760 IEEE Transactions on Energy Conversion, Vol. 10, No.

4,December 1995
NEURAL-NET BASED COORDINATED STABILZZING CONTROL FOR THE EXCITER
AND GOVERNOR LOOPS OF LOW EfEAD HYDROPOWER PLANTS

M. Djukanovic M. Novicevic Dj. Dobrijevic B. Babic Dejan J. Sobajic Yoh-Han Pao

Electncal Engmeenng Institute "Nkola Tesla'kJUGEL and EPS Power System Control Departments of Electrical Engineem and Computer Science
ul Koste Glavinica 8a EPRI Case Western Reserve University,
Belgrade, Yugoslavia palo &, CA 94304 and AI WARE, Inc., Cleveland, Oho, 44106

Abstract

This paper presents a design technique o f a new *ve optunal Em-excitation output (voltage regulation mput)
controller o f the low head hydropower plant using artificial neural x!,,,x2,, x, - voltage channel state variables
networks (ANN). The adaptive controller is to operate in real time to xlj,x2j,x+,x4r- frequency channel state vanables
improve the generating unit transients through the exciter bput, the guide T, ,T,. ,Th,k,. ,ku-voltage channel time constants and gains
vane position and the numer blade position. The new design procedure is 2;,,Tzj,~j,k~j,kj-frequ~cy channeltnne constants and gams
based on self-organization and the predictive estimation capabilires of
neural-nets implemented through the cluster-wise segmented associafive TW,k f l - f i htrme constant and gain
memory scheme. The developed neural-net based controller (NNC) whose E, -aw"ry mriable-mput to the exciter
control signals are adjusted using the on-line measuIements, can offer
better damping effects for generator oscillations over a wide range of
operating conditions than conventional controllers. D@al simulations of
hydropower plant equipped wah low head Kaplan turbine are p e r f o m d
and the comparisons of conventional excitation-governor control, state- 1. WTRODUCTPON
space optimal control and neural-net based conQol are presented. Results
obtained on the non-linear mathematicalmodel demonstraie that the effects Classical famdiar automahc voltage regulator (AVR) and speed
of the NNC closely agree with those obtained using the state-space governing system control theory has been widely used m the design
multivariable discrete-timeoptimal controllers. process. The speed control and the excitahon control were considered as
independent and decoupled processes charactenzed by different tnne
Kev words: Neural-nets, Adaptive pattern recognition, Optimal Conhol scales, since the speed loop was fwly slow m cwlpmsm .cvlth the
exatation loop. In recent years, considerable attention has been drected
toward Ihe problem of Improved desigmg turbme govemors, because of
NOMENCLATURE rts lnnuence on the d y n a r c performance of hydro power plant (IIPP)
udil recently analog d e n m have d o m t e d the scene of generatmg umt
h -plant's head controllers. D@al govemors opened the door for nnplementmg more
t,=h-l waterhammer complex algorithms due to their flernbility Digital adaphve control
h, -plant's steady-state net head schemes have demonstrated the efficiency of Implementahon of different
q-turbme flow identification techniques (maximum lkelhood, recursive least squares,
w ,w m ,w, - m a c h e speed, reference signal and synchronous speed &m"tal variable) and control tecbmques (pole shftmg, " m u m
a-gmde vane opemg variance, o m quadrahc regulators) [l] As a result the speed loop
y -main servomotor piston posihon starb3 interfering wah the excitabon loop, so that mutual mterachon of
y, - auxlliary servomotor piston position these control loops pronded possibihhes for obtaming better transients and
q-runner blade angle wider stabdity maxgnx.
p-output of temporary speed droop
Applications of modem optnnal control system design methods
T,,T ,Toad -turbine torque, eledncal torque, *ban= load t o r p
using the muhmiable system theory and linear optnnal regulator concept
T,-water s t a m tnne in plant ContrOp are well documented [1-61 These approaches are drected
bp,b,-permanent droop and temporary droop toward & that the controller parameters are opbmd for a wide
T q,Tl,T2-guide vane servomotor, dashpot, a d a r y servomotor, and range of operatmg condihons. The model reference control concepts and
runner blade time constants the self tuned regulator concepts have been found most effechve m
7; - T, turbme flow and torque sensitinty coefficients categories of adaptive stabhers
6-rotor angle
M-umt mertia constant On the other hand, intelligent control concepts such as
D-dampmg coefficient le- control and expert control are mtroduced m power system studm
-termmal voltage to overcome lnherent complexlhes and uncerknties that are always
present m a vanety of forms Learmng control methods are smlar to
e,-electromotwe force due to flux linkage which corresponds to the
adaptwe control methods Both are faced with the unknown or
transient period mcompletely h o m mformahon about system's dynamics and
h b a n w , and both attempt to change the control scheme durmg
95 SM 452-3 EC A paper recommended and approved system operation. Both classes of methods
by the IEEE Energy Development and Power Generation &or witzlout external supemsor The
Committee of the IEEE Power Engineering Society for that learmng control methods uhlize p
presentation at the 1995 IEEE/PES Summer Meeting, gradual improvment in controller perform
July 23-27, 1995, Portland, OR. Manuscript submitted
June 24, 1994; made available for printing This paper presents the low head hydropower plant nonlinear
April 27, 1995. mathemahcal model equpped mth the static (ST-1) excitation system with
multivanable voltage regulator, from whch a lmearized model is deduced
T h e theory of optimal control is applied to the lmearized model to generate
state space bencbmark model for comparison with neural-net based
controller
0885-8969/95/$04.00 0 1995 IEEE
 761

