You are on page 1of 5

Proceedings of The

IEEE International Conference


on Industrial Technology, 1996

Applying Structured Singular Values and a New LQR Design to


Robust Decentralized Power System Load Frequency Control

T. C. Yang H. Cimen
School of Engineering, University of Sussex, Brighton, BNl 9QT, UK

Abstract- The design of decentralized robust load Af2 A p ~ 2A P G ~APCzIT


frequency control for interconnected power sys- and A E RgX9, B E R g x 2and C E R Z x 9 .The system
tems is studied in this paper. It is shown that, sub- is stable and the control task is t o minimize the system
ject to a condition based on the Structured Singu- frequecy deviation Afi in area 1, Afz in area 2 and the
lar Values (SSVs), each local area load-frequency deviation in the tie-line power flow APti, between the two
controller can be designed independently. The ro- areas under the load disturbances A P D , and A p ~ 2 in the
bust stability condition for the overall system can two areas. Since the system parameters for the two areas
be easily stated as to achieve a sufficient interac- are identical and the Apt,, is caused by Af1 - Afz, the
tion margin introduced in this paper, and a suf- system performance can be simply tested by applying a
ficient gain and phase margin defined in classi- disturbance A P D to ~ the system and observing the time
cal feedback theory during each independent de- response of A f i . Some simulation results for A f1 when
sign. Within this general framework, each local a step disturbance of apD1 = 0 . 0 1 is ~ ~applied to the
area controller design is based on a new Linear system are plotted in Figure 9, 10 and 11 as dashed lines.
Quadratic Regulator (LQR) design approach [l],
B. Transform Into An Equzvalent Deszgn Problem
I. INTRODUCTION In general, an m-area power system load-frequency con-
In the dynamical operation of power systems, it is usu- trol problem can be modelled as a complex system consist-
ally important t o aim for decentralization of control action ing of m subsystems:
to individual areas. This aim should coincide with the k = A,x Bmu +
requirements for stability and load-frequency scheduling
y = c ,x (2)
within the overall system. In the load-frequency control w h e r e u = [ u I , ..., u m I T ; y = [ y l , ..., y m ]T =
function, it is necessary that the system frequency and
the inter-area tie-line power are kept as near to the sched- [ A f l , .'., Afm IT x = [ XI, ..., x, IT and x, are the
uled values as possible through control action. In this pa- state variables for the i-th area (i-th subsystem). The
per, the problem of decentralised load-frequency controller sample system used here is a special case of m = 2.
design is translated into an equivalent problem of decen- An m x m transfer function matrix G(s) linking U(s) =
tralised controller design for a Multi-Input, Multi-Output [ ~ I ( s ) !..., u m ( s ) IT and Y(s) = [ Y~(s), ..., y m ( 8 ) I T :
(MIMO) control system. The proposed design method = G(s)U(s)
Y ( S ) G ( s ) = [sij(s)lc,J=l,,m (3)
is base on SSVs and a new LQR design approach. The can be calculated as:
method is illustrated by a design for a sample two-area
power system, which is widely used by the researchers in G(s) = C,(SI - Am)-'Bm (4)
this area [3,4]. The design of m decentralized local controllerr now be-
comes the design of a m x m diagonal matrix F(s) =
IT. DECENTRALIZED DESIGN BASED ON SSVs
diag[fi(s)]a,l, ,m as shown in Figure 2.
A. A Sample System If all siJ(.) (i # j ) in G ( s ) were equal to zero, then
Figure 1 is a block diagram for the load-frequency con- each controller could be designed independently just as if
trol of a two-area power system. The nomenclature used it were in a SISO system as shown in Figure 3. However,
and the nominal parameter values, in per unit (pu), are since g,,(s) [i # j ) are not zeros, the following question
given in the Appendix. A state-space model can be con- must be resolved, i.e., ifeach f,(s) (i = 1,...,m) is designed
to form a stable closed-loop system as shown in Figure 3,
structed as:
x = Ax + Bu what are the additional conditions which can guarantee
that the global system of Figure 2 is stable? The answer
to this question is discussed in the next subsec'tion based
on the theorem given by Grosdidier and Morar (Theorem
3 in their paper [2]).

