I480 Proceedines of the 4& World Conmess on lntellieent Control and Automation June 10--14. 2002, Shanahai. P.R.

China

The third-order induction motor

parameter estimation using an adaptive genetic algorithm

Xiaoyao Zhou Haozhong Cheng Ping Ju

Department of Electrical Engineering, Department of Electrical Engineering,
Shanghai Jiaotong University, 200030, Hohai Univenity, 210098,
Shanghai, China Nanjing, China

Abstract: This paper presents a new adaptive genetic algorithm finding the global optimal solution. Genetic algorithm has
for third-order induction motor model parameter estimation. been applied to many fields of power s y ~ t e m , [ ~ ~in~ .[I41
'~'
The crossover and mutation probability of adaptive genetic ordinary genetic algorithm(0GA) has been applied to some
algorithm change according to the fitness statistics of popula- nonlinear load model parameter estimation, in this paper an
tion at each generation. The proposed algorithm can enhance
the convergence performance of GA and prevent premature adaptive genetic algorithm(AGA) is proposed and applied to
problem. This algorithm is successfully applied to the third- induction-motor parameters estimation.
order induction motor model parameter estimation. In section I1 the third-order induction motor model is
introduced and hasic procedure of an optimization based
parameter estimation method is given. Section Ill provides
I. INTRODUCTION an adaptive genetic algorithm based approach for parameter
estimation of induction motor model. Section IV studies the
To represent the inherent nonlinearity of electric loads, results of motor parameter estimation by AGA and OCA.
several nonlinear dynamic models have been devel-
oped[1,2,3]. Among them the induction motor models have
been long studied and widely applied[l]. How to estimate 11. MATHEMATICAL FORMULATION OF
the model parameters is a important problem, and now there INDUCTION MOTOR PARAMETER ESTIMATION
are three kinds of approaches: optimization based ap-
proach[4,5], analytical approach[6] and stochastic ap- A. The induction motor model
proach[7]. The optimization based approach is to search the
The induction motor load is an important component of
best parameters, which minimizes an errors function between
the measured output variables and simulated ones. Tradi- power system dynamic loads and bas significant effect on
transient voltage stability. By neglecting the stator tran-
tional search algorithms have been applied in these kinds of
approaches. Generally speaking, this sort of approach is sients, a third-order induction model can be derived from the
based upon the assumption of a smooth parameter search detailed fifth-order model [Io1. The equations of the third-
order models are:
space with ever present derivatives. When initial values of
the parameters are far away from their real values, the esti-
mation procedures may converge to local optimal values or
even diverge. The analytical approaches derive parameters
determinately according to the test results. This kind of d6 X-x' V.sin6
approach can be used in a special test such as step test. -=w-ws--.--
dt X' To'.E'
Usually, it is sensitive to measurement error. The stochastic
approach may he limited to the assumptions of noise. dw = - V.E'.sin 6
-
Genetic algorithms is a search algorithm based on the
dt X' - Tm (3)
principle of evolution and is used to search large, non-linear
search spaces where traditional optimization techniques fall
short. The basic principles of GAS were first designed by
P = -(VE'/X'). sin 6 (4)
John Holland['I. The genetic algorithm operates on a popu-
Q = V (V - E ' . COS S ) / X ' (5)
lation of potential solutions (chromosome), each represent-
ing a possible solution to a given problem. Each cbromo- where
some is assigned a fitness value according to the result of the
fitness function. Highly fit chromosomes are given more E', 6 voltages magnitude and angle behind transient
o p p o a i t i e s to reproduce and the offspring share features reactance
taken from their parents. These feature contribute to a ge-
WO,W , angular velocity of stator and, rotor [rad/s]
netic algorithm's robustness and increase the likelihood of

0-7803-7268-9/01/$10.00 02001 IEEE.