$- ; -Y)
The three control configurations (conventiom& state-space and
NNC) are then implemented on the non-linear simulation model in order to (Yl (6)
assess their dynamic performances when the system is subjected to large
dP 1
disturbances. The effectiveness of the proposed neural-net controller is -=-(Pk(YI>h")-d (7)
illustrakd through the comparison of dynamic performances obtained dt T2
using conventional controller of the stabilizer type and the state-space The constraints must be imposed on:
optimal controller. -the runner blade position ,p
, I p I pmax

-the speed of fhe runner blade servomotor

max
2. PROBLEM FORMULATION -the main servomotor piston position y,, < y < y,,
-the auxiliary servomotor piston position y,,, < y, IylmW
The basic components involved in the representation of a slngle
Kaplan turbine-generator of hydro power plant W P ) "Djerdap 1" with a -the main servomotor speed
local load connected through a trmsdssion line to an W e bus are
t
shown in the Figure 1 .

TURBINE
INFINITE
dwwwa BUS

1 I
GUIDE Y=G+jB
VANE S.M. BLADE S.M.
4 f
U
30-J
f -
INTERFACE
OUTPUT t
........_.
LOOP
FEEDBACK 4 comRoLmR I Figure 2: The Schematic Illustration of the Temporary Droop Type HPP
Govemor.
Figure 1: Coordinated Stabilizing Control Scheme for the Exciter and
Govemor Loops of Low Head Hydro Power Plant. Well known third-order synchronous generator model has been
widely used in many power systems dynamic studies [ 5 ] . This model
The regulation of water flow in hydro-electric power plant with incorporates the change in flux hnkage of the field winding and can be
Kaplan turbines is performed using both the guide vane position and the described by the following set of differential and algebraic equations:
runner blade position. When analyzing large disturbances one must take dw 1
- (Z-T-DW)
into consideration that the unit flow and speed of rotation move along dt M
experimentally obtained, the three dimensional "propeller" turbine
characteristics. These characteristics are obtained from prototypes of 2
dt = W , ( W - 1)
Kaplan turbines in the "propeller mode" (with fixed runner blade position
and for different values of the parameter (p) [7]. The HPP with a Kaplan
turbine is characterized by relatively short pipelines and incompressible
the water and pipes, so the stiff waterhammer is given by: 5 = vdid+ v,i, (11)
K=$j-q (12)
The generator is equipped with the static (ST-1) excitation s y s t h
The turbine torque and the flow in Kaplan turbines can be of the power system stabilizer type. This excitation system that includes
represented as a fimction of four variables: multivariable voltage regulator with four channels (voltage channel, first
T, = m(a,P>Ww,h) (2) derivative of voltage deviation, frequency channel, first derivative of
4 = d a , P>@>h) (3) frequency deviation) is shown in Figure 3. Mathematical model of this
The turbine characteristics are usually given only for a narrow region excitation system consists of 8 differential equations:
s u r r o u n d q the rated operating point. In order to obtain the turbine
characteristics for all possible operational conditions of the turbine we -EpD)
dt
used the method of linearizationby segments [7] based on the fact that the
flow and torque characteristics are smooth and change gradually with
speed in the turbine quadrant.

The mathematical model of the turbine govemor includes the

.temporary droop and the feedback taken from the auxiliary servomotor.
The structural block diagram of the govemor is shown in Figure 2.
The relationship between the guide vane opening and the servomotor
position a = a ( y ) and the relationship between the auxiliary servomotor
piston position and requested runner blade anglepk=pk(yl,hn)are
approximated by a set of linear segments. The mathematical model of the
turbine governor is described by the following set of equations and
constraints [SI:
- 1
4 1 =-- (W-"+P+bpY1) (4)
dt T,
(5)
762
A variety of methods are available for synthesizing a non-negative definite
qmnetric state weighting matrix Q and a positive definite symmetric
matrix R, which attribute costs to the states and control efforts. They may
VOLTAGE CHANNEL be chosen so as to shift the eigenvalues of the linearized system to the left
plane as far as it feasible. This, however, may not give the optimal
Iv
response because of the inherent non-lineanties in the modeling and some
ceding values, reahed after occurring severe dsturbances. Consequently,
in practice designers mainly use engineenng judgment and "trial and error"
approach by comidering good compromise between minimization of some
I variable's deviation, overshoot, available range of control signals and
'ref

