IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO.

2, JUNE 1995

23 1

Nonlinear Control Methods for Power Systems: A Comparison

V. Rajkumar and R. R. Mohler

Abstract-The design and performance of two nodinear controllers for power systems is explored in t i paper. The designs hs are based on nonlinear discrete-timepredictive control and feedback linearizingcontrol, consideringa simplifiednonlinear singlemachine, infinite bus (SMIB) power system model equipped with a series-capacitor controller. Simulations are presented for nonideal conditions such as bounded control, the uncertainties in the chosen model parameters, and the presence of inter-area-type persistent disturbances.The results indicate that superior performance can be obtained using the nonlinear predictive controller when compared with the feedback Linearizing controller and a standard linear quadratic regulator.

I. INTRODUCTION ARGE faults on the transmission system cause the electrical power to collapse to a small value during the fault. This leads to an acceleration of the rotors of the synchronous generators, due to unequal mechanical and electrical torques acting on the shafts. Rapid detection of the fault and fastacting circuit-breakers are normally used to isolate the faulted sections of the transmission system. In addition to the rapid isolation of faulted sections of the transmission equipment, it is generally recognized that auxiliary control action would be necessary to enhance the post-fault region of stability of the power system. Synchronous generators can be analyzed conveniently using Parks transformation [I], which results in a time-invariant nonlinear system. In the event that the post-fault system experiences small oscillations, simplified linear power system models such as [2] can be used with advantage for analysis and linear controller design. For example, the widely popular linear power-system stabilizers (PSS) [3] are used to provide supplementary damping through excitation control, thus enhancing the dynamic stability limit. Such linear control on the nonlinear power system generally provides asymptotic stability in a small region about the equilibrium and is thus appropriate for the dynamic stability problem, where the primary concem is of providing damping following small disturbances. Nonlinear controllers are suitable for the transient stability problem, by virtue of being designed to stabilize the system for large initial conditions (within the physical limits imposed by the controller). We note that when the impedance of the transmission lines are controlled, for example by a flexible ac transmission systems (FACTS) [4] device such as a thyristor-controlled series-capacitor (TCSC), the control
Manuscript received August 4, 1993; revised August 7, 1994. Recommended by Associate Editor, R. Ravi. This work was supported by NSF Grant ECS 9301 168, EPRI Conract RP 3573-05, and BPA Contract 94BI16785. The authors are with the Department of Electrical and Computer Engineering, Oregon State University, Corvallis, OR 9733 I USA. IEEE Log Number 9410593.

action is multiplicative (product of the state and control) and thus necessitates the consideration of nonlinear models in design for the most effective, economical use of the available control resources. In this context, nonlinear variable-structure control [5] and nonlinear model-based self-tuning control [6] have been proposed for a class of transient stability problems. In this paper, we illustrate the design of a nonlinear-timeseries, model-based, generalized predictive controller for a simplified power system, using rotor angle as the measured output and a TCSC controller. Generalized predictive controllers offer the advantages of being easy to implement in real-time and allowing systematic methods of handling input constraints. Recently, a number of papers have addressed a new class of controllers for the power system, based on the concept of cancellation of the nonlinear dynamics through feedback [7]-[14], using excitation or govemor control. Subject to the availability of an accurate reference model and measurements of the power system, the dynamics of the power system may be cancelled and replaced with some desired linear behavior. This paper also examines a very simple design of a seriescapacitor based feedback linearizing controller and subjects it to three types of uncertainties: 1) bounded control, 2) uncertainties in model parameters, and 3) the presence of unmodeled inter-area-type disturbances. The performance of the feedback linearizing controller is evaluated and compared with that of the nonlinear generalized predictive controller and an ordinary linear quadratic regulator (LQR), which are also subjected to similar conditions.

11. POWER SYSTEM MODELS FOR CONTROLLER DESIGN

The dynamics of the post-fault power system can be described by a set of nonlinear differential equations. Typically, the states can be associated with each machines armature currents and rotor currents resolved through Parks equations, rotor dynamics, automatic voltage regulator (AVR), and turbine-governor dynamics [ 151. Many controller design methods utilize simplified models of the synchronous generator with states composed of the rotor dynamics, change of internal voltage in the quadrature axis, and at least one state to represent the highly nonlinear closed-loop AVR. With the last consideration, we note that the field voltage is typically a state variable and not an independent agent for control, except for a short while during field forcing in the event that the terminal voltage feedback collapses. We address two issues regarding the simplification of power system models for controller design.

