You are on page 1of 6

3882

IEEE TRANSACTIONS ON MAGNETICS, VOL. 34, NO. 6, NOVEMBER 1998

An Evolutionary Algorithm for the Optimal Design of Induction Motors


Jan Pawel Wieczorek, Ozdemir G ol, and Zbigniew Michalewicz

AbstractThis paper describes the application of an evolutionary algorithm to the design of induction motors. It is shown that the use of an evolutionary algorithm offers advantages over other approaches. These include a high rate of global convergence and the ability to handle discrete variables. Index Terms Evolutionary algorithms, induction motor design, optimization.

I. INTRODUCTION

NDUCTION machines are used in large quantities in applications varying from household appliances to space technology; hence their design assumes great importance. Many conicting criteria have to be reconciled during design for an acceptable outcome according to a given specication. Device designers generally attempt to achieve this by set design procedures, usually with the aid of lumped parameter circuit models. Commonly, a trial and error approach is adopted in arriving at a solution which is deemed to best satisfy such conicting requirements. However, solutions thus found are unlikely to be the best possible; especially in view of commercial pressures which severely curtail the time available for search. Lumped parameter circuit models are popular, and intrinsically suitable, in the a priori evaluation during the design stage of the external behavior of induction machines. If used in conjunction with computer-based optimization techniques, they can vastly improve the design outcomes in that they facilitate the nding of optimal solutions with a modest investment of computational effort. One traditional approach to design optimization of electromagnetic devices has been to use nonlinear programming (NLP) techniques, which allow for fast convergence and are well established in engineering applications. However, discrete design parameters, such as the number of conductors and the wire gauge, cannot be drawn into optimization; the designer needs to perform a number of sub searches, each with a xed set of discrete design parameters. The search outcomes are then compared and judgement is made as to the best possible solution. Evidently, it is desirable for this entire process to
Manuscript received September 30, 1997; revised August 10, 1998. G J. P. Wieczorek and O. ol are with the Department of Electrical Engineering, University of South Australia, Mawson Lakes SA 5095, Australia (e-mail: o.gol@unisa.edu.au). Z. Michalewicz is with the Department of Computer Science, University of North Carolina, Charlotte, NC 28223 USA (e-mail: zbyszek@uncc.edu). Publisher Item Identier S 0018-9464(98)08252-1.

take place automatically without intermittent input from the designer. One way of doing this is to treat discrete variables as oating values within the optimization subroutine. Once the search terminates, these sub-optimal oats are rounded off to the nearest integer (e.g., number of turns) or to the nearest value determined by availability (e.g., conductor size) before they are used within the design algorithm. The globality of the search outcome needs to be checked by initiating the search from various starting points. This is tantamount to approximating an actual mixed integer programming (MIP) problem with an NLP one, which occasionally may result in pseudo-optimal solutions. Furthermore, the method still necessitates that the designer exercise judgement in deciding whether the search outcome represents a valid optimum. A better way in dealing with the MIP nature of the electromagnetic device design would be to use a globally convergent optimization solver, without having to resort to an NLP approximation [1]. Discreteness of variables precludes the use of optimization methods which rely on gradient evaluation; optimization methods capable of effectively dealing with this class of problems are still being developed. However, recent developments in evolutionary optimization promise to lead to long-sought-for global optimization tools which can handle MIP problems [2]. In recent years, evolutionary algorithms (EAs) have been recognized as potent tools in optimization. They exhibit a high rate of global convergence across the broad spectrum of optimization problems [3]. Since the search progress is based on function evaluations, no gradient evaluation is required. These features make EAs eminently suitable for use in the design optimization of electromagnetic devices; a dedicated algorithm can be devised to provide a powerful solver. In achieving this, the EA needs to be enmeshed with a model which adequately represents the device for design purposes. The model complexity must be counter-balanced with execution time, since the number of function evaluations may be high depending on the number of populations and the size of each population. This paper is organized as follows. The next section introduces induction machine design as an optimization problem. Section III presents the proposed algorithm together with experimental results. Section IV concludes the paper. II. INDUCTION MOTOR DESIGN: AN OPTIMIZATION PROBLEM The suitability of EAs in induction motor design optimization will now be demonstrated. In doing so, the so-called

