Comparison of Classifiers for Power Quality Disturbances with Wavelet Statistical Analysis

In this paper, an attempt is made to compare the Power Quality (PQ) recognition accuracy with efficient features. Initially, a few of the reliable statistical parameters such as the mean, standard deviation, Root Mean Square (RMS), form factor, crest factor, Shannon entropy, log entropy, normalized entropy, skewness, and kurtosis of eight synthetically generated PQ disturbances and the pure tone signal are computed. These statistical parameters are simple to compute and are of low-dimension. A host of classification techniques such as the K-nearest Neighbor (KNN), Discriminant Analysis (DA), Decision Tree (DT), Support Vector Machine (SVM), Naïve Bayes' (NB), and Random Forest (RF) have been put to test for performance appraisal to determine the discriminating ability of these parameters. Further, the applicability of multi-resolution Wavelet Transform (WT) has been explored to extract these chosen statistical parameters in the WT domain for a possible enhancement in recognition accuracy. The result shows, the RF remains the slowest with the highest accuracy while the performance of the DA remains the poorest. The WT has outperformed the baseline method of statistical analysis as revealed from our results. However, the KNN tends to provide the highest classification accuracy among all others for low feature dimension whereas the speed of response of DT has been fastest.