakia
Abstract.The paper deals with the interactive cOTIp ~ter aided control design alsori tlms for the continuous linear dynamic systems . The decentralized pole assiGr~1'1ent problerll for tne strongly connected suhsystems using the dynamic compensation a:::Jo. the t'No-stage optinal control desi gn al~orithD reliable under tbe structural pe rturbations of the interconnections are cOl".sidered.:rhe iilteractioll by the COTJputetio n concerns the system strl1cture Bnd tbe data specification, the eigenvalues specification and the Graphical representatio:: of the desired veriables as e step responses. The case study of the power system mode l is included. Ee;{\'lords. AIcori tl1nlS for digital computer control; interactive compllter aided desien, large-scale systems; mu l tivar:!_ab l e control; opti1:lal control. I N':"CP..ODTJ CT ION
There i s increasing i n terest in tbe large-scale sys tem Llethodolog::r . A large-scale system may be considered as an interconnection of the s u bs~rste!:1s. ~he co~plexitv of the svstem is nain ly dete~mined - by t ~ ~ in~erconnections . A desirable feature is the separation at the subsyste~s a n ~ the i~te~con nection level. ':'he :::ost 13:i.nple approach is to ne~lect the interconnections and treat t~e lower order s ubsystems only. J uch a n approach works well if there is little interaction occ uring between controllers. Let us deal VIi th two procedures for the control design of t he large-sc/)le linear d;,rnaT!1ic s ysteDs, \'lhe re th.e i ::1terac tior.s infl1..1el:ce Ol·, t h0 s:',Tste r,~ activity. ~hc dcca n tr3li3ed polo ass:i_cr.!lellt proced :.~ ro cleels ;-:i th t~l.C c1:'nan~,c co:-.rpel~satio :: desiGn for the controllers so that t ~ e closed-loop eic;envalues 2~e prescribe d for the k -cha::1nel , j ointl~ co n trollable, jointly observa ble, si;ron~ly connected s~rsteY.ls [ 5]. :;"11e co nce ~t of t~~e cO:',lplcte s:rster.: is 1.' sed to :Ja l:e the systen bot~ co ntrollable and observa b le throl'3> a s inJle c::a~c -,-, e I b:r t:'-~e applicatioE of t~le :".oi:d:,r:1a:-:ic dece :1tralized feedbac :: to 011 c:Oan':e1s . Then t he Bra sc ~ - ?earso ~ procedure [4Q ca ~
be
't~sed
to construct t :J e
The two -stage optimal control procedure is s upplied by the power system Dodel control del3iGn example.
JJet us fori:iulate the froblern and the l-'ro b J.elii :2 . Problem 1. The system 3 is described as l:iJleer k-ch arcnel system of a form
',Jr=Ax+t 3.u., i= 1 1. 1. ~Ti = Cix, ., WDe r~.
xl:
';'!D \1.
i=1, ••• ,k~2, .
(resp.
11
1li.€ i
(1)
rH.
~
; resp.
y.£R ~ ) is t~e sta te (resp . the co~ ~
trol, resp. t )1 0 ont p,; t) vector. AE. f{n x n , I\€ Rn x nli , C € R !J i x n i are the consta~t matrice s describinG the systefl d;pa mi!:: s. Le t us suppose that S is a joint ly co~tro lla ble , j ointly obs ervable,i .e.
d:n.1 a l~-:i c
lo ca l control resulti ~G in a Cl0BGdJ_oop systen \',i t1'2 prescribed spectr1l!n.
S.F.(,.('. - M*
The two-staGe optimal co n trol desi~n proc ed1.1 re reliable under the str'.1ctl1 ral perturbati ons.Such perturbation C9n resul"; in i :::.stabj.li ty of the S;,TStem ['7J. The fi_rst stage deals \':1 to the local control desien, the second stage deals with the cont rol design minimizing the interconnec tion effect. Its i n fluence o~ the cost criterion is eva l119ted by t?J.e CN1PlltillG of the Sllboptil'!1ali ty index (7).