00189464/98$10.00 1998 IEEE

WIECZOREK et al.: AN EVOLUTIONARY ALGORITHM

3883

TABLE I DESIGN PARAMETERS WITH THEIR DOMAINS

Fig. 1. Exact equivalent circuit model of induction machine.

exact equivalent circuit model of the machine, depicted in Fig. 1, will be used. This model is basically a per phase representation of a balanced polyphase machine in the frequency domain, comprising six elements, or model parameters. It elegantly represents the electromagnetic interactions taking place in an induction motor, albeit with substantial simplications and idealizations. Overall, the model is popular and well understood among engineers and, despite its shortcomings, offers a reasonably good prediction accuracy with modest computational effort. The six model parameters shown in Fig. 1 are predominantly dependent upon machine topology, although the materials used also have a signicant effect, especially if the machine is driven into saturation. For instance, the stator (one of the six model parameters) is leakage reactance obtained as a complex function of a number of quantities (design parameters) in the form (1) represent, respectively, the where number of turns, core length, stator bore diameter, stator slot height, stator slot width, etc. [4][6]. For simplicity, not all are shown in (1); an example is quantities which affect that of skewing of slots. In order to optimize performance, these design parameters need to be selected in such a way for the best possible overall as to produce a value for performance. Similar considerations apply to the remaining model parameters which too are complex functions of many factors related to machine topology and material properties. A more detailed description of how the model parameters were obtained in this case is deemed to be outside the scope of this discussion due to both the complexity of the design algorithm (which resulted in some 60 pages of MATLAB code) and intellectual property considerations. The high level programming language MATLAB has been adopted as the programming platform on account of its numerous advantages including rapid code development, ease of debugging, and excellent graphical interface. In the case presented here, 11 design parameters as specied in Table I have been drawn into the optimization process as variables with a view to obtaining values for the six model parameters (as per Fig. 1) for optimal performance. The selection of parameters was based on authors experience in induction machine design. These 11 parameters are known to have the most signicant effect on the performance of an

induction motor. The inclusion in the optimization process of further design parameters results in a slight improvement in the returned optimal value of the objective function. However, the cost in terms of computational effort is disproportionately high. Table I also shows the practicable domains for the design parameters. The rst two of the above, i.e., the number of conductors per stator slot and the stator wire gauge, are discrete variables; the former is an integer, whereas the latter assumes a discrete value from a list of available wire gauges. All of the above parameters are subject to a set of design constraints. Such constraints include threshold values for power factor, efciency, starting current, torque, slip, cost of materials, and machine size and weight. III. AN EVOLUTIONARY ALGORITHM An evolutionary algorithm has been developed to optimize the design of an induction motor. The algorithm uses binary representation; the individuals are formed by appending strings, delineating the eleven design parameters drawn into the process. The algorithm is based on a roulette wheel selection, single-point crossover, bit mutation, and elitism. The population size, bit string length and the number of generations can be selected depending on the requirements and available computer resources. Each parameter can be varied within its domain, the boundaries of which are user-specied. To enable the EA to handle the constraints, a nonstationary penalty function was used [7]. In this case, an evaluation devised in the form function if if (2)

3884

IEEE TRANSACTIONS ON MAGNETICS, VOL. 34, NO. 6, NOVEMBER 1998

is used, where is an objective function (to be maximized) is the penalty function which is dened by and gen with if constraints not violated otherwise (4) (3)