Rocdings of the 4* World Congress on Intelligent Control and Automation June 1&14, 2002, Shanghai, P.R.China 1481

X, ,Xs,X, magnetizing, stator and rotor reactance Y, ---Measured (or desired) values of output
vector
X ' = X , t X,X,I(Xm+ X,)transient reactance Y, --Computed values of output vector
X=X,tX, For the third-induction motor model
Y= k', QP
Td=(X, + X,)/mnRrtransient open-circuit time con-
stant
Z=[M, T ' , C, X ' , T,,X T
4,R" stator and rotor resistances C. GA based parameter estimation
M motor inertia
It is obvious that better parameters usually result in less
load torque constant error function. In CA, large fitness would reproduce more
Tm off-springs. This will most likely contribute to better
P3Q
induction motor active and reactive power parameter estimation. Noting that the error function is al-
ways positive fitness is chosen as inverse of error function.
let T'= T,'.X'/X, C = (X- X ' ) / X the induction
J=l/E(10)
motor model can be rewritten as
So the procedure of searching minimum error function
dE' equivalent to searching maximum fitness function
TI- = -E'+CV.COS~
minimize
z=z'zEs
E(Z) Q maximizeJ(Z)
Z=Z. 2 s
(11)
cv
= @ - U , ---sin6
.
(6) The binary codes in CA should converted into decimal
TIE' parameters used in model and fitness computations. At first
VE' .
l M dz o = -F-sm6 - T, the decimal integer Z"" corresponding to a binary code
z"'with length L can be obtained
B. The major procedure of an optimization based
parmeter estimation method Where, a, --- a single binary feature or detector

The major procedure of an optimization based parameter The maximum values are
estimation method is to search the best (or optimal ) z; = 1 1 ...1, Z Z )=2 L -1
parameter vector Z ' , which minimizes an error function Then, the decimal integer can be converted into the deci-
E mal parameter by
E = minimize E(Z)
zcz 26s
(7) z, =zjmi.+(Zj,, -z,&)(zj / (10) z'lo'
mT)

Where s a set of admissible parameters Where parameter, Z, ..,J'* parameter

The enor function E is usually taken as a nonnegative
and monotonously increasing function of output error
1
,Zjm, - Search range of Zj
[Zjmin

E = jF(llY,(t)- Y,(/)l()dt (continuous time)@)

to
111. AN ADAPTIVE GENETIC ALGONTHM
N
E = CF(/IY,(k)-Y,(k)l/) (discrete time) (9) There are usually there operators in a typical genetic algo-
k=O rithm, the first is the production operator which makes one
Where, [to,T ] --_timeregion of observation or more copies of any individual that posses a high fitness
value; otherwise the individual is eliminated from the solu-
/I 1 ---norm tion pool the second operator is the crossover operator. This
F ( e ) Monotonously increasing function operator select two individual within the generation and a
crossover site and carries out a swapping operation of the
k --- k'' Time sample string bits to the right hand side of the crossover side on both
N ---Number of all samples individuals. Crossover operations synthesize hits of knowl-
1482 Pmceedings of the 6 World Congress on Intelligent Conml and Automation June 10-14, 2002, Shanghai, P.R.China

edge gained from both parents exhibiting better than average which the mutation is applied.
performance. Thus the probability of a better performing of When GA converges, the fitness differences be:
offspring is greatly enhanced. The third operator is the mu-
tation operator which introduces occasional change of a
random string position with a specified mutation probability.
The influence of crossover probability p , and mutation the normalized fitness is defined as:

probability p, on controlling CA performance has been

acknowledged and many methods of setting the optimal
values has been rep~rted~".'~~.
In this paper we improve the
AGA (adaptive genetic algorithm) first proposed by Srini-
Vas["]. it is been reponed[91 tho effectiveness to employ
c(f)to regulate the mutation operator.
I

The key idea of the AGA is changing p , ,p , according

to the fitness statistics of population at each generation. It is
has been observed that the difference between the maximum
fitness and the average fitness of the generation decrease
when the GA converges to the optimum solution. So the
p , ,p , should be changed according to the f,, - f .On
in this paper it is found that if the c(f3 is employed to
p , ,p , at same time, it will d a n c e the GA convergence
the other hand if are the same p , ,p , for all the chromo-
performance.
some of one generation, it means that solutions with high
fitness and solutions with low fitness have the same value to
do crossover and mutation, this will influence the effective-
ness of GAS. So the adaptive strategy for updating has been
proposed in [13].

