Hybrid soft computing techniques for feature selection and parameter optimization in power quality data mining

Recognition of the presence of any power disturbance and classifying any existing disturbance into a particular type is the first step in combating the power quality problem. In spite of the extensive number of power disturbances classification methods, a research on the selection of useful features from the existing feature set and the parameter optimization for specific classifiers was omitted. The kernel parameters setting for support vector machine (SVM) classifier in training process along with feature selection will significantly impact the classification accuracy. Two novel wrapper based hybrid soft computing techniques are proposed in this paper for feature selection and parameters optimization to classify nine types of power disturbances without degrading the SVM classification accuracy. The feature items were selected from discrete wavelet transform across several decomposition levels of the disturbance signals and from the duration of disturbance occurrence. This analysis selects the more useful feature set and optimized parameters for two types of kernels namely the polynomial kernel and radial basis function kernel for SVM. Compared with the traditional grid algorithm the proposed genetic algorithm and simulated annealing based approach significantly improves the classification accuracy rate by eliminating relatively useless feature items and proper parameter selection for the classifier.