0-7803-3104-4 ' 880 -


C. Decentralazed Controller Desagn for MIMO systenbh (c-4)*
umaz(W-' ( j w )H( j u ) )< p-' ( E ( j w ) W ( j w)) V u
In Grosdidier and Morar's paper, a transfer function
where W(s) = diag[wi(s)],,l,z, is a properly chosen
G(s) = [g2J(s)]z,3=1,2, for a m x m MIMO plant is de-
composed into: diagonal weighting function matrix. (c-4)* can also be
replaced by:
1 G(s) = G(s) G(s) + (5) (c-4) / h i ( j w ) w i l ( j w ) l < p - l ( E ( j ~ ) W ( j w ) ) Vw
where G(s) = diag[g,,(s)],,l,z, h
is a diagonal matrix; ( i = 1 , 2 , .."
all diagonal elements in G(s) are zeros and off-diagonal Due t o w,rl(jw) in (c-4), although
elements in G(s) are equal to those in G(s). p-l(E(jw)(W(jw)) is still the same for all SISO loops,
Using the notations: the restrictions on Ihi(jw)l are different. In fact, (c-3) is a
special case of (c-4) with W = I .
- E(s) = e ( s ) G-'(s) (6) Before applying the above results developed in [a] to our
H(s) +
= G(s)F(s)(I G(s)F(s))-l = diag[h,(s)] (7) system, it is necessary t o consider the issue of the robust
~ ( s = +
) G ( ~ ) F ( ~ ) ( I G(S)F(.))-~ (8) stability.
where F(s) = diag[f,(s)],,l,2, is a diagonal transfer The stabilitjr condition (c-4) is given for the nominal
function for a decentralized controller in Figure 2; H(s) or plant G(s). If the state space model of equation (1)
H(s) is a closed-loop transfer function matrix for a feed- changes, the plant model G(s) will also change. It is
back system consisting of F(s) and G(s),or F(s) and G(s) genarally not possible to establish a clear relationship be-
respectively. Grosdidier and Morar have proved the follow- tween the change of values in (1) and the change of values
ing theorem: involved in condition (c-4). For this reason, we specify the
The closed-loop system H(s) is stable if robust stability conditions as:
(c-1) G(s) and s ( s ) have the same number of Right (r-1) Condition (c-4) is satisfied with a sufficient mar-
Half Plane (RHP) poles; gin. This can be checked by plotting Ihi(ju)l and
(c-2) G(s) is stable; and p-'(E(jw)W(jw))lwi(jw)I on the same graph and an in-
(c-3)* G 7 z a z ( H ( j w ) ) < p - l ( E ( & J ) ) vu teraction margin for loop i can be defined as the shortest
where nmazdenotes the maximum singular value; and vertical distance between the two curves.
1-1 denotes Doyle's structured singular value with respe$ (r-2) There are sufficient gain and phase margins in each
to the decentralized controller structure of F(s). Since H SISO loop for the stability. This can also be checked by a
is a diagonal matrix in this paper, condrtion (c-3)* can be Bode or Nyquist plot of f % ( j w ) g , , ( j w ) .
replaced by: For the given sample system, m = 2 and the equivalent
(c-3) lh,(jw)l < p-l(E(jw)) Vw (i == 1 , 2 , ...m ) MIMO system can be represented by Figure 4. A plot
where 1 1 denotes the magnitude. of p-l(E(ju)) is given in Figure 5 as a solid line (W is
This theorem gives suflczent conditions for the system chosen as I and is neglected).
H(s) to be stable if the controller design is based on the
w 111. LOCAL AREA CONTROLLER DESIGN
fully non-interactive model G ( s ) , i.e. each f i ( s ) is de-
signed, independently, based on an SISO model g;,(s). In Now consider a system enclosed by a block with bro-
particular, condition (c-3) states that the magnitude of the ken lines in Figure 1, the system has an input u1 and an
frequency response of SISO closed-loop transfer function output y1, and a state-space model for this system can be
constructed as:
X0 = A'x" + b"U1
scalar frequency dependent function p-'(E(ju)). Grosdi-
dier and Morar have also proved that [2], although (c-3) y1 = coxo (9)
where
is a sufficient condition and therefore may have some con-
servativeness, compared with the other conditions devel- XO = [ Afi APT^ A P G ~APC1lT
oped for the independent decoupled design, for example A" E R4X4, bo E R4 and CO E R4.Inspection of Figure 1
the diagonal dominant condition and the generalised diag- shows that, a transfer function:
onal dominant condition, (c-3) gives the tightest restric-
gyl(s) = co(sI - A0)-lbo (101
tive band and is the least conservative. However, since
the same restriction p-'(E(jw)) is applied to all hi(s) in is equal to gll(s) given in Figure 4 if AP,i,(t) E 0. It is
condition (c-3), a modification on this condition [2] can well known that, for the system represented by Figure 1,
be made to provide more flexibility and to reduce further APtie(t 00) = 0 after step load disturbances. However,
the possible conservativeness caused by the inflexibility in in general:
condition (c-3): Sll(S) = SIl(S>+Am(.) (11)