In the above equations, x represents the vector of design is a constant, chosen in a parameters as per Table I. way that amplies the constraint violations so as to make them signicant enough for the EA to conne the search is related to the to the feasible solution space; value of value of objective function. gen is the current generation represents a constraint violation within that number, being the total number of constraints; generation with is expressed as a per unit value to render constraints is the value resulting comparable with one another. is the user-dened from the current evaluation whereas constant. In general, constraints are dened as

for for

for Type 1 constraints for Type 2 constraints

(5)

and allude to the existence of two different In (5), types of inequality constraint in an engineering problem of this kind, namely: constraints which constitute the Type 1: There are permissible lower limits. For instance, high values for power factor are desired for good performance in an induction motor (the higher the better); hence the need for dening an acceptable lower limit. If the user specied constraint for the power factor were 0.85, then anything less than that would be a violation. Thus, for a returned value of 0.83, the numerator in (4) would be 0.02; a positive value, indicating a violation. Type 2: There are constraints which constitute the permissible upper limits. For instance, stator current density may not exceed certain values on account of permissible machine operating temperatures; hence the need for dening an acceptable upper limit. If the user specied constraint for the stator current density were 5 A/mm , then anything more than that would be a violation. Thus, if the algorithm returns a current density value of 6 A/mm , the numerator in (4) would be 1; this time a negative value indicating a violation. The latter type of constraint explains the use of the absolute notation in (4). On the other hand, constraints cannot be equal to zero for practical reasons. For instance, zero current as a constraint would mean that there is no current in the windings when the machine operates at full load; a very desirable but impossible

situation. Similarly, a zero constraint value for the stator slotlling factor would imply that no windings have been tted into stator slots. To avoid such absurdities, a provision is made for the algorithm to block any constraint from being active if the corresponding . At the beginning of program execution, the designer species the constraints to be present during optimization. If a nonzero value is entered for a constraint, this constraint is present during optimization; if the value is zero, the constraint is disabled, i.e., it is not present during optimization. This offers the user a choice by blocking selected constraints from being active. Consequently, division by zero will never occur in (4) during algorithm execution. If a constraint is not violated, then the corresponding becomes zero, implying that there is no penalty associated with this constraint. If a constraint is violated, then the corresponding value of violation in per unit terms, as dened in (3). For in (4), is introduced into the calculation of example, if the current density in stator conductors resulting from a design evaluation is 6 A/mm and the user-specied constraint for it is 5 A/mm , then the resulting violation will . Thus, can only assume be equal to abs either positive values in the case of nonfeasible solutions or is equal to zero in cases when none of the constraints is violated. The constraints are somewhat exible in that the devised penalty function allows for slight constraint violations. This means that the nal solution resulting from the optimization may lay slightly outside the feasible space. The use of such exible constraints facilitates the nding of optima, which may be located just outside of a specied feasible region. Often, designers will be prepared to relax some of the constraints in order to gain signicant improvements in the value of the objective function (e.g., machine cost or efciency). If such an infeasible design cannot be accepted, the designer can force a reduction in the constraint violation by multiplying the by a suitably selected constant to increase such relevant a constraints weight. In the case presented, efciency was selected as the objective function to be maximized subject to the following ten nonlinear inequality constraints: maximum stator slot lling factor 0.69; maximum magnetic ux density in stator teeth 1.7 T; maximum magnetic ux density in stator yoke 1.55 T; maximum magnetic ux density in rotor yoke 1.5 T; maximum current density in stator conductors 6 A; 7; maximum per unit starting current minimum per unit pull-out torque 2.2; minimum per unit starting torque 1.3; maximum full-load slip 0.03; minimum full-load power factor 0.87. Fig. 2 illustrates the program structure. Execution of the program starts with the specication of the nonvariant design parameters such as the number of poles, materials used, and magnetization characteristics of core materials. This is followed by the selection and specication of the constraints, as well as the number of generations, bit-string length, and population size.

WIECZOREK et al.: AN EVOLUTIONARY ALGORITHM

3885

Fig. 2. Program structure.