345
346
L. Bakule
the triple (e,A,B), C = col (C 1 ' ••• ••• ,c1,),:8 ~ ~O\" (B1"",l):) is cor.Jplete15 ,8J.(.) means transpose (.). decentralized pole assignment p!'oblen is to find the dynam:Lc local controllers of c form
~'he
.
zi
=
Sizi + RiYi'
Uin~ Qizi + FiYi' Z~E.TI.
i
= 1, ••• ,k (2)
so thet the closed loop spect.rum of the system (1),(2) is freely assiGnable. Note that the nondynamic loc::ll control has a for]"! l
1.\
= FiYi
min ,T
(4)
U
s\.lbject to
,
' E T) n.x n. -.., ".,·,m. x r.J. n >0 ,./; le l l,hi,,"n l l , ·oi- ,
~
n.
ts both cOY'.trollable and observable t~lrOl.lgh a si::Jgle channel denoted .;i. '1'he11 the dynanic conpensator can De applied to construct a control closedloop systen with prescribed spectrum [~. Theorem 1 (resp. Corollary 1) SOlve this problem [SJ.I'he triple (C,A,B) is complete if (A,B),(A,C) is controllable, observable[s]. n
,
Dj
.
Ri>O; XiE:R l i, resp. uiE:H -) lS the state vector (resp. the control).'l'he system dynamics is described by the constant Elstrices Al' ,B-;... ,A; -;. I,et us suppose k peirs (Ai,n i ) cont!'olable.
n+,.
4C
uj.n
11 A
BFII ,
}3
1
Let us suppose the control D = 1} + 1.11 •
(7)
;3CJIilJIn OH - ALGOEITIIIVI3
I ·et ,'S fo:.~r.;uJ.[lte t~1 e alsori tl1ms on the basis of the knoqn res ults. The solution consists of the solution of Problem 1 (resp. Problem 2) as the Algorithm 1 (resp. Algorit hm 2). The basic approacb to the Problen 1 solution is to apply first local nondynamic controls (J) so tllSt the closed loop system
'"
IIAII~II>B·tl·r.;·II;A '[' l l l c
4)
=
=A +lB.H.r: - - l l.• If it is satisfied Tgo~to 4) else go to end). If j = k-1 go to 5) else go to 3) •
i£(1, ••• ,k} A = A ,n = B., C = C .• c 'it' a-
5) a) Choose
6)
- Q U~ G U~ 1 ' • • • , Al ), = c1i8ij ( B , ... , I\J. (
tJn",/l.. =1
where ~o (resp. ~c) denotes the observability index (resp. the controllability index) [4,5]. Algorithm 1. 1) a) Give A,Bi,Ci,~i,j=O,xo = x(t o ) 2) Test ths joint controllability, joint observability (C,A,]), .£ = (C1"",Ck)'~ =~TIi • If it is not satisfied go to end) • J) j = j + 1. Test the completness (Cj,Ac,B j ) using the pseudorandom generator. So that
(6)
·..,here J,. l\..
.
K.!: ,a 0 ).'f3i
the desired set, cenerally complex conjugate pairs, of the eigenvalues A = [ .A~, ... ,).n+1-1; t- min f ~Ol ale},
~<.l
I,et us find the control optimizing ooth the stlbsystems Beti vi ty and the i nfltiGl1ce of tbe L1terconnections. ']'hei:: find F
.