* 881
~

where Agll(s) # 0. For the purpose of a local controller The transformation yields the state vector of the bal-
design, we can consider gil(s) as a nominal model for anced system xb having its elements ordered according to
gll(s) and with an uncertainty Aglljs). their combined measures of controllability and observabil-
Figure 6 gives the magnitude frequency responses of ity, reflected by its associated HSVs. The most (the least)
gfl(s) and gll(s) and it can be seen that, g:l(s) is a rea- controllable and observable state will appear as the first
sonably good approximation of g11(s). (the last) elenent in the state vector of the balanced sys-
Since the plant parameters in the two areas are identi- tem. Based on the above analysis, the choise of the Q
cal, we only need t o consider decentralized local controller matrix is made as follows:
for area 1. To design a local controller based on the nomi- (i) The first 1 states of the balanced system are those
nal model of g:,(s), i.e., the state-space model of equation states which are deemed t o contribute most to the dynami-
(9), it is necessary to obtain sufficient stability margins as cal behaviour of the system. Thus they should be weighted
stated in (r-1) and (r-2), in order to have a robust perfor- according to their contribution.
mance for the global system represented by Figure 1. This (ii) Ignore the last n - I states of the balanced system
is also because that the gil(s) is an approximation of the by placing zero weighting on them. This is because those
“plant mode!” gll(s) given in Figure 4. last n - 1 states are poorly controllable and/or observable,
It is known that, for a SISO system, a state feedback and therefore play a minor role in the dynamical behavior
based linear quadratic regulator design to minimise: of the system. Thus it is impractical and useless t o expend
, r03 energy, which has to be very high, on these states.
With regard t o the I retained states, obviously the j-th
( j = 1,..., 1) state is more controllable and observable and
leads to an equivalent SISO feedback loop as shown in Fig- therefore plays a more important role in the system dy-
ure 7, where k is a state-feedback vector obtained from the namics than the following I - j states. This means that
LQR design. In has been proved that [5], for the feedback the j-th state requires less control effort than the follow-
loop shown in Figure 7, the resulting optimal control sys- ing ones t o affect a change in its status. (i.e., it is more
tem has an infinite gain margin, a phase margin of at least economical to use to affect a change in the system dy-
60° and a significant amount of system nonlinearity can be namics). Therefore, to exploit the physical characteristics
sustained without instability. The special characteristics of the states and their role in the system dynamical be-
of the Nyquist plot for the open-loop system in Figure 7 haviour, each state should be weighted according t o the
can also provide some information about the closed-loop ratio of its contribution with respect t o the most control-
system frequency response of - yl(s) i.e., the frequency re- lable and observable one, i.e., the first state:
U 1 Is’)’
- \ I

