You are on page 1of 6

2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.

Paper Identification No. 79

Induction motor Parameter Estimation Using Hybrid


Genetic Algorithm
K.Sundareswaran, H.N.Shyam, S.Palani Joby James
Department of Electrical and Electronics Engineering Viswajyothi College of Engineering and Technology
National Institute of Technology Kerala, India
Tiruchirappalli, India

Abstract—The main objective of this work is to develop a cost determination of circuit parameters of induction motors which
effective off-line method for determination of induction motor are running continuously in industries is not possible using
equivalent circuit parameters by conducting a single load test on traditional methods.
the motor. The proposed scheme is an alternative viable method
to conventional means of no-load and blocked rotor tests. The The problem of induction motor parameter determination
identification of motor parameters is redrafted as a multi- has been addressed extensively by many researchers in the past
objective optimization problem and solution is sought through [1-8]. Deep bar machines were considered in [1] and motor
conventional optimization method as well as Genetic Algorithm parameters were identified from standstill tests in [2].
(GA). The conventional method employed is the well known Parameter estimation was carried out with emphasis on leakage
Rosenbrock’s (RB) rotating coordinates method. When the reactances in [3]. Rotor time constant calculation for vector
results of the two methods are analyzed, it is observed that while controlled drives is available in [4]. The method of finite
GA offers near optimal solution to the problem, the method of element analysis was employed in [5], while an iterative
RB always results in global optima, provided initial values are method was seen in [6]. A standstill frequency test was utilized
chosen judiciously. Hence, it is proposed to combine these two in [7] for motor parameter determination. This method requires
methods to gain the advantages of both the methods. In such a sophisticated equipments such as frequency spectrum analyzer.
hybrid optimization method, the task of global search is carried Another approach used a single test on the motor with an
out by GA, while Rosenbrock’s method is devoted to local search. inverter for parameter estimation [8]. Even though these
Comparison of these two techniques are discussed and presented
methods are very rigorous in principle and procedure, the
in conjunction with computed and practical results. It is shown
that combination of GA with conventional method yields
results are highly accurate, which is a basic denominator for
improved results. vector controlled applications.
However, such a high degree of accuracy is not demanded
Keywords-Induction motors, Equivalent circuit parameters, in many other applications such as relaying and power system
Genetic Algorithms protection [9, 10] and also in analysis, performance evaluation
and simulation studies. Further in many of the references
I. INTRODUCTION mentioned above, a pertinent drawback is that iron loss
Induction motors have been extensively used in fixed and component is neglected in the computation. Even though, this
variable speed applications for many years owing to their size, component may not have an impact in many drive control
cost, weight, reliability, ease of maintenance and efficiency. characteristics, this parameter plays a vital role in the
Recent advanced control techniques such as vector control determination of motor efficiency and power factor, in
make them to compete with DC motors in many aspects. The particular, at light load conditions. In this paper, a novel off-
information regarding motor circuit parameters is very much line method suitable for determination of steady state
essential for design, manufacture, performance evaluation and equivalent circuit parameters of three-phase induction motors is
vector control applications. reported. The proposed scheme possesses a few distinguished
features. The prominent one is that it requires only a single load
Generally, the equivalent circuit parameters are predicted at test on the motor at the existing supply voltage and frequency.
the design stage by the manufacturer and subsequently This makes it possible to estimate parameters of induction
confirmed by test. However, designer’s data may not be motors running continuously in a plant. Another distinct aspect
available in many cases especially for an old motor in a plant. of the new dispensation is that the estimation process is
The classical technique for the determination of induction redrafted as an optimization task and conventional as well as
motor parameters is based on no load and blocked rotor tests. evolutionary methods such as genetic algorithm (GA) are
The main disadvantage of blocked rotor test is that the motor employed for the solution. The conventional optimization
has to be locked mechanically and performed by skilled method employed is the well known Rosenbrock’s rotating
operators. Further, this test requires a lower voltage than the coordinates method [11, 12]. It is found that this method
existing supply voltage. With motor permanently coupled with performs well, when initiated with suitable starting values. In
load, it is not feasible to conduct no load test also. Thus, fact, the method starts diverging, with random initial guess.