Denote: I.AI - AI =.[ ).1 ~i the · desired ,-0 characteristic polynomial
(3)
resulting from (2) for n i = O. :Get us formllate the Problem 2 now. Problem 2. Let us ~ind the control fi for the cost
VI.~lGre
(8)
Test the A-cyclic property. If it is not cyclic go to 7) else go to 8). Using the pseugo-random genera.7) tor geDerete H 52 that U,,'..II=IIBHclI, A:=A+BHC; c.;o to 6). G) a) Choose c' = d' C so that (A,c') is ohservable. 9) Compute a column vector :L. I = [L i]
i= 1, ••• , n;
- d'A i - 1 n iri J d'A n - k - 1B·, 1";~+'-~ k -
10 )
ka f - -
Compute q. 11)8) Give A. 12) Solve equation Pe = f, r= m~, ". w~lere e (resp.f) is q+r(q+1)
Large scale system control design
(resp. n+q) column vector, F=[Pij]; Pj=col (P1j"" ,Pn+q, j ); Pj=col( O, ••• ,O,p .'j =1, ,(11-1' ••• , «()' 0, ••• , 0) , j £ [1 , ••• , Pj=col (0 , ••• , 0 , Pjj=L 11 ,L21""
qJ ;
te-time problem reletion [1J. I.et us ~ l rther suppose IBI = m. Algorj. thm 2. 1)8) Give ,\,Bi,Aii,Qi,Ri,i"i,xo=x(to)' tJ
2)
3)
,nJ
fi= ;1q_i_n,iE [n+1, ••• ,n+ql. Compute H = [hi i]; a qxq matrix
u·l = - R.B.P.x. = G.x .• l l J.. J. 1. 1
4)
5)
is dcterdned: ,je:{1, ••• q],
h~j= -e j
h " _0 = 1 J J- J
h~~
=
-J
h j =-
H
6)
0 else.
q+rq+j
= 111 11.:) L..
-Hr =
G-F.
Compute ~ ec $. ZIb-ac> } t" 1/t. -, l [li.. . A a "*L, '1'1\ . .J ,1\. . =1\. . - B . F. . i.j
; ..j
lJ _
lv
_
J-J
lJ
-1
l
lJ
eq
c=mrx\[Ii] ; ~ L.] (resp. AM l.] )
Ho
.... I
1\ q
= ~(A+BG), G =
b=mi n ~m [p i J ,:? i= Qj _+1' i Bi Ri Bi Pi
.
0
~
...
. - - ..... - .. - hr -. d'
h _ r 1 IT 0 'r The feedback matrix A~
( , , ••• ,u , )' ; u = u 1 + 1.1 2 , u 1=.u 1 k J = ,..,0 x'Px 0 ,P=diag (P 1 " " ' ? lC );61= IV =v(A+BG) ,G=dia gJG , ••• ,G ),A= 1 k A
fj ~2 ,
e q+ j
(B'B)-1 B 'A, u 2 =-Fx
A-A,
+e q+ + 0 rq j e q+r (q_1)+j e 1
8
Compvte F
v
Eo = [h~j]
Compute -1
f.= Al1 +q-l._J., iE $1, :t. "'n-J_ t ••• 1])
Solve the Riccati equation AiPi+PiAi-PiBiRiBiPi+Qi=O, ~i •
• • • , J~n 1 ' 0 , ••• ,0),
j€[q+1, ••• , ( q+1)r1; f=col (f 1 " •• ,fn+q)'
347
=lAc o
15) a) 'i;est a nd e-valuete step responb) se. If it is not satisfied choose A ' new, A -A new, go to 11) else go to end). 16) End. IJet us further forJl11Jlate Proble!n 2 solution. It is based on the Riccati algebraic subsystem solution and on the global control design minimizing the interconnection effect. 'i:he suboptimali ty index ~ evaluates the interconnection effect on the criterion value(r6J .1.The Riccati equation solution is a doubling continuous - time algori thnl effectively using a discre-
denotes the minimal (resp. maximal) eigenvalue of the corresponding matrix • 7)a) Test and evaluate step response, b) ~ , J ,
End.
The IBM 2250 display interaction is inclvded in two wa:'iTs. The step number of both al~orithms is followed by a) (resp.b»). a) (resp.b» means that the interaction in the problem data can occt1r (resp. the specified variables can be displayed as the time functions-step response).