FREQUENCY CHANNEL) k I lw damping. In this paper the matrices Q and R are defined through :
(i) the " i z a t i o n of excursions of the rotor angle deviations about the
steady-statevalues;
(ii) an inspection of the closed loop system eigenvalues,
(i) use of the fdlavailable range of control signals U
If In order to achieve the compromise between these targets we used non-
diagonal matrix Q obtained from the output matrix C as.
Q = CTC (27)
C = [CI, cz, c 3 , c 4 >C~,CS, c 7 ]
Determination of the feedback matnx IC that minimizes the above PI
is the well-known state-space opt& regulator problem. K is obtained
Figure 3: The Structural Block Diagram of the Static Excitation System from solution of the discrete algebraic matrix Riccati equation as'
(ST-I) of the Unit with Multivariable Voltage Regulator.
S - A'SA +ATSB(R+B'SB)-'BSTA -Q =0
(28)
The complete non-linear plant model in the state space is described K = (R + BTSB)-'BTSA
by a set of 16 differential equations and appropriate algebraic equations
(1-21). Since the design procedures of the optimal controllers are based on This statespace optimal controller was used as.
the theory of linear systems it is necessary to linearizenon-linearities about (i) the benchmark for comparison with the neural-net based controller;
steady-state operating point. The discrete linear state space model (given in (ii) the source of patterns that are used in the Supervised Learmng (SL)
Appendix 1) can be organized in the following matrix form: mode of neural-nets.
x( m + 1) = Ax(m)+ Bu( m) + Fw(m)
where: 3. DIRECT MODE COORDINATED STABILIZING HPP
CONTROL WITH NEURAL-NETS

Artificial neural networks (ANN) [12-171 are parallel hstributed

The mathematical model given by (22) mcludes the model of the processing systems, capable of synthesizing a complex, transparent, and
conventional multivanable voltage regulator (mth voltage channel and lnghly nonlinear mapping from space of input features to output
frequency channel) and the model of a u d a r y servemotor If the space. The unsupervised leaning concept usmg neural-nets proposed in
convenhonal control scheme is to be replaced by the optunal state-space L3,14] can be eEectwely used to dmover smilarities among unlabeled
control scheme, as shown m Figure 1, equahons (4)45) as well as patterns. The FLN-net architecture mtroduced by Pa0 [13] allows
equations (14)-(21) correspondmg to the convenhonal voltage regulator unsupervised learning, supervised leannng and associative retrieval to be
and the auxiliary servo-motor are onutted [SI In th~sway the system state carried wth the same net configuration and with the same data structure.
vector x ,the system control vector U and the system hsturbance vector w More detailed descnption of neural-net technology is beyond the scope of
are defined as follows tbs paper, and an interested reader can found complete information in Ref.
~31.
X = [ A @ AJ, &,,MRD
,A9, AY, A v r (24) Neural net technology is eminently suited to the organmition and
management of sensor data It is also well suited to the momtoring of
systems, identifying system states, interpretmg the sigmficance of groups
of sensor readmgs and predicting the llkely evoluhon of the system
Clearly discrepancy between predicted and observed values can alert the
where U _ is control input to excitation system, ugm is control input to neural net to c h g e s in the nature of the process bemg momtored and can
occasion the posting of alerts or adaptive updatmg of the net's model of the
speed governing system. In the optimal control scheme the control inputs
observed process.
are generated by means of feedback from all the seven states (24) with In t h i s paper the feedback coordmated stabdmng control law is
feedback constants to be determined in accordance with an
treated as the mapping from the space of actual system
optimality criterion.
measurements to the space of syntheslzed control actions as shown in
However, having in mind that the stab- signals w
ill be of Figure 4 The neural-net control system shown m Figure 4 consists of
discrete type and renewed at every sampling time-AT,, we used well- predzctor and controller The role of neural predictor is to learn the
known discrete time state-space optimal controller. For hll state feedback, mapping of the closed-loop system dynmcs under dfferent operating
the control vector U is constructed as a linear combination of all states, i.e. conditions "his mappmg can be descnbed as the transfoimation.
u(m) = -Kx(m) mAT,,, I f I ( m + l ) O s w m=1,2 ...N (25) ~ (=
iP) ( F ( Y , ~ , J ' , Q ~ J '=>p)( ~ ( t ) , ~ ( t - l A f ) , u (W
t- , .
The feedback matrix K is to be determined so that a certain (29)
,~ (-fkk), (f - k k ) , ~ ( -f nAt),J', Q. 2,Y)
performance index (PI) is minimized. A convenient dmcrete optimal where.
control performance index has the quadratic form:
y(i) = [ m ( t ) , y ( t ) , m ( t ) - m(t - A~)]T- is pattem vector at tnne (t)
+u'(m)Ru(m)) Q = Q' 20 R =RT)O (26)
PI = O.S~(x'(m)Qx(m)
0 In the tralIllng phase p is identified on the basis of the hstory data as
 763
In the training phase g is identified on the basis of the history data as:
A
while in the operation the estimate of u(t) is calculated given function (33)
A

PNNETas: while in the operation the estimate of u ( t ) is calculated given function