10634536/95$04.00 0 1995 IEEE

232

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2. JUNE 1995

A. Considerations Regarding Voltage Regulators

Most modern generators are equipped with AVRs. These devices maintain the terminal voltage of the generator at a specified value and in the process modulate the field voltage and hence the field current, thus supplying the required reactive power to the load. It is well known that the AVR can inject negative damping into the system at high power loadings, leading power factors, and large tie-line reactance [2]. In linearized-analysis terms, this can result in unstable rotor mode eigenvalues at some power loadings. This phenomenon cannot be simulated in models that neglect the AVR dynamics. An adhoc simulation of unstable rotor modes may be done, however, by making the mechanical damping constant negative in the swing equation. The neglect of the AVR in power system models used for feedback control design purposes and the use of field voltage as an independent control variable as proposed in the feedback linearizing controllers of [7]-[ 141 cannot guarantee terminal voltage regulation, except by adding nonstandard AVRs by the incorporation of additional control loops in the feedback linearizers. But in this case, the nonlinear dynamics are not cancelled completely, since the feedback of terminal voltage results in the injection of significant nonlinearities into the system. B. Considerations Regarding Infinite Bus

the origin, and controllable series susceptance

- = wbw dt

d6

where

T , is the prime mover torque, ,,

E is the internal bus voltage of the generator,
is the infinite bus voltage, is the series susceptance, is the machine inertia, is the mechanical damping factor, is the rotor angle in radians, w is the rotor speed in per-unit, W b = base speed, 377 radslsec, U is the control input (additional series susceptance), and suffix e represents equilibrium value. An input-output model can be obtained for the system ( ) ( ) by choosing 1-2,

V B M D 6

Y(t) = 6 ( t ) .
Differentiating (3) twice and eliminating w

(3)

d26
~

dt2

(g)2 +
- -

wbVEsin(6,

+ S)(B,+ u ) / M

- wbTm/M = 0. (4) The assumption of a single-machine, infinite bus (SMIB) may be valid for faults around low-inertia systems which Some properties of linear controllers with the linearized are connected to high-inertia external systems by long tie- input-output model of (4) can now be illustrated, which will lines. There is concern, however, that the SMIB-designed provide insight into local behavior of the controllers to be controllers may not perform well in the presence of inter-area designed. modes. It will thus be necessary to evaluate the SMIB-designed Consider a linear model of (4) as nonlinear controllers on more realistic multimachine models d2A6 with different oscillation frequencies. For an ad-hoc appraisal dtz U-dA6 bA6 = CAU (5) dt of controller performance, a simulation test may be devised on SMIB systems, where the infinite bus voltage is modulated where a, b, and c are appropriate linearization constants. A simple proportional output feedback may be designed for in magnitude by an inter-area-type frequency, an approach ( 5 ) by choosing termed single-machine, quasi-infinite bus in [ 161. There are additional considerations with respect to neglected AU = -KA6 (6) turbine-governor dynamics, reactive power demand, and load modeling, which are not addressed in this paper for brevity. where K is some feedback gain constant. In the following sections, we design a candidate nonlinear Substituting (6) in (3,the roots of the resulting characterpredictive controller and a candidate feedback linearizing istic equation are obtained as controller for power system transient stability. The designs are based on series-capacitor control, and for simplicity of X l , 2 = ( - U f du2- 4(b cK))/2. (7) illustration of the concepts involved and in the interest of For normal operation of the power system, the rotor modes brevity, use a second-order model of the power system. The candidate controllers are subjected to the testing criteria exhibit oscillatory damping, indicating complex conjugate such as those listed above, and for comparison purposes are roots. Clearly, the real parts of (7) are unaffected by the evaluated alongside an ordinary LQR which is also subjected selection of K. Thus controller (6) cannot stabilize system (4) locally, as long as u2 - 4(b c K ) < 0. Second, the roots to the same test conditions. of ( 5 ) cannot be placed at will in the state-space, owing to the dependency of the roots on the single tunable parameter K. 111. DISCRETE-TIME NONLINEAR PREDICTIVE CONTROLLER Now consider the controller as a PD-a proportional + Consider a simplified power system model of the post-fault derivative controller with

single machine connected to an infinite bus, represented with the rotor speed and rotor angle, with equilibrium translated to