The S8Hpling j.nterval din the both cases is ad justed on 0:-\1 (resp.J;\.2) in the Algorithm 1 (resp. Algorithm 2)
dA1 :. 0.5/ AM [A~] r1,
r1 .
0A2· 0.51 AM [A + BG 1
Of c01.1rseo~JA1(resp.J"~dA2) but by the grap..hical representation we can choose as:icT,i ts an integer. The IBM SSP/)60 and the user-simplified GS? for :PJJ 1 have been used.
348
L. Bakule
Example. Using the Algori tblTI 2 der.lonstrate the influence of t he state penalization lJsing Q on the s ubo ptimali ty indexJ. Let us con sider power systeM model formed by two cOl'.nected plants [2,6J. The considered system:
:~ 11 02.72 o
o
1 0 0 j'BI1= - (.!. 04 5 0 n -3.33 3.33 5. 2 C.I 12.5
,2 o 0J' 000 A" [~'72 000 0
R1= [1J •
000
t
Q1=
[Vi0 w0 00 goO]' 000 000
Vie;. 1. b.
A1=A2 ,B 1=B 2 ,A·12 =A 21 ,Q1=Q2 ,R 1=R 2 , xo=col(O 0
The corlputation ~. as been perforr::ed on the IBM 370/ 13 5 computer.
0.01 0 0 0 0 0).
For w=0.15 (resp.15; resp.150) the s uboptimality index ,=0.23 (resp. 0 .07;resp. 0 ). Fig.1a (resp.Fig.1b) denonstrates the freql1 ency deviatio!l for y/=0.15 (resp.w=15).By the increasing state p en alizatio ~!. the system is better daMped and logi c ally t he intercon~e ctio~ effect decreases. It practically vanisbes for ~&150.
t
Fig.1.a.
Af(Hz] (resp. A f ; resp • .A f 2 ; resp. t ls]) denotes thJ frequency deviation (resp. x 2 = .4 f 1 of the first subsystem, resp. x6 = A f2 of the second subsystem; resp. time). t max = 8ls1, d = 0.1 in all cases.
CONCJJUSION Two interactive procedures have beer described. The decentralized pole assj.gnme!lt problem "l nd the two-stage optimal cOlitrol desi.gn alGorithnlS. The simple use of the second procedure is given on the example. 1) Jl.nderson,B.D.O.(1 978). Second-order converge nt algorithms for the steady-state Ri ccati equation. Int.J.Control,vol.2 0 ,g,pp.2 95-306. 2) Bakule L. ( 1978 ) Two-level control Generation of nu ltiarea electric enerey systems. In I.Troch (Ed.), 0ir.1UletioYl of Control Sys tems, North-Holla nd ,pp. 161-1 GJ . 3) Bakule L.(1 979 ). On decentralized lerge-sca1e d:~lnBmj. c s~rstem control. In J.Bene~,L.Bakule (Eds.), 3rd Formator Symposium, Academia:--
205-218.
4) Brasch,P.H. and J.B. Pearson (1970). Pole plecenent using dyns@ic co~pen sators. IEEE AC ,vol.15,1,J 4- 4J . 5) Corfmat,J.!'. s i:d A. S. IV\orse (1975). Decentralized con trol of linear !!l1..11t1variable syster:s. Proc.6th I!:'i'.C ConGress, Boston, paper 43 .3. 6) Fosha,C. and O.L E13erd (1970). The r:!egawatt-freql1er.c~T control proble]!!. IEEE Pl\.;3,4,pp.%3-57 1 • 7) ~iljok,D.D., Bnd S.K. Sundaresha l1 (1976). OptirneIity vs. reliability. In Y.O.Ho, S. K. Mitter (Eds.), Directions in :r,arf?e 3c31e s ~stems, Plenum Press, Hew York,pp.1 4-152. 8 ) Wonham, W.M. (1974). 1inear multivariable control. A ~eo me tric approach. Springer Ver ag, Berlin.