,g, as:
?t) = pNNEr (F(y,uJ',Q,Z,Y)) (3 1)
M e r this mapping is learned, the predictor infers and maintains a u(f) = s,%?,%.T(WY>u)) (34)
distributed representation of a dynamic model of a system using the real- The "hats" in equations (30-34) are mathematical symbols for training
time information. The training samples for predictor are collected from the phase (equations (30) and (33)) and consulting phase (equations (31) and
off-line closed-loop system operation, controlled by optimal state-space (34)). During the training phase the neural-net is confronted with
controller. In consulting, the predictor provides the estimate of y(t), representative examples of associated inputloutput pairs. Consequently ,
calculates the prediction error and decides which supervised learning (SL) the target during training phase is to leam autonomously inputloutput
modul should be used by neural controller. If the prediction error is A
smaller than prespecified tolerance, the control is continued with current mapping that is denoted by,,g . During the consulting phase
SL modul. Otherwise, the predictor activates on-line re-clustering process synthesized weights and thresholds are used for estimation of output
that identifies SL modul which will produce adequate (near optimal) A

control signals. Consequently, the predictor serves as the real time pointer control values, denoted by U([). Functionality of the training and
to the "best" SL modul that contains (memorized in the hetero-associative consulting phase is used in our approach to support a critically important
memory of the functional link net) the sequence of control signals for near- task performance capability, namely that of learning a model of a process.
optimal trajectory. It should be noted that control action generated by neural-net
Unsupervised learning within this concept activates one of many system is of discrete type. Consequently, the effects of this control are
indvidual neural-nets (SL1, SL2,... SLN) trained with supervised calculated at the end of each discrete time interval. New values of
leaming for different operating scenarios. The neural-net trained for control errors are then evaluated and control error buffer is refreshed
unsupervised leaming calculates Euclidean distances of each input providing new control vector for the next time interval.
pattern to any of the existing cluster centers, and assigns the cluster A direct mode coordinated stabilizing HPP control system shown
number (one of the N previously trained neural-nets) to which the input in Figure 4, is intended for real-time adaptation of control action after
pattern is most similar. It has been shown that learning complex concepts occurring unpredictable disturbances using basically the same limited
based on presented examples is greatly facilitated if data can be number of measurements as a conventional control scheme. The
partitioned in clusters, and if every cluster is related to a different sub- requirement for additional variables to be introduced in the state-space
concept [13]. This reduces the dimensionallity of the problem, and optimal controllers, either measured or estimated, is avoided. Therefore,
enables subsequent learning paradigms to perform in well-defined the use of state estimators or observers in the reconstruction of
circumstances. additional variables required by state-space optimal control is not
The neural-net controller is designed to generate the stabilizing necessary.
control actions, utilizing a distributed model of the inverse system The approach proposed in this paper is the development of our
dynamics. It can be described as the transformation: ideas presented in [21]. However, it should be noted that completely new
role of the neural predictor is formulated, input-output mapping is
u(f) = g(H(y,u)) = g ( Y ( t ) , Y ( t - 1Af),u(t- w,... (32)
designed in different way, and the concept of near-optimal (discrete Riccati
equation based ) NNC is introduced.
..., ~ ( - kAf),u(t
f - kAt),._.... ~ (- ntA f ) )