AU z= -KIA6 - K2-

dAS dt

___

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2. JUNE 1995

233

By substituting (8) into (9, again check the roots of the we closed-loop characteristic equation to yield
~ 1 , = 2

u ~ + ~ needed -I

(-(a

+ C K ~ )J ( a +

C K Z ) ~-

4(b

+ c ~ 1 ) ) / 2 .(9)

past output measurements, b k , and b k - 1 . The control values for the predictions are obtained iteratively by minimizing (13). Consequently
6k+j

Thus by introducing feedback of the derivative of the output, the real part and the imaginary parts of (9) can be adjusted independently, by an appropriate choice of the gains.
A. Discrete-Time Nonlinear Input-Output Model

= alSk+j-l+

+ b l sin(8, + + ci sin(6, + 6k+j-i)uk+j-i + d .

azSk+j-a

Sk+j-l)

(14)

D. Minimization of the Criterion Function

Criterion function (13) can now be minimized with respect to the controllers U k + j - 1 over j = 1 , 2 , . . . , N , to obtain N nonlinear equations in the future controls. Consider the ith such equation obtained by minimizing (13) with respect to
uk+i-l

System (4) can be discretized locally by using Eulers backward shift approximations for the differentials, using sufficiently small step-size h , i.e.,

Substituting (lo)-( 11) in (4)and rearranging the time-shifted terms, we obtain the discrete, nonlinear input-output equation for the power system as
6k+i = a i S k

x [Ae+j - Ae+j-lm])

+ ~2ur~+i-1 0 =

(15)

+ aaSk-1 + b i sin(6, + S k ) + CI sin(& + 6 k ) u k + d

a2 = -.;/ab; d = -d/ab

where
a1 = - a i / a b ;
bl

= -6\/ab

i.e., fi(u) = 0, for i = 1 , 2 , . . . , N , where U = [ U k , uk+l,. . . , uk+N-l] T . We define the derivative of the jth prediction of the output with respect to the ith control recursively as follows (12) F o r j = 1 , 2 , . . . , N ; i = j , j + l , . . . , N ; S e t A i - A i p l = 0 when j = i , A i + j = CI sin(6, 6k+j-1) when j < i , A i + j = 0 (causality) when j > i , use (16)

c1 = - c i / a b ; and

D / ( h M ) ; U: = - ( 2 / h 2 + D / ( h M ) ) ; l / h 2 ;b: = wbVEB,/M; e: = w b V E / M ; d = -wbTm/h/l.

U;

= l/h2

a; =

where A,+j-1 =[a1

+ bi cos(Se + + ci +
COS(&

6k+j-1)

6k+j-i)uk+j-i].

(17)

B. Generalized Predictive Control (GPC) Criterion The generalized predictive controller is now based on the model (12), by minimizing the following generalized predictive cost function
J = (112)

J=1

EG+]+ (Pl/2) E ( S k + j 3=1

N

- &+j-1)2

+ ( ~ 2 1 2 )C u E + j - l .
3=1

(13)

We note that cost function (13) involves quadratic terms in the output, the discrete rate-of-change of the output, and the input. The rate-of-change of the output terms are included to have derivative effect as discussed earlier for the linearized model. N is the prediction horizon of the GPC.
C. Predictions o the Nonlinear Discrete-Time Reference Model f

Nonlinear equations (15) can be solved numerically [17]. When the controller saturates, additional considerations are necessary to ensure that the controller remains a minimizing controller [18]. The control loop iS closed using the control U k , and the whole procedure is repeated at the next time step. We note that this method provides a powerful method of nonlinear output feedback control. In general, large prediction horizons will be necessary to assure stability. The framework developed allows the implementation of receding-horizon controllers. In this paper, we have considered N = 2 and have not made any special consideration for controller saturation, to keep the procedure simple. LINEARIZING CONTROLLER DESIGN IV. FEEDBACK Nonlinear systems and control have received renewed attention in recent years. For certain kinds of nonlinear systems where the control variable appears linearly, it has been shown that subject to certain conditions being satisfied by the model, it would be possible to devise a transformation to design a feedback control law, which cancels all the dynamics of the system, and replace it with a desired linear dynamical system [191, P O I .

