Artificial neural network based implementation of space vector modulation for three phase VSI

Abstract

SVM (Space Vector Modulation) is among the most common PWM technique for three phase Voltage Source Inverter, SRM (switched reluctant motor), BLDC(Brushless DC) motor, and permanent magnet motor. Space vector PWM (SVPWM) has better harmonics performance along with higher rms voltage. Command voltage reference for all the three phases of three phase SVPWM is actuated as a whole. The logic behind this algorithm is that it averages the output vector of the inverter equal to the reference voltage vector. Due to complex nature of its operation, there is a computational delay involved in SVPWM, this often bounds its working up to few kHz of switching frequency. An Artificial Neural Network is used to solve this particular problem in this paper. The computational delay is negligible in case of feedforward neural network especially when a parallel architecture based dedicated application-specific IC (ASIC) chip is used. The conventional back propagation method undergo drawback such as overfitting the network. In this paper an alternate learning algorithm Bayesian Regularization method is used for training of the Neural Network. Bayesian regularized artificial neural networks (BRANNs) does not require cross-validation and are extra robust than conventional back-propagation neural network.