Prekcbon error based

on-lme re-clustering

Figure 4 Block diagram of a hrect mode neural-net based coordmated stabilizmg control system
764
It is also mterested to note some Important remarks comparing the Fun&onal Link Net was more than five tnnes faster than w t h the classical
proposed approach with the adaptive control algonthm using a look-up GDR neural network configuration w t h 50 nodes m one hidden layer The
table, presented m Ref [18] A set of controller gams m [18] is first training b e s per iteration on the IBM PC486/33 were 0 14 seconds and
computed off-lme usmg the pole-assignment method for each zone of P-Q 0.77 seconds, respectively for FLN and GDR neural-net archtectures
plane, and stored m computer memory The digtal computer monitors
system operatmg conhtions and picks up a set of suitable gam sethngs Finally, the proposed neural-net based controller was confronted
from the look-up table at each samphng instant The on-line use of a look- with several small as well as large hsturbances Due to the limited space,
up table by an excitation controller was described also m (191 However, m t h s paper just the results of a smulation of 3-phase to ground short-
m these mvestigations it was found that the gam changmg by the mcGt on generator terminal bus, cleared after 0 10 seconds, are
controller, as the system moves from one operatmg zone to another during pTESen&d:
a transient, m itself constituted a msturbance and would under The maximum excitation voltage E,,, was specified m
condtions result m a system transient response &at was less damped than surtulatons to 2.73 pu For the purpose of comparison the response
wth a fixed gam controller [20] cuves for the system's variables equipped with conventional multivanable
In t h ~ spaper we avoid the need for the gain changmg by the voltage regulator, o p M state-space controller and neural-net based
controller, havmg m mmd that look-up tables are, m some sense, replaced eontroller are shown m Figures 5-6 On these Figures rotor angle,
by parallel dstnbuted hetero-associative memones reahzed on the t e d voltage, electrical torque, excitation control signal, turbme flow,
computational platform of Functional Lmk Neural-Nets Moreover, the guide vane opening, mechanical torque and govemor control signal are
performance of the NNC was exarmned lfthe &sturban= was as severe as presented
a temporary three-phase short-circuit, thus causmg much faster transients Comparing w t h the case of applymg the conventlonal
Although the behamor of A N N s is condtioned through the multivanable voltage regulator, the system performance 1s hgHy
t r m g phase, the mportant charactenstic of the neural controller 1s that unproved in case of applymg the proposed controller At the same tune
the tramed NNC is capable of generalmtion over the set of previously comparing wth the case of apphcation of optimal state-space controller,
leamed mstances and is able to perform m new c i r c e c e s The whch was used as a benchmark m t h ~ swork, the system performance
advantage of neural-net control design techque 1s also that gene-on closely agree when NNC is apphed Very slight degradation of oscillatory
capabilities of neural-nets enable the possibdity of coordinated optimal processes was expected h a m g m mmd that optvnal state-space controller
control of exciter and govemor loops This is m clear contrast to the is based on fult states feedback, which is Impossible to Implement, whle
lmtation of the smgle-mput-smgle-output (SISO) design techniques, the NNC is based on three discrete-type measured signals, i e the speed,
whch require optmation of two control loops separately lerminal voltage and the speed acceleration of the generator
The focus in this paper was to propose a neural-net based
COoIdinated stabilizing control as an alternative, emergmg from the
4. RESULTS OF TESTING THE NEURAL-NET BASED iutdigent ~ontrolconcepts [22,24, 251 Certainly, the proposed control
CONTROLLER scheme don't pretend to be general m a sense that it fits any possible
operating regime or any disturbance The generalization capabilities of
To demonstrate the efficiency of the proposed neural-net based neural nets depend on how well the selected patterns m the tranng phase
controller, several smulations were performed By following the design represent the entire pattem space The larger number of t r m g patterns
procedure explained m Section 2 and after some trial and error, the mcreases the likewload of the precise generalization However, an
followmg C and R matnces were chosen excessively large t"g data set may be tnne consurmng to obtam
C=[15,30,05,0 1,l 0,2 0,l 01 Generalization is achieved when the information processmg system
0
presumes to formulate a mappmg between the mput and output pattem
Q = CTC spaces based on h t e d knowledge made avsulable to it by the traunng set
R = dzagonai[1,500] of pattem pam For practical purposes, it is hghly desirable that neural
nets leam from a small number of examples and are able to mterpolate and
On the basis of these values for C, Q and R matrkes the gam
extrapolate in the space of its mtemal representation so that large regions
matnx K of optunal state-space controller is found usmg equation (28) A
of the control space can be covered. A smgle neural network, irrespective
companson of the eigenvalues of the HPP system equpped wth the
of a sm, IS the least &ely canhdate for this concept It is obvlously
optmal state-space controller as well as with conventional multivanable
unrdshc to require that the proposed control scheme should be tramed
voltage regulator of the stabilizer type shows that the damping of the aU
slightly damped modes is sigmficantly mcreased Th~sd provide that the for each possible combination of operatmg condtions and dsturbances
transient response of the HPP equipped wth the optrmat statespace This problem is approached using a herarchcal neural net system m
controller will be supenor companng wth a system con- whch the jkst level momtors the operatmg regme and activates one of
multivanable voltage regulator many mdividuat neural nets, each focused on one, narrow, local
In order to generate the t r a m g set of patterns, the state- subproblem and trained to handle a specific control sequences [I71
space optmal regulator problem is solved ( u s q solutions of the dwrete State-space optimal controller is utilized as initial approximation
algebrax matnx bccati equation (28)) for " m t operating of the neural controller. State-space feedback controllers are based on the
conditions and dfferent types of dsturbances In the experiments SEX knowledge of the system model and depend on its linearlzation In real
hstory data pomts for state vanables and three hstory data pomts for world systems, neither system model nor its linearizatioii are dependable
control variables are used to formulate the model (i e in equations (29) However, wing &te optimal feedback as t r k g mtiation mechanism is
and (32) n=6, k=3) whle the t m e mterval chosen for dscretmtion was a possibllity whch we explore m tJns work It is not reahstic to expect that
At=O 05 s The traunng set for each SL modul consists of 150 t r m g neural controllers could be efficient m operating regmes m which the
pattems Each traunng pattem contams (as shown m Figure 4) 25 mput optnnal state feedback controller is not satisfactory However, the method,
based on computational archtecture of neural nets, whch describes the
features and 2 output features (control mputs u-(t)and u g o v ( t ) )A t r m g process m the situation when it is mpossible to describe real-
Functional Expansion FLN model of order-2 was used w t h all world n o d " d y n m c processes by exact mathematical model is
enhancements of the second level pruned, except for ~ ( t )and presented m [26] The problem of impromg performance, when a neural
controller is mplemented in a multnnachine power system environment
[w ( t ) - w(t - At)] Consequently, the enhanced representation (wth 49 requires a more suitable method for synthesimg traunng pattems (whch
mput features, no hdden nodes and 2 output nodes) has been used The are to be used dunng the t r a m g phase of the learmng process) as a
neural-net was allowed to train itself usmg the l e a m g rate q = O 25, benchmark The method of sequential design of decentralized multivanable
momentum m=O 9 and sigmoidal slope 0 10 The least square error was controllers m multmachme power systems by projective controls [27] is
reduced to 0 000007 m 1161 iterations wth the m a m u m pattem error one of possible cand~datesDue to the lmted space, the results of t h ~ s
6 9x10-' and the mm" pattern error 2 0x10" The t r m g wth the mveshgation are not presented in t h ~ paper
s
 765

100.000 -

----I.,
...~-. __ - - -___-.._
__ .=--.- -__ _______ --.