978-1-4244-2806-9/08/$25.00© 2008 IEEE 1


2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.
Paper Identification No. 79

Secondly GA is used for this optimization problem. It is III. EVALUATION OF MOTOR PARAMETERS USING
observed that GA has several advantages, when compared to CONVENTIONAL OPTIMIZATION METHOD
conventional methods, such as reduced computational burden,
The conventional optimization method selected in this work
faster convergence, convergence to near optimal solution in
is Rosenbrock’s rotating coordinates method. This method
most cases with random initial values etc. However, GA has
comes under unconstrained optimization techniques. For
the main disadvantage of converging to near-optimal values
details, one may refer to [11]. Dedicated software is developed
rather than true optimal values. Hence, it is proposed to
in C programming language for this optimization problem. The
combine the two methods such that GA is used first with
accelerating factor α and the decelerating factor β employed in
random initial values; after certain number of iterations,
Rosenbrock’s optimization scheme are fixed at 3.0 and 0.5
Rosenbrock’s method is applied using the solution obtained
respectively and the program is executed for 300 iterations.
with GA as starting values. This hybrid approach allows GA to
Load tests are conducted on a 5 hp and 2 hp induction motors
search for near global optima, while conventional method
and the data obtained from the test is used for the optimization.
carries out local search, leading to true optimal values.
The initial values are selected based on certain standard
Computed results for two induction motors are presented with
assumptions [13]. The no load power factor for the test motors
experimental results to validate the new dispensation.
is assumed at 0.25 and no load current is taken as 30% of rated
current. These two assumptions help in determining initial
II. PROBLEM FORMULATION values of Rm and Xm. Starting value of R2 is taken nearer to
The well known per phase equivalent circuit of a three- known value of stator resistance; first value of leakage
phase induction motor is depicted in Figure 1. There are six reactance is computed assuming full load power factor around
unknown parameters in the circuit. However, stator resistance 0.80. With these initial values, optimization is carried out and
of the motor can be measured using a DC test and as usual, the the results are tabulated in table I. In order to verify the results,
ratio X1/X2 is assumed unity, leading to only four unknown no load and blocked rotor tests are conducted on the test motors
variables in the motor model. and the parameters thus obtained are shown within the
brackets. It is seen from table I that the motor parameters
These variables are X1, Xm, Rm and R2. In the proposed estimated by Rosenbrock’s method agree to a larger extent to
method, a single load test is conducted on the motor at the the one obtained with conventional method with marginal
existing supply voltage and frequency. Input current, I1, motor difference. It may be noted that the values obtained from
terminal voltage V1, and input power factor angle, together conventional tests need not be taken as bench mark values,
with motor speed are measured. Thus, the objective of the since, motor parameter estimation in these tests are based on
method is to obtain four unknown equivalent circuit certain assumptions. Another important point observed is that a
components which are compatible with the measured current judicial choice of initial values, as done in table I, is a must for
and power factor. In order to proceed with the optimization convergence of this conventional optimization problem. For
process, the problem is formulated as follows: example, when starting values are assumed randomly, the
results are converging, as shown in table II. The result is that
Find ( φ ) such that
even after 1000 iterations or so, the method does not converge.
Minimize F (φ ) = e1 + e2 Thus, for this method to yield satisfactory results, a prior
knowledge of variables is a pre-requisite. However, it may be
Subject to φ min ≤ φ ≤ φ max.
noted that this is not a major drawback, since, in the present
Where problem, suitable initial values can always be obtained, as
F (φ ) is the objective function, mentioned earlier.
φ is a set of machine parameters and is {X1, Xm, Rm, R2},
e1 = I 1 − I 1∗ , I 1 is the measured stator current and I1∗ is
estimated current,
TABLE- I
e 2 = θ1 − θ 1∗ , θ1 is the measured power factor angle and θ1∗ RESULTS OF ROSENBROCKS METHOD WITH JUDICIAL CHOICE OF
INITIAL VALUES
is the estimated power factor angle, and the subscripts, min
and max represent the lower and upper bounds of the variables Motor 1 (5 hp) Motor 2(2hp)
and are obtained from the prior knowledge. Parameters Initial
Final values
Initial
Final values
values values

X1(=X2) 7.599 5.462(4.947) 25 12.566(12.9716)

R2 1.321 2.386 (2.6782) 20 7.473(7.807)

Rm 407.187 408.108(320.666) 1140.246 1029.296(1223.82)

Xm 105.135 118.458(79.268) 294.183 489.703(356.86)

Figure 1. INDUCTION MOTOR EQUIVALENT CIRCUIT

978-1-4244-2806-9/08/$25.00© 2008 IEEE 2


2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.
Paper Identification No. 79