sponse of h l ( s ) concerned in the robust condition of (c-1). &b = diag(1, -,a1 -,a1 ..., -,61 0, 0, ..., 0)
U2 a3 Cl
The main difficulty of LQR design is to choose the Once Q b is obtained, it is transformed back into the
weighting functions in (12). Aldeen and Crusca has pro- original system coordinates:
posed [l]a systematic approach t o choose these functions Q = T*&bT
and applied this approach to power system stabilizer de- where T* is the conjugate transpose of T .
sign. This approach is also adopted here and the main T in (12) is selected to make an appropriated “gain” of k
procedures are illustrated below. and an appropriated interaction margin as stated in (r-1).
The model of equation (9) is first transferred into its A smaller r implies a weak penalty on the control signal, +

balanced form: and hence leads to a “bigger” k. Since the frequency re-
X b = AbXb + bbU
sponse of hl(s) is affected by the feedback control, this in
Y = cbxh (13) turn will affect the interaction margin. In practice, r can
where
xo = T-lXb Ab = TA0T-’
be chosen as:
r=y-
e
61
b b = TbO cb = C O T - ’
where B = trace@) and y is a parameter chosen by the
and T is a transformation matrix. In this balanced form, designer.
the controllability and observability gramians, denoted by For the system considered here, it is found that
W, and WOrespectively, of the system eqn. 13 are equal: C = diag( 3.435, 2.6782, 0.7871, 0.0299 )
W, = WO= C = diag(a1, ..., a?,a1+1,..., a,) 1 is therefore chosen as 3. This leads to:
where n is the order of the system; ui 2 ui+l 2 0 ( i =
1, ..., n - 1) are called the Hankel singular values (HSVs);
and I is the number of most dominant (most controllable
and observable) modes.
[ 3.278 -0.567 -1.420 4.535
-0.567
& = -1.420
4.535
8.278
12.52
43.42
12.52 43.42
19.39 70.34-
70.34 344.6 I
- 882
By choosing y = 1 and r = 2.027, the feedback gains system performance is robust and better than that if an-
obtained are: other robust design method proposed in [3] is used.
le = [ 4.847 6.373 2.743 - 12.11 ]
VI. REFERENCES
The frequency response of
[l]M. Aldeen and F. Crusca, “Multimachine power sys-
Ihl(jw)l, where h l ( s ) = -
y l ( s ) is obtained from Figure 1 tem stabiliser design based on new LQR approach,” IEE
U1 (4
when the above state feedback k is connected, is plotted Proc.-Gener. Transm. Distra., Vo1.142, No.5, Sep. 1995,
in Figure 5 as a dashed line. The interaction margin de- pp .494-502.
fined before is 10.26db at a freuency of LI M 2.95 rad./sec. [2] P. Grosdidier and M. Morari, “Interaction mea-
sures for system under decentralized control,” Automatica,
IV. SIMULATION RESULTS V01.22, No.3, 1986, pp.309-319.
[3] Y. Wang, R. Zhou and C. Wen, “Robust ioad-
To test system performance, a step load disturbance of
frequency controller design for power systems,” IEE
APD, = 0.Olpu is applied t o area 1 and the system output
Proceednngs-C, Vo1.140, No.1, Jan. 1993, pp.11-16.
of Af1 is observed. An Integration-Absolute-Error-Time
[4] C.M. Liaw and K.H. Chao, “On the design of an
(IAET) criteria of the following form is also used:
optimal automatic generation controller for interconnected