Trml
0.00 2.00 4.00 6.00 8.00 10.00 0
TIME ( s ) TIME ( s )

2 3.000 7
J

TIME ( s )

Figure 5 System responses (rotor speed, ternnnal voltage, electrical torque, exciter mput) to a temporary
(0 10 seconds) short-circult
----- ConvenQonalcontroller, - optnnal state-space controller, ------------- neural controller

1.ooo

0
$ 0.600
Y
0
W

Figure 6 System responses (turbme flow, gulde vane openmg, mechanlcal torque, and govemor mput) to a
temporary (0 10 seconds) short-circult
-----conventional controller, -optunal state-space controller, -------------neuralcontroller
766

The behavior of neural nets is strongly conditioned through the ACKNOWLEDGMENT

training process. Consequently, changes or modi6cations in neural-net
behavior are obtained through the continuation of the training process and M Djukano~lcgratefully acknowledges support provided by AI WARE
the presentation of new data obtained from new operating conditions. It is Inc, Cleveland, 0% USA, with the N-NET Neural-Net System
shown in Ref [2 11 that, when presented with the same situation a number Development s o h a r e which was used to calculate all the numencal
of times, the NN controller can improve its performance. Neural examples in this paper Authors also gratefully acknowledge explanations
controllers are a large dimensional nonlinear controllers, with all the and parameters of HPP "Djerdap" obtamed from Dr Dusan Amautovic,
capability of providing performance, and all the difsculty in tuning it, that Inshtute "kola Tesla"
t h ~ simplies. However, in contrast to controllers which rely on specific
functional relationship, the strength of this approach lies in the adaptability
of the network structure, a functionality is synthesd implicitly by the REFERENCES
weights, the thresholds and the activation functions. As examples change,
that is as the net learns through more and more observations, the mapping 1. GhandakIy, A ,Kroneger, P , "An Adapbve Time-Ophmal Controller for
&om input to output is also adaptively refined [14]. Generahag Umts Stabilizer Loops" IEEE Trans on Power Systems, Vol
2, NO 4, NOV1987, pp 1085-1090
2. Ghosh, A , Ledwich, G , Mal&, O P , Hope, G S , "Power System
It should be noted that a more robust control scheme may be
S t a b h r Based on Adaphve Control Techmques", IEEE Trans PAS-
synthesized using several conventionally measurable quautities, such as 103, No, 8, Ang 1984, pp 1983-1989
accelerating power, reactive power or terminal current. On other side, 3. Mao, E, h4a& O P , Hope, G S , Fan, J , "An Adaphve Generator
good multivariable optimal controller should be designed having as a Exatahon ControllerBased on Lmear Ophmal Control",EEE Trans on
simultaneous targets suppressing mache-to-system, shaft torsional or Energy Conversioq Vol 5, No 4, Dec 1990, pp 673-678
inter-area oscillations. The role of neural-net predictor and controller in 4. Harley, R G , Lahond, M A , Secker, A , "Ophmal and Mulhvanable
such circumstances is currently under our investigation. The focus in this Control of a Tu&ogenerator", Electnc Power Systems Research,
paper was to overcome some limitations of classical optimal control theory lO(1986) pp 35-46
approaches, regarding the requirement for measuring all state variables or. 5. Yu, Y N , "Electnc Power System Dynanncs", New York, Academc
introducing the observer for the reconstruction of full state feedback Press, 1983.
variables, and to make the step toward further investigations for their 6. Chapman, J W , Ilic, M D , Kmg, C A , Eng, L , Kaufman, H ,
" S t a b W g a Mulhmachme Power System ma Decentralized Feedback
practical applicability. In that respect the main advantages of the proposed
Lmeanzlng Excltabon Control",IEEETrans on PWRS, No 3, Aug 1993,
neural-net based controller are: pp 830-840
-NNC avoids data complexity and requires a small amount of 7. h u t o w c , D , Ivklqanov~c,R , "An Approach to the Analysis of Large
information like conventional controllers (angular speed, terminal voltage Perturbahom m Hydro-Electnc Plants wth Kaplan Turbines",
and acceleration of angular speed); ElectricPower SystemsResearch, 9(1985), pp 115-121
-NNC provides improved system transients with large stabdity 8. Skatanc, I), Calowc, M , Arnautomc, D , "Ophmal Control of Low Head
margins like state-space optimal controllers. Hydropower Plants", Proc&gs of XV IAHR Symposium, Belgrade,
11-14 September 1990, Paper D4
9. Fu, K S , "Learmng Control Systems-Renew and Outlook", IEEE
Tmmchons on Pattern Analysis and Machme Intelligence, vol '
PAMI-8, NO3, May 1986
10. Berger, K M , Loparo K A , "A Survey of Techmques to Handle
Uncertamty and Complemty wth Applicahon to Intelligent Control",
5. CONCLUSIONS CWRU, Techn~cal Report TR 87-139, October 1987
11. Ghandakly, A , Idowu, P , "Design of a Model Reference Adaphve
A new design of the flexible neural-net based coordinating S t a b h r for the Exciter and Governor Loops of Power Generators",
stabilizing control of the exciter and govemor loops of low head hydro IEEE Trans onpower Systems, Vol 5, No 3, August 1990, pp 887-893
power plants has been proposed in the paper. The control system design 12. RnmellmC D.E , Hmton, G E ,and Williams, R J , "Leanung internal