IV. APPLICATION OF GENETIC ALGORITHM TO MOTOR implemented for this selection process. Each chromosome is
PARAMETER IDENTIFICATION assigned a sector in this virtual wheel and the area of the
The GA is a search mechanism based on the principle of sector is proportional to their fitness value. In Roulette wheel
natural selection and population genetics. The primary concept selection, an angle is generated randomly and the chromosome
of GA was first proposed by Dr. Holland in 1975 [14, 15]. GA corresponding to this angle is selected. The chromosomes thus
constitutes biologically inspired multi-parameter selected are called parent population and are subjected to
search/optimization algorithms that have proven to be undergo crossover and mutation to produce offspring for the
effective in solving a variety of complex problems where other next generation. Conventional method adopted in GA is
algorithms have either failed or faced difficulties. There are Roulette wheel selection and in this work, this selection
many advantages that make GA attractive [14]. GA does not method is modified by combining it with Elitism. Using
require the use of derivatives. They offer a parallel searching Elitism, a definite number of best solutions are retained and
of the solution space rather than the point-by-point searching are re-used in the next generation without undergoing the steps
in a small region. Since the location of the optimal solution is of selection, crossover and mutation.
unknown before the search, it is very likely that the point-by-
point search needs to search through the entire solution space Following the selection of parent population, offspring
in order to find the optimal solution, not mention the traps of population is generated by performing cross over and mutation.
In crossover, randomly selected sub sections of two individual
the local optima. Hence, GA can find near-optimal solution for
chromosomes are swapped to produce the offspring. In this
a complex problem very quickly. GA is started by possible work, multipoint crossover is adopted for increased efficiency
sets of solutions to a problem; each set is termed as a since four variables are embedded in one chromosome.
chromosome. A group of chromosomes is called population. Mutation is another genetic operation by which a bit within a
Each chromosome consists of combinations of individual chromosome may toggle to the opposite binary. The crossover
solutions, called genes. Genes are normally represented in and mutation are performed based on the probability of
binary form. In this work, each unknown motor parameter is crossover and mutation.
coded as 10 bit unsigned binary number and as such each
chromosome has a bit length of 40. The maximum value of Dedicated software in C programming language is
each coded parameter is taken as double the expected developed for the implementation of GA to the present
problem. No attempt has been made to identify optimal GA
parameter value. Chromosomes in a generation are forced to
parameters. The parameters are selected via a trial and error
evolve toward better once in the next generation by three basic
process to achieve best solution set [16]. The parameters used
GA operators, namely reproduction, crossover and mutation. in the implementation of GA are listed below:
Reproduction, also termed as selection, is based on the Population size 30
suitability or fitness of each chromosome to the problem. For Coding Binary
this, a fitness function pertaining to the problem is to be
defined. Only a certain number of chromosomes screened and Number of generations 300
selected by the fitness function can survive and pass their Selection scheme Combination of Roulette
genes to the next generation. Since, GA can be used only for wheel selection with
maximization, the following fitness function F is used in the Elitism
present method:
Crossover operator Multipoint crossover
1
F = (1) Crossover probability 0.7
1 + F (φ )
The Roulette wheel selection is the most common and easy-to- Mutation probability 0.01
implement selection mechanism. A virtual wheel is Termination criterion 300 iterations

TABLE- II TABLE III


RESULTS OF ROSENBROCKS METHOD WITH RESULTS OF GENETIC ALGORITHM
RANDOM INITIAL VALUES
Parameters Motor 1 (5 hp) Motor 2(2 hp)
Parameters Motor 1 (5hp) Motor 2 (2hp)
Initial values Final values Initial values Final values
X1(=X2) 3.90039 (4.94) 7.524 (12.9176)
X1=X2 5 3.11(4.947) 10 14.063(12.9716)

R2 5 4.178(2.6782) 10 12.332(7.807) R2 2.6686 (2.678) 7.3313 (7.807)

Rm 100 98.608(320.666) 250 270.464(1223.82)


Rm 280.0586 (320.666) 1372.277 (1223.829)

Xm 50 49.219 (79.268) 100 120.030(356.86) Xm 98.826 (79.268) 398.816 (356.86)

978-1-4244-2806-9/08/$25.00© 2008 IEEE 3


2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.
Paper Identification No. 79