The predictions of the plant output as demanded by (13) over the horizon N are highly nonlinear in the control and may be obtained by running the time index j in (14) from one to N . Note that when j = 1, the initial conditions for the predictions are available as the current and immediate

234

EEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2, JUNE 1995

System ( ) ( ) can be represented as 1-2

d6 - = wbw dt

(18)

dw - = a ( w , 6)

+ b(6)U

dt and is in a form amenable to direct feedback cancellation. Assuming that b(6) # 0, a direct feedback linearizing control law takes the form
U

1-2, pair of (25). When this controller is used on system ( ) ( ) asymptotic stability is guaranteed in a small region about the origin. While this is not suitable for the transient stability problem, it would serve our purpose of comparison with the nonlinear predictive controller and feedback linearizing controller.
RESULTS, AND DISCUSSION VI. SIMULATION DATA,

For easy reproduction of these results, a detailed account of the simulation experiments and some numerical results are provided for validation. The following simulation parameters i.e., XI X, - X,,), are chosen in this study: Be = l/(X; U = [w-T,+VEsin(6,+6)B,+Dw]/VEsin(6,+6). (20) where X i = 2.0 p.u. is the direct axis transient reactance, Xl = 0.5 p.u. is the equivalent transformer reactance, X , = 1 Substituting (20)in (18)results in the equivalent system p,u- is the line reactance, and X,, = 0.5 p.u., is the equilibrium value of the series capacitor reactance, corresponding to 50% dS - = wbw dt (21) compensation of X,. Let the capacitor reactance have a range of &loo% percent compensation of X,. Thus the maximum dw -= v/M. (22) and minimum allowable incremental series susceptance is dt U,, = 0.06667 p.u. and umin = -0.04762 p.u. The machine Now choosing the linear control law as inertia M = 3.25 seconds, internal bus voltage E = 1 p.u. w = -k66 - k,w (23) infinite bus voltage V = 0.99 p.u. and synchronous speed W b = 377 radskec. The damping factor D is varied in each system (21)-(22) can be provided with some desired linear simulation and is listed below appropriately. response. The linear component of the control can be deAll simulations to follow are done at the nominal operating signed by solving the well-known algebraic Ricatti equation, condition, 6 = 50 degrees, w, = 0 and prime mover torque , considering the feedback-linearized plant (21)-(22) for the T, = 0.252795. At a sampling (and numerical integration) (A, b)-pair. time step of h = 0.01 second, and damping factor D = 0.5, the Remark 4.1: Controller (19) fails whenever coefficients of the model (12)are obtained as a1 = 1.99846, S = fn7r - 6 ; n = 0,1,2, , (24) a2 = -0.99846, b l = -0.00382, ~1 = -0.01147, d = 0.00293. which may also be found to be the values at which the columns of the nonlinear controllability matrix [20]fail the A. Effect of Bounded Control linear independence test. As mentioned earlier, the control demanded by (19) can be Remark 4.2: Controller (19) has two components-a feedexcessively large and, in practical systems, would be restricted back linearizing component and a linear feedback component. to the control saturation values. For the above operating The control must be restricted to the admissible control set. conditions choosing D = 0.5, LQR state weighting matrix In the event that U ( X ) is restricted to the boundary of the Q = diag(1,l) and control weight R = 1, the closedadmissible control set, (19)no longer guarantees feedback loop feedback linearized system (21),(22) has the following linearization. In the transient stability problem, the initial posteigenvalues: X l , 2 = -7.76173 f j7.6142, corresponding fault states are typically large, and thus the control demanded to the controller gains K = [1,49.5126]. For the same by (19) to linearize the system can be very large and can LQR weights, the linearized closed-loop system (25)-(26) cause controller saturation. In this event, the residual nonlinear has the following eigenvalues: Xl,z = -5.7785 f j7.613, dynamics will not behave as the original nonlinear power corresponding to controller gains K = [.758679,48.8674]. system, but with different properties. The GPC uses a weight of p1 = 100 on the rate-of-change of output and p2 = 0 on the control. The fault is a shortV. LINEAR QUADRATIC REGULATOR circuit simulated by setting E = 0 P.u., cleared after 10 AC A linear quadratic regulator may be designed on the nominal cycles, simulated by resetting E = 1.0 p.u. Without control, linearized system of (1)and (2)around the origin. In this case the power system (1)-(2) is unstable. Fig. 1 shows the application of the feedback linearizing dx -=Az+bu controller ( 19) (FBLC), the nonlinear generalized predictive dt controller (13)(GPC), and the LQR (26).The nonlinear GPC and the controller is chosen as yields the best damping among the controllers, in the shortest U = -Kx (26) possible time, using the least amount of control. The FBLC demands more control (needs three positive peaks of control) where K is the vector of feedback gains, obtained by solving to do the same task less efficiently than the LQR (needs the algebraic Ricatti matrix equation, considering the (A, b) two positive peaks of control), as shown by the less damped