is based on the pattem recognition principle and the implementation representahon by error propagahon in parallel distnbuted
on the parallel distributed computational architecture of neural-nets. processmg",Explorahons 111 the Wcrostructures of Cogmhon, vol 1
Foundat~ons,pp 318-362, MIT Press, 1986
Self-organization and the predictive estimation capabilities of neural net 13. Pao, Y E,"Adaqtwe pattem recogmtion and neural networks", Addison
computing are utilized using the cluster-wise segmented associative Wesley Pubhshg Company Inc , 1989
memory system. The direct mode NNC is intended for real-time adaptation 14. YHPao, and D J sobajic, "Combined Use of Unsupemsed and
of control action using discrete measurements of speed, terminal voltage Supemsed Learrungfor Dynam~cSecunty Assessment", BEE Trans on
and speed acceleration. Power Systems, Vol PWRS-7, No 2, May 1992, pp 878-884
15. D J U ~ ~ ~ O MW, Sobap,
C, D J ,Pao, Y H , "Neural-Net Based Calculahon
Digital simulations of hydropower plant equipped with low head of Voltage Dips at W m u m Angular Smng 111 Direct Transient
Kaplan turbine and the comparisons of conventional excitation control, Stabd~tyAnalysis", Electncal Power and Energy Systems, vol 14, No 5,
state-space optimal control and neural-net based control show that NNC October 1992, pp 341-350
provides improved system transients with large stability margins like state- 16. Djukancvlc, M , Sobajlc, D J , and Pao, Y H , "Predxtion of Power
System Frequency Response after Gnerator Outages Using Neural-
space optimal controllers. Once trained, the neural-nets are able of Nets",IEE P " g s , Pan C , Vol 140, No 5, Sept 1993, pp 389-398
providng the control action in real-time based on on-line measured 17. Djukanovl~,M , SobaJic, D J , and Pao, Y H , "Neural-net Based
operating conditions. Deteiuunahon of Generator-Sheddmg Reqwrements in Electnc Power
Systems", IEE Procedngs-C, Vol 139, No 5, Sept 1992, pp 427-
We have shown that complex mapping relating the operating 436
conditions and measurements to control actions can be effectively and 18. H% Y Y ,Wu, C J , "Adaphve Control of a SynchronousMachne Using
accurately stored in the hetero-associative memory using a feedforward the Auto-Searclung Method",lEEE Trans on Power Systems, Vol3,
neural-net system. Neural-nets have proved to be a powerful tool for NO4, NOV1988, pp 1434-1440
building unconventional space transformations such as from parameters 19. ElmetwaUy, M M ,Rao, N D ,Mal& 0 P , "ExpenmentalResults on the
wbch characterize the HPP operating conditions and on-line Implementahon of an Ophmal Control for Synchronous Macbnes",
measurements to the space of control actions. The examples of possible IEEE Trans on Power Apparatus and Systems, Vol PAS-94, 1975, pp
1192-1198
procedures for the design of these controllers with incomplete feedback, 20. Harley, R G Discussion on [ E ]
employing essentially the same variables as in conventional control 21. sobajic, D J, Pao, Y H , Lee, D T , "Autonomous Adaptive Synchronous
schemes, give a good sense of their suitability for practical implementation Mactune Control" Electncal Power and Energy Systems, vol 14, No 2,
in low head hydro power plants. ApnYJune 1992, pp 166-174
 767
22. Pao, Y.H., Phillips, S., Sobajic, D.J., "Neural-Net Computing and the = 0.2447;q = 0.6208;q = 0.31 13;q = 1.2593;
Intelligent Control of Systems", in the SpeciaI Issue on Intelligent
-?; = 1.1421;z = 0.7756;q = -0.7439;G = 1.2706
Control of The International Journal of Control (IJC), Vol. 56, No. 2,
August 1992, pp. 263-290.
23. Hiyama, T., "RuleBased Stabilizer for Multi-Machine Power System", K,= 0.6851;K; =0.9962;K3 =0.5813;K4 = 0.6862;
IEEE Summer Meeting, 89 SM 691-7 PWRS. & = -0.0815;& = 0.6239
24. Hsu k .Y, Chen, C.R., "Tuning of Power System Stabilizers Using an By eliminating equations and variables corresponding to the conventional
Artificial Neural Network", IEEEPES 1991 Winter Meeting, New excitation and govemor controllers we obtain the 7th order linearized
York, Februasy 3-7, 1991.
model which was used for synthesizing optimal state space controller
25. Weerasooriya, S., El-Sharkawi, M.A., ""Identification and Control of a
DC Motor Using Back-Propagation Neural Networks", IEEE Trans. on described in Section 2.
Energy Conversion, Vol. 6, No.4, Dec. 1991,pp. 663469. A real low head hydropower plant "HPP DJERDAP"was selected
26 Sobajic, D.J., Pao, Y.H., Lee, D.T., "Robust Control of Nonlinear for the illustration of approach presented in the paper. The unit parameter
Systems Using Pattern Recognition", 1989 IEEE Intemational for the case study are specified in Table 1 [8].
Conference on Systems, Man and Cybernetics, Cambridge, MA, Nov.
14-17, 1989, pp. 1-7. SYNCHRONOUS GENERATOR, EXCITATION SYSTEM
27 Arnautovic, D., Medanic, J., "Sequential Design of Decentralized AND TURBINE GOVERNOR SYSTEM DATA
Multivariable Excitation Controllers in Multimachine Power Systems by
Projective Controls", IEEE Trans., EC-2, 1987, pp. 598404.
1
P=0.773 p.u Sn=190MVA D=O. 1 1
Tf,=0.090 s bF=0.061p.u 1
Q=0.015 p.u I <=15.75 kV I T,,=O.Ols 1 T,,=0.047 s I b,=0.873 p.u
~ ~~