JIAET = 1000
lo lAfl(t)ltdt
In the simulation study, the linear model of a non-
power system” Int. J. Contr., Vo1.58, 1993, pp.113-127.
[5] B.D. Anderson and J . Moore, Linear optzmal control,
Prentjce-Hall, 1971.
reheating turbine -in Figure 1 is replaced by a non- VII. APPENDIX: NOMENCLATURE
APG
linear model of Figure 8 with 6 = 0.015. This is t o take TT : Turbine time constant, T T ~ =T T ~
= 0.3s.
into account the Generating Rate Constrain (GRC), i.e., TG : Governor time constant, T G =~ T G =~ 0.08s.
the practical limit o n t h e response speed of a turbine. Tp : Power system time constant, Tpl = Tp2 = 20s.
A number of simulations, using the nominal plant pa- R : Regulation paremeter,
rameters and those changed by some percentage, have R1 = Rz = 2.4Hz/puMW.
been carried out for three different cases. -The JIAET val- Kp : Power system gain,
ues obtained are listed in Table 1, where: It>l = Kp2 = 1 2 0 H ~ / p u M W .
(1) Case A: two identical controllers designed in Section T 1 2 : Synchronising coefficient, T12 = 0.545puMW.
3 are added to the system of Figure 1, B : Frequency bias parameter,
(2) Case B: two decentralized area controllers designed B1 = B2 = 0.425puMW/Ht.
by the method of “Designing stabilizing controllers for un- PD : Load disturbance.
certain systems using the Riccati equation approach” [3] K : Integration gain, K1 = K2 = 1.
are connected to the system of Figure 1, and a12 : The ratio between the base values of two areas,

(3) Case C: no additional controller is connected to the a12 = -1.

system represented by Figure 1.


Some selected time response plots are given in Figure Table 1: JIAET values (*: the system is unstable)
-
9, 10 and 11. In these figures, solid lines are for Case A, Test Parameter Case Case Case
dashed-dotted lines are for Case B, dashed lines are for -No Changes A B C
Case C. 0 0% 2.5 6.9 34.4
The above results show that, in comparison with both 1 +5 % 2.7 6.6 60.8
no additional controller and with the cointrollers designed 2 -5 % 2.6 7.2 20.8
by the method give in [3], the system performance is sig- 3 +10 % 2.6 6.3 105.0
nificantly improved by the controllers designed in this pa- 4 -10 % 2.5 7.8 13.2
per; and that this performance is robust against the plant 5 +15 % 2.6 6.1 172.4
parameter changes. This can be seen from the values of 6 -15 % 2.6 8.3 9.3
JIAET listed in the column of “Case A” in Table 1, which 7 +20 % 2.7 8.0 (281.1)”
are all between 2.5 and 2.9. 8, -20 % 2.5 9.0 6.5
9 +25 % 2.8 5.9 (471.2)*
V. CONCLUSION 10 -25 % 2.6 9.5 4.7
A new robust decentralized power system load- 11 +30 % 2.9 5.8 (686.9) *
frequency controller design method is proposed in this pa- 12 -30 % 2.7 10.4 3.7
per. The simulation results have demonstrated that the
10.' ion io' IO'

Figure 6. Bode plot of E; (solid line) and (dashed line)

7q-p (SI-A ')-I bo

~ ~

Figure 7. An equivalent SISO control system


Figure 1. Block diagram of a two-area power system

Figure 8. A nonlinear turbine model with GRC

P i p r e 2. An equvilant MIMO system

I I

Pigwe 3. Independent SISO system design


- 0
h",*<*fO >

Figure 9 Dynamcresponse of A fi Test No.0

, . . . .. . .a. , .. . .. ..
r
I".{S.C >
Figure 10. Dynamicresponse of A fi Test NOS

Figure4. A two-input two-output system


. . .
1W

..... .
a
E
.E 0
d

-100 I
io-' ioo io' io2 .-**a3
Frquency (radisec)

Figure 5. Bode plot of It-' (Rbw))(solid line) and hl (dashed line) Figure 1 1 . . Dynamic response of A f, Test N0.6

You might also like