Optimized Neural Network Speed Control Of Im Using Genetic Algorithm

SPEEDAM 2006 International Symposium on Power Electronics, Electrical Drives, Automation and Motion

Optimized Neural Network Speed Control of Induction Motor using Genetic Algorithm W.S.Oh *, K.M.Cho*, S.Kim*, and H.J.Kim** * Yuhan College (Korea) ** Hanyang University (Korea) Abstract-G For the high performance drives of induction motor, recurrent artificial neural network (RNN) based self tuning speed controller is proposed. RNN provides a nonlinear modeling of motor drive system and could give the information of the load variation, system noise and parameter variation of induction motor to the controller through the on-line estimated weights of corresponding RNN. Self tuning controller can change gains of the controller according to system conditions. The gains are composed of the weights of RNN. For the on-line estimation of the weights of RNN, extended kalman filter (EKF) algorithm should be used. In order to design EKF with optimal constants, simple genetic algorithm is proposed. Genetic algorithm can follow the optimal estimation constants without trial and error efforts. The availability of the proposed controller is verified through the MATLAB and Simulink simulation with the comparison of conventional controller. The simulation results show a significant enhancement in shortening development time and improving system performance over a traditional manually tuned EKF estimation algorithm based neural network controller. Index Terms—neural network, induction motor speed control, genetic algorithm, extended kalman filter

although it requires the same derivative information. The EKF does not require batch processing of data and therefore it is very suitable for on line training. But according to parameters that are needed in using EKF, the performance of EKF is varied. So it takes a time consuming effort to find good parameters. In order to design EKF with optimal constants, simple genetic algorithm is proposed. Genetic algorithm is well established techniques which have recently found extensive application in an intelligent way to find values close to the global optimum. Genetic algorithm (GA) can follow the optimal estimation constants without trial and error efforts. The availability of the proposed controller is verified through the MATLAB and Simulink simulation with the comparison of conventional controller. The simulation results show a significant enhancement in shortening development time and improving system performance over a traditional manually tuned EKF estimation algorithm based neural network controller. Experimental results are under test. II. NEURAL NETWORK SPEED CONTROLLER

I. INTRODUCTION

A. Self tuning controller

In recent years artificial neural network (ANN) has gained a wide attention in control applications.[1][2][3][4] The ANN provides a nonlinear modeling of motor drive system without any knowledge of predetermined model and thus makes the drive system robust to noise, parameter variations, load changes. The concept of model reference adaptive control is used in training ANN to achieve trajectory control of induction motor. Most of the ANN based adaptive control approaches use off-line system identification. Because of this separation, it is impossible to effectively cope with the system parameters that are changed dynamically during operation. This makes the tuning of the respective controller parameters difficult. Thus, an on-line learning process is desirable.[5] EKF algorithm has shown significant merits for training both feedforward and recurrent neural networks. The training algorithm based on the EKF is shown to require significantly smaller training data than the pure gradient descent algorithms [2][4] and is simpler than gradient descent algorithms

1-4244-0194-1/06/$20.00 ©2006 IEEE

In order to design speed controller of the induction motor, we define the following performance index.[1]

[

]

J (u , k ) = y ( k + d ) − y ref ( k + d ) + ρ u [Ru ( k ) ] 2

+ ρ v [v e ( k + d ) ]

2

2

(1)

y (k ) is the output value and u (k ) is the input value. y ref (k ) is the desired reference trace. ρ u , ρ v are the weighting factor used to adjust the tradeoff between control signal and accuracy of system response. If the derivative of J (u , k ) with respect to the control signal u (k ) at the time k is equal to zero, the control signal is obtained.

B. Recurrent Recursive Neural Network Model The dynamic system representation capabilities of RNNs have been shown to be considerably greater than those of purely static networks.[2][5] Fig.1 shows the

S14 - 12

general architecture of RNN, where Z −1 indicates one sampling time delay. u M (k ) is the network input, and Y N ( k + 1) is the output of the network and is produced one step ahead at the discrete time (k+1). WN ( N + M ) (k ) are weights of the network. We can design an induction motor model which has one input u1 ( k ) which is current I qs (k ) and two output Y1 (k ) and Y2 ( k ) . Y2 (k ) is the speed output of the induction motor. C. Design of Induction Motor Speed Control System We could show that the speed output model of the induction motor is a function of speed and input like Eq.(2). y2 (k +1) = f [y2 (k), y2 (k −1), u(k), u(k −1)]

III. GENETIC ALGORITHM

(2)

To minimize the performance index J (u, k ) of Eq.(1), The control input is shown like Eq.(3) at sampling time k. u(k ) =

1 a1 y2 (k ) + a2 y2 (k − 1) + a0 yref ( k + 1) + b1u (k − 1) + b2ve (k ) b0

[

Fig. 1 Basic structure of RNN model

]

(3) Where, a 0 , a1 , b0 , b1 , b2 is constant composed of weights of the neural network. Using EKF algorithm, we could get the weighting matrix W . Then, we have control input value u (k ) through Eq.(3) that is composed of weighting values. u (k ) is the torque current component of induction motor, I qs (k ) .