The initial population of (in our experiments, 200) individuals is generated next. Each chromosome consists of 11 genes representing randomly selected values for the 11 variables from within specied domains. Every gene is made up of 20 bits, including those representing the two discrete design parameters (the number of conductors per stator slot and the stator wire gauge). These discrete design parameters require modication for each population in order to only assume the permissible discrete values. In the case of the number of conductors per stator slot the above is achieved by rounding off the random number to the nearest integer. The other real number, delineating the stator wire gauge, is used after rounding off to the nearest integer as a pointer to the table of standard wire gauges setting the stator conductor size for the subsequent design evaluation. All other variables are continuous and do not require any special treatment. Both the probability of crossover and mutation rate are user denable; in the case presented in these experiments they were 0.9 and 0.005, respectively. The design is evaluated for every individual of a population; the algorithm terminates after performing the specied number (in our experiments, 20) of generations. At this point, the designer is offered the option to view the performance characteristics for the proposed design. Table II contains the results of eight consecutive program executions. Due to the random nature of genetic algorithms, the computation time of a single execution varies slightly, with the approximate average value being just over three hours

on a desktop computer with a Pentium 120 MHz processor. Values for the eleven variables are presented for each run, together with the values of wire gauge and airgap ux with which the optimal value for the objective density function (Efciency) has been achieved. The algorithm has returned an acceptable solution every time, which is indicated by a good value for the objective function with no constraint violations. The optimized efciency is within a narrow band with a mean of 92.316%, which is entirely satisfactory for this frame size. The ensuing air-gap magnetic ux density values are included in the table to indicate that the algorithm naturally selects values that have been historically established as being suitable for this frame size. Fig. 3 depicts examples of performance characteristics displayed at the program termination, in this case for run number 8 from Table II. The solid curves represent the optimized performance, whereas the curves with broken lines represent the performance of the existing motor calculated using the same design algorithm. Test results for the existing motor are also superimposed (shown as ). The inclusion in the graphical representation of constraints together with the full load operating point (shown as ) is most useful as it allows for a better interpretation of the results; constraints (shown as dotted lines) are as follows. Line current versus speed characteristic: horizontal line represents the maximum starting current constraint; the vertical line represents the maximum slip constraint at full load. Shaft torque versus speed characteristic: the upper horizontal line represents the minimum pull-out torque; the lower horizontal line represents the minimum starting torque constraints. The vertical line depicts the maximum permissible slip at full load as in the rst characteristic. Efciency versus shaft torque: no constraints are indicated. Power factor versus shaft torque characteristic: minimum acceptable power factor at full load constraint is shown. This facility may prompt the designer to review constraint specication and may wish to relax some of the constraints toward improving the value of the objective function. It is observed that the performance of the existing machine does not satisfy the criteria set in this case. For example, the constraints of maximum starting current and minimum power factor at full load are both violated (sub-plots 1 and 4 of Fig. 3). The performance of the optimized machine is seen to be improved respectably when compared with the existing machine. The motor used as a benchmark for algorithm verication is the product of a long evolutionary design process. The use of the EA-based optimization approach enables similar or better results to be obtained without the benet of accumulated design experiences over a hundred years! A further signicant aspect is that of minimal need for interactive design input. IV. CONCLUSION It has been shown that EA based algorithms constitute a viable and powerful tool for the optimal design of induc-

3886

IEEE TRANSACTIONS ON MAGNETICS, VOL. 34, NO. 6, NOVEMBER 1998

TABLE II EXECUTION RESULTS

(a)

(b)

(c)

(d)

Fig. 3. Sample performance characteristics: optimized motor, - - - - - existing motor, 3 full load operating point,  measured results, and constraints.

tion motors. The approach described is suitable for handling MIP problems, implicit in electromagnetic device design, without having to resort to NLP approximation techniques. The approach is universal and, although demonstrated here

for induction motor design, may be applied to the design optimization of other types of electromagnetic device. Another valuable feature is that of being implementable on a desktop computer.