The load test data obtained from the 5 hp and 2 hp motors load and blocked rotor tests. These graphs give better
are used for the optimization. The results are tabulated in table understanding of parameter identification process. As can be
III. The values in brackets indicate parameters obtained seen, the variation of parameter value is little with traditional
through no load and blocked rotor tests. It can be seen that optimization scheme; with GA, it is more dynamic and
there is a good agreement between parameters predicted by GA oriented towards global optimum values. In order to gain both
and conventional tests. Further, comparing results of table I and the advantages of faster initial convergence of GA and steady
table III, it is evident that the method of GA results in better rate of convergence of conventional method, it is now
estimation of motor parameters than conventional optimization proposed to combine the two schemes. In such a hybrid
method. One of the chief advantages of GA is that there is no scheme, GA is first used with random starting values in order
necessity of judicial choice of initial values to the problem;
to reach a stable low objective function value. The optimal
with random initial guess, the program converges satisfactorily.
This can be attributed to the fact that GA use multiple starting chromosome, containing the information regarding the optimal
points (different chromosomes in the initial population) as machine parameters is then used as the initial starting point for
opposed to a single starting point in conventional optimization Rosenbrock’s algorithm, which then takes over.
method while searching the solution space. Since, initial values
are scattered randomly in the whole of the solution space, GA The change over from GA to Rosenbrock’s method can be
investigates simultaneously much larger portion of the solution initiated as soon as a low and stable objective function value is
space. This helps GA from falling in to the trap of local optima. achieved with pure GA.
Thus, conventional techniques can get trapped in one of the
local optima. The solution is then to try different starting points In order to illustrate performance comparison of the three
and run the same program several times. schemes, the transition is demonstrated at the 200-th iterative
step in Fig. 2(a). The fusion of GA with conventional method
V. DEVELOPMENT OF COMBINED GENTIC/ROSENBROK’S enables the optimization process to speed up considerably, as
ALGORITHM AND RESULTS depicted in Fig. 2(a) and attains a very low objective function
value. The novel optimization algorithm prevents
After analyzing the results obtained with the two methods, it is
Rosenbrock’s method to end up in local optima and also
interesting to compare them. The comparison is performed
greatly reduces optimization time of pure GA schemes. Table
based on certain characteristic curves pertaining to the
IV shows motor parameters identified through the combined
problem. The graphs in Fig. 2(a) compare convergence
algorithm. Comparing tables I, III and IV, it is obvious that
characteristics of the two methods. This graph shows variation
better parameter estimation is performed with the combined
of objective function of conventional method and GA. It is
algorithm.
seen that GA converges faster, resulting in a reduced objective
value of 1.4 in the 10-th iterative step and as low as 0.2011in
For further verification and illustration, tests were carried out
the 200-th iterative step. It is obvious that GA achieves larger
on a 2 hp motor and certain performance characteristics of the
changes at the initial part of the convergence process, whereas
motor were measured. These performance indices are now
it tends to slow down at the final part. This is mainly due to
calculated using the machine parameters obtained through
crowding factors of GA, which have a tendency to attempt
combined GA/Rosenbrock’s algorithm. The results are shown
newer solutions in the surroundings of the best solution
in Fig. 3. It is obvious that there is a good agreement between
attained till then [17]. On the other hand, the convergence in
measured and computed results, validating the veracity of the
conventional method is slow, with arbitrary starting points,
new method.
which were taken from the initial population of GA. A
significant observation is that even though conventional
method converges slowly, the process is at a steady pace.

The variation of motor parameters for 2 hp motor with each


iterative step using the two methods is depicted in Fig. 2. The
point “c” represents the value of the parameter obtained via no
TABLE IV
RESULTS OF THE COMBINED METHOD
Parameters Motor 1(5 hp) Motor 2(2 hp)

X1(=X2) 3.8665 (4.94) 9.622 (12.9176)

R2 2.6687 (2.678) 7.597 (7.807)

Rm 294.553 (320.666) 1412.75 (1223.829)

Xm 80.7908 (79.268) 355.193 (356.86)

978-1-4244-2806-9/08/$25.00© 2008 IEEE 4


2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.
Paper Identification No. 79