In order to design EKF with optimal constants, simple genetic algorithm (GA) is proposed in this paper. Fig.2 shows the flowchart of GA [10]. In this paper, the crossover possibility is 0.85 and the mutation possibility is 0.002. The generation number is 100 and the individuals in generation are 20. The individual bit length is 24 bit which represents 8 bit normalization factors for S (k ) , 8 bit normalization factors for η (k ) and 8 bit normalization factors for Q(k ) . The composition equation and fitness value in ref.[10] are used for composition of proposed GA. Start Gen =0

.

Initialization

D. Extended Kalman Filter

In this paper, we use EKF algorithm to train weights of neural network. It is good for on-line learning [2][4]. But EKF has constants S (k ),η (k ), Q(k ) that are weighting matrix, learning parameter and covariance matrix respectively. These constants are chosen by user with trial and error method. The performance of the controller depends on the values of constants.

Reproduction Gen=Gen+1

No

Crossover & Mutation Fitness evaluation End ?

Yes

Result End

G

G

Fig. 2 Flowchart of genetic algorithmG

IV. SIMULATION RESULTS In order to verify the validity of the proposed GA based neural network controller, several simulations are carried out using MATLAB and Simulink software. Fig.3 shows the block diagram of proposed control system. In

S14 - 13

RNN, the weights of the neuron are calculated using EKF algorithm and go to the Self Tuning Neural Network Controller to tune its gains. Simulations are focused on the facts whether the proposed GA based Neural Network (NN) algorithm is better than normal NN algorithm or not. For the comparison, simulations of the speed response are performed in case of the speed command variation, the load variation of induction motor. Specifications of motor are 1hp, 230V, 4 pole. The sampling time of the controller is 100 μ sec . Fig.4, 5 show speed response waveforms in case of load variation using a normal NN speed controller where EKF gains are designed arbitrary and proposed GA based NN speed controller respectively. The command speed is 180rad/sec that is increased from zero speed and the 50% disturbance load of the rated torque is applied at 7sec. Fig.6,7 show speed response waveforms with applying full load. In all point of view such as variable speed characteristics and speed recovery time, proposed GA based NN controller is better than normal NN controller. We can observe that proposed controller is more robust to the inertia variation.

Fig.5 Speed response I of NN with genetic algorithm (50% load)

G

Fig. 3 Block diagram of proposed control system

Fig. 6 Speed response II of NN controller with arbitrary constants(full load)

G G

Fig. 4 Speed response I of NN controller with arbitrary constants(50% load)

Fig.7 Speed response II of NN with genetic algorithm(full load)

S14 - 14

V. CONCLUSION In this paper, a GA based neural network controller for induction motor drives was presented. The structure of proposed controller is based on the self tuning controller that can tune its control gain to be autonomous according to system parameter variations. A RNN induction motor model is used to identify the variation of induction motor system. EKF algorithm is used for the training of the weights of RNN that represent parameter variations. Genetic algorithm is used for EKF constants to be optimal. The designed controller also can compensate for uncertainties of nonlinear induction motor control system since the real output values are directly used for parameter identification and tuning. Simulation results using MATLAB verify the effectiveness of proposed controller.

REFERENCES [1] M. A. El-Sharkawi and S. Weerasooriya: Developmentand Implementation of Self-tuning Tracking Controller for DC Motors, IEEE Trans. on energy conversion, pp.122-128, March 1990. [2] Joao O. P. Pinto, Bimal K. Bose, Luiz E. B. Silva: A stator flux oriented vector-controlled induction motor drive with space vector PWM and flux vector synthesis by neural networks, IAS2000, pp.1605-1612, Sep. 2000. [3] K. S. Narendra and K. Parthasarathy: Identification and control of dynamic systems using neural networks, IEEE Trans. Neural Networks, vol.1, pp.4-27, Jan.1990. [4] M. A. Hoque, M.R. Zaman and M.A. Rahman: Artificial Neural Network Based Permanent Magnet DC Motor Drives, IAS95, pp.98-103. [5] G. V. Puskorius and L. A. Feldkamp: Neurocontrol of nonlinear dynamic systems with Kalman filter trained recurrent networks, IEEE Trans. Neural Networks, vol.5, pp.279-297, Mar. 1994. [6] Goodwin, G. C. and K. S. Sin, Adaptive Filtering Prediction and Control, Prentice Hall, 1984. [7] B. K. Bose, Power Electronics and AC Drives, Prentice-Hall, 1986. [8] Goldberg, David E., Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley, 1989. [9] Hsu, Y.P., Tsai, C. C.: Autotuning for Fuzzy-PI control using genetic algorithm, IECON96, pp602-607. [10]W. S. Oh, Y. T. Kim, C.S. Kim, T. S. Kwon, H. J. Kim: Speed control of induction motor using genetic algorithm based fuzzy controller, IECON99, pp625-629. .

S14 - 15