Feature selection for power quality event identification

This study investigates the performance of a filter based feature selection approach for PQ event identification. The filter based approach is independent of the nature of induction algorithms used for the classification purposes and therefore offers several advantages over other approaches such as wrappers. This property of filter approach has been exploited in this study to obtain a generic feature subset that can be used with any induction algorithm. For this purpose, fourteen distinct single and simultaneous PQ events were simulated following IEEE Std. 1159. The feature selection of these events is accomplished through a combination of meta-heuristic search method and correlation based feature evaluation. Two meta-heuristic search methods based on Genetic Algorithm (GA) and Binary Particle Swarm Optimization (BPSO) have been included for the comparison of search performance. The efficacy of reduced feature subsets is evaluated through induction algorithm based on Naive-Bayes. The results convincingly demonstrated that it is possible to obtain significant reduction in the features (approx. 50% with GA, 86% with BPSO) without any compromise in the classification performance.