Figure 3.COMPUTED RESULTS ON A 2HP MOTOR

VI. CONCLUSION
This paper has presented a new dispensation for determination
of induction motor equivalent circuit parameters. A single
load test data alone is sufficient for parameter identification
and this enables the application of the method for an old motor
in a plant or a motor which is continuously running, where
conventional methods of parameter determination cannot be
easily applied. The parameter identification is redrafted as an
optimization task and conventional and evolutionary methods
are used for the solution. It is observed that both the methods
possess several merits as well as demerits and it is natural to
see that the two schemes as potentially complementary
methodologies. Hence, the two methods are combined to gain
advantages of both the methods such that genetic algorithm is
Figure 2.PARAMETER IDENTIFICATION THROUGH ITERATIVE used for initial search and conventional method for the final
STEPS stage of optimization. This hybrid optimization algorithm
a)OBJECTIVE FUNCTION b) X1(=X2) c) R2d) Rm e) Xm prevents conventional method to be “trapped” in local optima
and also significantly reduces optimization time of pure GA
schemes, besides settling with global optima. Computed and
experimental results are supplied to validate the proposed
method.

978-1-4244-2806-9/08/$25.00© 2008 IEEE 5


2008 IEEE Region 10 Colloquium and the Third ICIIS, Kharagpur, INDIA December 8-10.
Paper Identification No. 79

REFERENCES
[1] Z.Zhang, G. E Dawson, and T. R. Eastham, “Evaluation of dynamic
parameters and performance of deep-bar induction machines,” in
Proc.IEEE-IAS Annu Meeting, 1993, pp. 62-66.
[2] S. Moon and A. Keyhani, “Estimation of induction machine parameters
from standstill time-domain data,” IEEE Trans. Iind. Applcat., vol. 30,
pp. 1609-1615, Nov./Dec. 1994.
[3] T. A. Lipo and A. Consoli, “Modeling and simulation of induction
motors with saturable leakage reactances,” IEEE Trans. Ind. Applicat.,
Vol. IA-21, pp. 180-189, Jan/Feb. 1984.
[4] L. Zai, C. L. de Marco, and T. A. Lipo, “ An extended Kalman filter
approach to rotor time constant measurement in PWM induction motor
drives,” IEEE Trans. Ind. Applicat., Vol. 28, pp. 96-104, Jan/Feb. 1992.
[5] S. Williamson and M. J. Robinson, “Calculation of cage induction
motor equivalent circuit parameters using finite elements,” IEE
Proceedings-B, vol. 138, No.5, pp. 264-270, Sept 1991.
[6] B. K. Johnson and J. R. Willis, “Tailoring induction motor analytical
models to fit known motor performance characteristics and satisfy
particular study needs,” IEEE Tans. On Power Systems, Vol. 6, pp. 959-
965, Aug. 1991.
[7] M. Bertoluzzo, G. S. Buja, and R. Menis, “Inverter voltage drop-free
recursive least-squares parameter identification of a PWM inverter
induction motor at standstill,” in Proc. IEEE. ISIE, 97, vol.2, pp.649-
654, 1997.
[8] A. Gastli, “Identification of induction motor equivalent circuit
parameters using the single-phase test,” IEEE Trans. Energy
Conversion, vol. 14, pp. 51-56, mar. 1997.
[9] T. A Higgins, W. L. Snider, P. L. Young, and H. J. Holley, “Report on
bus transfer: Part 1-Assessment and application,” IEEE Trans. Energy
Conversion,, vol. 5, pp. 462-469, Sept1990.
[10] S.S. Mulukutla and E.M. Gulachenski, “ A critical survey of
consideration in maintaining process continuity during voltage dips
while protecting motors with with reclosing and bus- transfer practices,”
IEEE Trans. Power Syst., vol. 7 pp. 266-272, Aug. 1992.
[11] G.S.G. Beveridge and R.S. Schechter, Optimization: Theory and
Practice, McGraw Hill Inc. 1970.
[12] S.S.Rao, Optimization theory and applications, Mohinder Singh Sejwal
publications, Wiley Eastern Limited, 1987.
[13] M.G.Say, “Alternating current machines”, 5th edition, ELBS, 1992.
[14] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and
Machine Learning. Reading, MA: Addison-Wesley, 1989.
[15] Z. Michalewicz, Genetic algorithms + Data Structures = Evolution
Programs. Berlin, Germany: Springer-Verlag, 1992.
[16] J.J. Grefenstette, “Optimization of control parameters for genetic
algorithms,” IEEE Trans. Syst., Man, Cybern., vol. SMC-16, pp. 122-
128, Jan./Feb. 1986.
[17] T.Furuhashi, Advances in Fuzzy logic, Neural Networks, and Genetic
Algorithms. Berlin, Germany: Springer-Verlag, 1995.

978-1-4244-2806-9/08/$25.00© 2008 IEEE 6

You might also like