An Investigation on Feature Extraction and Feature Selection for Power Quality Classification with High Resolution and RMS Data

This paper shows different state-of-the-art machine learning methods for structured data, applied to classification of power quality data sets. k-Nearest Neighbor, Support Vector Machine, Random Forest, XGBoost and LightGBM are chosen for comparison of classification of high resolution and root mean square data. Discrete wavelet transform and TsFresh are chosen for the pre-processing of the high-resolution data and the extraction of features. For feature selection, mutual information filtering, feature importance und sequential feature selection are tested. Special to this investigation is the use of both - highresolution waveform data (sample rate 5 kS/s) and root mean square data (20 ms). The input data were synthesized by mathematical equations. The highest score is achieved by the XGBoost classifier in combination with the LightGBM feature selector. Accuracy shows 97.71% for root mean square data and 98.96% for the high-resolution data. Furthermore, the results illustrate the dependency of the classification result on the data structure, feature extraction, feature selection and the classifier.