= [w - a ( w , 6 ) ] / b ( 6 )

(19)

+ +

s . .

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2, JUNE 1995

235

1001

0.021

1001

0.021

'

2 0

'
0 . 0
10

10

Time (seconds)

Time (seconds)

Bold Line - Nll Predictive COntKll

Bmken Une LQR contrd Dolied Une - UneadzlngControl

I .

10

!o\r
-1001

5 Time (seconds)

10

-0.02'

10

Time (seconds)

Bold Line - NII Ptedtcbve Contml

Broken Line - LOR Control Doned Une - UneadzingContml

I' I 1 1 ;) I

I I 11'1

1 1 1 1

-0.05

10

Time (seconds)

Time (seconds)

Fig. 1. Effect of bounded control on nonlinear GPC,feedback linearizer, and LQR. Response to 10 cycleshort-circuit.
1

Fig. 3. Response to nine cycleshort-circuit with nonlinear GPC,FBLC and LQR. Based on imperfect reference model-10 percent error in damping factor D , for unstable operating condition.

I
5 10

-0.021

I 10

Time (SeCMlds)
0.11

Time (seconds)

g o

1b+
l I I' I , >

~otted~ne-~lnearizingcontrol

,$. _- - - - - - -

-O050

'

10

Time (seconds)

Fig. 2. Response to 10 cycleshort-circuit with nonlinear GPC, FBLC and LQR, based on imperfect reference model-20% error in line susceptance B.

With D = 0.5, Q = diag(1, l ) ,R = 1 for the linear part of (19) and for (26), and p1 = 100, p 2 = 0 for (13), the result of a 10-cycle short-circuit is shown in Fig. 2. While all controllers stabilize the system, the offset seen in the rotor angle and the control input, caused by the FBLC and nonlinear GPC controllers, bring out the next important feature of nonlinear controllers. If the reference model is inaccurate, the feedback linearizing controller would feedback-linearize the system around a wrong post-fault equilibrium. Similarly, the nonlinear GPC would conduct the post-fault system to a wrong equilibrium. In contrast to this, the LQR which was also designed on a similar imperfect reference model is seen to bring the power system to the proper (only) post-fault equilibrium at the origin. It may be noted that the LQR exhibits poor damping compared with the nonlinear GPC. Similar observations can be also made with regard to uncertainties in the infinite-bus voltage, a fictitious far-end quantity.

swings of FBLC. The LQR neglects all nonlinearities in the system, and as a consequence, displays degraded performance when compared with the nonlinear GPC. This study brings out the following features-i) the control input demanded by the feedback linearizing controller (19) is greater than the control input demanded by LQR (26), since (19) spends considerable effort in linearizing the system, and ii) LQR's exhibit degraded performance on the nonlinear power system when compared with the nonlinear GPC. B. Imperject Reference Models The nominal susceptance term Be is seen to incorporate the line reactance X,, the measurement of which involves considerable uncertainty. With this consideration, the plant (1)-(2) uses the Be as outlined earlier, while FBLC (19) and nonlinear GPC (13) uses a value of Be which is 20% in error, i.e., B' = 0.8Be. For appropriate comparison, the linearized model (25) also uses B' = 0.8Be to design the LQR (26).

C. Effect o Uncertain Parameters for a Dynamically f Unstable Situation By virtue of its simplicity, model (1)-(2) can never represent the dynamic instability seen in AVR-equipped power systems. We may artificially endeavor to make (1)-(2) unstable, however, by choosing the damping factor as a negative quantity. It may be noted that the damping factor D is a fictitious quantity, lumping the effects of mechanical and electrical damping in (1)-(2). Thus it can be determined only approximately for a practical power system. Let D = -1.91 in the model ( l t ( 2 ) . Let D = l.lD, representing a 10% error in the ' estimation of the damping factor, be the value of D used in the feedback linearizer (19), the nonlinear GPC (13), and the linear reference model (25). For Q = diag(1,I),R = 1,p1 = 100, p2 = 0, and a nine cycle short-circuit, the response with the controllers (13), (19), and (26) on the system (1)-(2) is shown in Fig. 3.