IV. SIMULATION RESULTS

and The OCA and AGA-based parameter estimation algo-

rithms have been implemented to third-order induction
motor based on the flowchart shown in Figure 1. The origi-
nal motor parameters and search ranges are shown in table I.
The adopted parameters in t h o algorithms are given in Ta-
Where k,,k,,k,,k, E (0,l) toensure p , , p , E [0,1] ble2:
p , is the large of fitness values of the individual selected
for crossover and p , is the fitness of the chromosome tu

M T, T' C r: ' X
Real value 3.2 1.o 0.00685 0.9685 0.1344 3.72
Search range(max) 1 0.2 0.01 0.5 0.05 1
Search range (min) 8 1.8 2 1 0.5 5

No. Of Length Popula- No. of

variables ofchrom tion size iteration pm k1 k2 k3 k4

AGA 5 48 50 100 0.95 0.05

OGA 5 48 50 100 0.86 0.5 1.0 0.05

Rofeedings of the 4' World Congress on Intelligent Control and Automation lune 10-14. 2002, Shanghai, P.R.China 1483

1 SCleSt s o l u t i o n s for

tnka the, lr,sf solution recorded

as the f i n a l solution
I

Fig 1. The flow chart of AGA

M Tm T' C X' X Error

Best 3.278 0.9894 0.0068 0.9631 0.1365 3.701 0.0185
worst 3.335 0.9768 0.0076 0.9602 0.1413 3.552 0.2576
Average 3.30 0.9812 0.0073 0.962 0.1401 3.673 0.1583
1484 Proceedings of the 4* World Congress on Intelligent Control and Automation June 10-14, 2002. Shanghai, P.R.China

Method Shortest execution time (sj Longest execution time (s) Average execution time (sj
AGA 15.12 16.74 15.53

I CGA I 27.46 I 35.65 I 30.49 I

tion of nonlinear dynamic loads in power systems”
V. CONCLUSION: IEEE trans, vol PWRS-9 Feb 1994, pp 479-493
This paper presents a new adaptive genetic algorithm for
third-order induction motor model parameter estimation. The [7] J.H.Holland “Outline for a logical theory of adapive
crossover and mutation probability of adaptive genetic algo- system”, Journal ofACM, Vo1.3, 1962, pp 297-314
rithm change according to the fitness statistics and normal- [8] K.Iba “Reactive power planning in large power system
ized fitness distances between the solutions of population at using genetic algorithm” Trans on IEE Japan, Vol 113-
each generation. This algorithm is successfully applied to the B 1993 pp 865-872
third-order induction motor model parameter estimation. The
performance shows that the AGA is able to enhance the [9] Q H Wu, Y J Cao and J Y Wen “Optimal reactive
convergence performance of GA and prevent premature power dispatch using an adaptive genetic algorithm”
problem. Electrical Power & Energy Systems, Vol. 20, No. 8, pp.
563-569,1998
[IO] A.H. Mantawy, Youssef L. Abdel-Magid , Shokri Z.
VI. REFERENCE Selim “A new genetic-based tabu search algorithm for
[I] H.C. Stanley, “An analysis of the induction motor”, unit commitment problem” Electric Power Systems
AIEE Trans., Vol57,1938, pp 751-755 Research49(1999)71-78

[2] W. Xu, Y. Mansour, “Voltage stability analysis using [ 1I] Grefenstene, J. I., “Optimization of control parameters
generic dynamic load models” , IEEE Transactions on for genetic algorithms”. IEEE Transactions on System,
Power Systems, Vo1.9, No.1, Feb 1998 pp 479-493 Man and Cybernetics, 1986,16(1), 122-128

[3] D. Karlsson, D.J. Hill, “Modeling and identification of [I21 Schaffer, D. and Morishma, A., “An adaptive cross-
nonlinear dynamic loads in power systems”, IEEE over mechanism for genetic algorithms”. Proceed-
Transactions on Power Systems, Vo1.9, No.1, Feb 1994 ings of the 2nd International Conference on Genetic
pp 157-166 Algorithms, Cambridge, MA, July 1987, pp. 3 6 4 0

[4] P.Shackshaft, 0.C.Symons “ General purpose model of [I31 Srinivas, M. and Patnaik, L. M., “Adaptive probabiliti-
power system loads” Proc.IEE Vo1.124, 1997, pp.715- es of crossover and mutation in genetic algorithms”.
771
I __ IEEE Transactions on System, Man and Cybernetics,
1994,24(4), pp656667
[5] JA de Kock, FS van der Menve “ Induction motor
parameter estimation though an output errOr tech. [I41 P.Ju , E.Handscbin , D.Karlsson .. “Nonlinear dynamic
nique” IEEE Trans, Vol EC-9, March 1994, pp 69-75 load modelling: model and parameter estimation”
IEEE Transactions on Power system , Vol.1 I No 4 No-
[6] D.Karlesson, and D.J.Hill ‘‘ Modelling and identifica- vember 1996 pp1689-1697