WIECZOREK et al.: AN EVOLUTIONARY ALGORITHM

3887

The EAs intrinsic tendency to search for global optima requires some caution in the implementation of design algorithms. For example, in attempting to explore the feasible space, EA may select a set of parameters for which magnetic properties ( - curves, iron losses) are unknown; a situation which may occur when the value of magnetic ux density exceeds, due to selection by the algorithm of very narrow teeth, the available values. Such problems are overcome by a range of measures including the allocation of a low tness value each time such an anomaly is encountered. The results obtained for a realistic device such as the induction motor presented here substantiate the usefulness of evolutionary optimization methods in electromagnetic device design. The algorithm is capable of nding near-global optima with a considerable level of condence. The accuracy of convergence can be improved further by additional enhancements to the EA solver. Improvements may include the use of oating point representation of individuals (including integer representation for two discrete variables), ranking or tournament selection, etc. Also, it is possible to accelerate the convergence by means of hybridization [8]; a near-optimal solution found by the EA can be pursued by a suitable gradient method such as the sequential quadratic programming for quick and accurate convergence. REFERENCES
G [1] O. ol and J. P. Wieczorek, Use of MATLAB in induction motor design optimization, in Proc. 1st Australian MATLAB Conf., Melbourne, Australia, 1996, pp. EE1824. [2] , Application of nonlinear programming techniques in electromagnetic device design, in The Role of Mathematics in Modern Engineering, A. K. Easton and J. M. Steiner, Eds. Lund, Sweden: Chartwell-Bratt, 1996, pp. 417424. Data Structures = Evolution [3] Z. Michalewicz, Genetic Algorithms Programs, 3rd ed. Berlin, Germany: Springer-Verlag, 1996. [4] W. N urnberg, Die Asynchronmaschine, 2nd ed. Berlin, Germany: Springer-Verlag, 1963. [5] W. Schuisky, Induktionsmaschinen. Vienna, Austria: Springer-Verlag, 1957. G [6] O. ol, Induction machine models for design, in Proc. Int. Conf. Evolution Modern Aspects of Induction Mach., Turin, Italy, 1986, pp. 510.

[7] J. A. Joines and C. R. Houck, On the use of nonstationary penalty functions to solve nonlinear constrained optimization problems with GAs, in Proc. 1st IEEE Int. Conf. Evolutionary Computation, Z. Michalewicz, D. Schaffer, H.-P. Schwefel, D. Fogel, and H. Kitano, Eds. Piscataway, New Jersey; Orlando, FL, 1994, vol. 2, pp. 579584. [8] H. Myung, J.-H. Kim, and D. B. Fogel, Preliminary investigations into a two-stage method of evolutionary optimization on constrained problems, in Proc. 4th Annu. Conf. Evolutionary Programming, J. R. McDonnell, R. G. Reynolds, and D. B. Fogel, Eds. Cambridge, MA: MIT Press, 1995, pp. 449463.

Jan Pawel Wieczorek received the Master of Engineering Science degree in electrical engineering from the Technical University of Lodz, Poland. He is currently working towards the Ph.D. degree. His current research is in the eld of electrical machines. His interests are focused on a range of areas related to electrical machines such as electromagnetic device design, magnetic materials, modeling and simulation, energy efciency in electromagnetic energy conversion processes, and the use of mathematical optimization in engineering.

Ozdemir G ol received the MESc degree from the Technical University of Istanbul, Turkey, the M.E. degree from the University of Melbourne, Australia, and the Ph.D. degree from the University of Adelaide, Australia, all in electrical engineering. He has had extensive experience in the eld of electrical machines, both in industry and as an academic. He is currently Head of the Electrical Engineering Department and Head of Electrical Machines and Drives Research Group at the University of South Australia. His interests include dynamic modeling of electrical machines using numerical methods, design and analysis of novel electromechanical energy conversion devices, and application of mathematical techniques to design optimization of electromagnetic devices.

Zbigniew Michalewicz received the M.Sc. degree from the Technical University of Warsaw, Poland, in 1974 and the Ph.D. degree from the Institute of Computer Science, Polish Academy of Sciences, Poland, in 1981. He is Professor of Computer Science at the University of North Carolina at Charlotte. His current research interests are in the eld of evolutionary computation. Dr. Michalewicz was the General Chairman of the First IEEE International Conference on Evolutionary Computation, Orlando, FL, June 1994. He is a member of the editorial board of several publications including IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION and the Journal of Advanced Computational Intelligence.

You might also like