236

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2, JUNE 1995

100,

0.02 i

1-0 0

10

-0.02

Time (seconds)

5 Time (seconds)

10

stabilizes the system, and the nonlinear GPC offers the greatest effectiveness for stabilization and damping. Similar observations can be made with respect to unmodeled prime-mover dynamics when comparing controller perfor0.05cos(wrt) in the mance (e.g., consider T,(t) = T,, simulations, and use only T,, in the control design, where w, is a frequency close enough to resonate with the generator shaft mode, say w, = 4.54 radshec).

7 k d B o
E

Bold Line Nn Predictive COnIrol

VII. CONCLUSIONS
This paper investigates the performance of a nonlinear generalized predictive controller, feedback linearizing controller and an ordinary LQR for a simplified power system under nonideal conditions. Owing to physical limitations on the magnitude of the control, transient stability controllers initially operate in a bang-bang mode. Nonlinear GPC is seen to provide a powerful, effective option for the nonlinear output feedback control of the power system. The nonlinear GPC is seen to possess excellent robustness properties and requires the least control when compared with the LQR and the feedback linearizing controllers. A major advantage of the nonlinear GPC over the other two controllers is that it allows a systematic way of handling control constraints albeit numerically intensive, which may restrict the look-ahead horizon for real-time implementation. Feedback linearizing controllers demand excessive control effort, compared with LQRs, due to the additional control effort expended in feedback linearizing the nonlinear dynamics. When the control saturates, the nonlinear dynamics are not cancelled entirely, and the system is left with residual nonlinear dynamics with properties that are different from the original nonlinear power system. Feedback linearizing controllers need a perfect reference model and measurements to provide exact cancellation. Uncertainties in the reference model can lead to deteriorated robustness of the controller. There exists the danger of destabilization, by imperfectly cancelling power system dynamics which possess an unstable equilibrium, e.g., dynamically unstable situations. The presence of dynamic uncertainties, such as time-varying infinitebus voltage, can lead to nonrobust performance of the feedback linearizing controller. LQ regulators perform worse than the nonlinear GPC, but better than the feedback linearizers, for identical conditions. Although the performance of LQR on nonlinear systems is suboptimal, it is seen to have good robustness properties. ACKNOWLEDGMENT The authors wish to thank W. A. Mittelstadt and D. Maratukulam for their contributions. REFERENCES
[ I ] P. M. Anderson and A. A. Fouad, Power System Control and Stabiliry. Ames, Iowa: Iowa State Univ. Press, 1979. [2] F. P. deMello and C. Concordia, Concepts of synchronous machine stability as affected by excitation control, E E E Trans. Power Appr. Sysf., vol. 88, pp. 316-329, 1969. [3] E. V Larsen and D. A. Swann, Applying power system stabilizers, Pts. I, I1 and 111, IEEE Trans. PowerAppar. Sysf., vol. 100. pp. 3017-3046, 1981.

0.05

Broken Line - LOR Control

Dotted Line - Linearizing Control

-0.05

5 Time (seconds)

10

Fig. 4. Response to nine cycleshort-circuit with nonlinear GPC, FBLC and LQR, in the presence of persistent inter-area-type disturbance.

It may be observed that nonlinear GPC (13) performs best, and rapidly damps the rotor oscillations, using the least control input. The feedback linearizing controller fails in stabilizing the system and results in an unstable transient. Under identical conditions, the LQR controller is seen to stabilize the system to the correct equilibrium, however, using twice the control demanded by the nonlinear GPC. This result indicates the potential danger that exists by improper feedback cancellation of unstable dynamics, a situation similar to the inexact cancellation of a nonminimum phase zero by an unstable pole in linear systems. We emphasize at this point that for values of D more positive than - 1.91 and for shorter-duration faults, the feedback linearizer may stabilize the system, but is inferior in performance when compared with the nonlinear GPC and the LQR controllers.