R4.034 p.u xd=1.245p.u. T,=0.09 s Tf3=0.031s T,=15.4 s

X=0.700p.u ~ ~ 4 . 9 2 5 p . u T2=O.047s
. ku1=50p.u. c=2.23 s
APPENDIX 1: The linearized HPP state space model and system G-0.249 p.U id=0.373p . ~ . T,=0.031 S k"4.5 P.U. T=OS s
parameters: B4.262 P.U 7',,=6.5 s kfl=l P.U. kfl=Sp.~./Hz TI=0.024s
V,=l.05 p.u M=7 s T,4.03 s kf=3 p.u./Hz/s T,=1.4 s
The state space model for the HPP plant with conventional controllers can
be organized in the following form [SI:

= A x ( t ) + Bu(t) + Fw(t)
dt BIOGRAPHIES
Miodrag B. Djukanovic was bom in Visoko, Yugoslavia on June 4,
1959. He received his B.S., M.Sc. and Ph.D. degrees in Electrical
Engineering from the University of Belgrade in 1982, 1985 and 1992
respectively, specializing in electric power systems. In 1984 he joined
Electrical Engineering Institute "Nikola Tesla". At present he is also with
The elements of matrices A, B and F can be easily identified from the Electrical Engineering Department of Belgrade University as lecturer at
following linearized equations in whch the symbol A (that denotes postgraduate studies. His major interests are in the area of power system
deviation from the steady-state operating point) will be omitted: steady-state and dynamic security and applications of neural networks
and fuzzy set theory in electric power systems.
dm
_
dt
=-
M
((q -a
T
-D)w- K16- K2e;+'qT
I;
+(Ti % +
--)U
T Milovan Novicevic was bom in Yugoslavia in 1954. He graduated and
TT received his M.Sc. degree from the University of Belgrade. In 1981 he
+(T,-e-r)g,-T,) joined the Union of Yugoslav Electric Power Industries where he was
I; working in the Dispatching Centre. His present research interests include
dS -
_ 1 adaptive control, load frequency control and operations planning.
- Ox@; 3= 1 1
T, (kpE, - EpD)
(EFD- -e; - K4s) ; =
dt dt eo K, dt
Djordje Dobrijevic was bom in 1967 in b i n , Yugoslavia. He received
B.S degree in electrical engineering in 1992 from Belgrade University.
In 1993 he joined Institute "Nikola Tesla". His current research interests
include hydro power plants modelling, power system protection and
optimal power flow analysis.

Borivoje S. Babic was bom in Bileca, Hercegovina, in 1941. He

received B.S and M.S degree in electrical engineering in 1966 and 1980.
Today he works as a head of Study and Research Department, Electric
Power Board of Serbia, Belgrade. His current research interests include
real-time preventive control of electric power systems and optimal power
flow analysis.
Dejan J. Sobajic (M'80-SM189) received the BSEE and the MSEE
degrees from the University of Belgrade, Yugoslavia in 1972 and 1976,
respectively and the Ph.D. degree from Case Westem Reserve
University, Cleveland, Ohio in 1988. At present, he is with Power
System Control, EPRI. Dr. Sobajic is a member of the IEEE Task Force
on Neural Network Applications in Power Systems and of the IEEE
Intelligent Controls Committee. He is the Chairman of the Intemational
Neural Networks Society Special Interest Group on Power Engineering.
Yoh-Han Pao has been a Professor of Electrical Engineering and
Computer Science at Case Western Reserve University (CWRU) since
1967. He is the George S. Dively Distinguished Professor of Engineering
at CWRU, is a Fellow of IEEE and of the American Optical Society.