D.Effect of Inter-Area-Oscillation-Type Disturbance

In the context of interconnected operation of power systems, the presence of various inter-area and local modes of oscillation is inevitable. This situation is simulated approximately, by modulating the infinite-bus voltage in magnitude by a small frequency, representative of an inter-area mode. Accordingly for the simulation of the plant (1)-(2), we choose V ( t ) = VO- 0.125cos(3.77t), corresponding to an inter-area fundamental frequency of 0.6 Hz. Further assume that nonlinear GPC (13), controller (19), and linear reference model (25) are provided with the constant value of VOonly. Choosing D = 0.5; Q = diag(1, l ) ,R = 1 forboth controllers (19) and (26), and p 1 = 100 and pz = 0 in (13), the response to a nine-cycle short-circuit is shown in Fig. 4. Again, the feedback linearizer can be observed to fail in stabilizing the system, while the LQR stabilizes the system under identical conditions. The nonlinear GPC performs best, and nearly rejects the effect of the persistent disturbance. It may be noted that for smaller faults and lower inter-area frequencies, the feedback linearizer may stabilize the system. Under identical circumstances, it may be noted that the LQR

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 3, NO. 2, JUNE 1995

237

[4] N. G. Hingorani, Flexible AC transmission, IEEE Spectrum, vol. 30, pp. 4 M 5 , Apr. 1993. [5] Y. Wang, R. R. Mohler, R. Spee and W. Mittelstadt, Variable structure FACTS controllers for power system transient stability, IEEE Trans. Power Syst., vol. 7, pp. 307-313, 1992. [6] V. Rajkumar and R. R. Mohler, Bilinear generalized predictive control using the thyristor-controlled series-capacitor, in Proc. IEEE PES Winter Meeting, Paper 94 WM 194-1 PWRS, New York, 1994. [7] R. Marino, An example of a nonlinear regulator, IEEE Trans. Automat. Contr., vol. 29, pp. 276-279, 1984. [SI Q. Lu, Y. 2. Sun, and G. K. F. Lee, Nonlinear optimal excitation control for multimachine systems, in Pmc. IFAC Conf Power Sysr. Modeling Contr. Applicat., Brussels, Belgium, 1988, pp. 27-32. [9] Q. Lu and Y. Z. Sun, Nonlinear stabilizing control of multimachine systems, IEEE Trans. Power Syst., vol. 4, pp. 236-241, 1989. [IO] M. D. Ilic and F. K. Mak, A new class of fast nonlinear voltage controllers and their impact on improved transmission capacity, in Proc. I989 Amer. Contr. C o n f , Pittsburgh, PA, 1989, pp. 1246-51. [ I 11 J. W. Chapman, M. D. Ilic, and C. A. King, Stabilizing a multimachine power system via decentralized feedback linearizing excitation control, in Proc. IEEE PES Summer Meeting, Paper # 92, SM 540-5 PWRS, 1992.

[12] Y. Wang, L. Xie, D. J. Hill, and R. H. Middleton, Robust nonlinear controller design for transient stability enhancement of power systems, in Pmc. SIst IEEE Con$ Decis. Contr., Tucson, 1992, pp. 11 17-1 122. [13] J. W. Chapman and M. D. Ilic, Some robustness results for feedback linearizing control of generator excitation, in Proc. 3Ist IEEE Conf D e c k Contr, Tucson, 1992, pp. 1123-1 128. [I41 F. K. Mak, Design of nonlinear generator exciters using differential geometric control theories, in Proc. 3Ist IEEE Con$ Decis. Contr.. Tucson, 1992, pp. 1149-1 153. [I51 Y. N. Yu, Electrical Power System Dynamics. New York: Academic, 1983. [I61 A. M. A. Hamdan and A. H. Nayfeh, The effect of nonlinearities on the response of a single-machine-quasiinfinite busbar system, IEEE Trans. Power Syst., vol. 4, pp. 843-849, 1989. [I71 G. R. Walsh, Merhods ofOptimization. United Kingdom: Wiley, 1975. [18] T. T. C. Tsang and D. W. Clarke, Generalized predictive control with input constraints, in IEE Pmc., part. D, vol. 135, 1988, pp. 451-460. [ 191 A. Isidori, Nonlinear Control Systems: An Introduction. Berlin: Springer-Verlag, 1989. [20] J. J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs, NJ: Prentice-Hall, 1991.