Performance Evaluation of Machine Learning Approaches for Fault Classification in Distribution Networks

This research presents how to evaluate the performance of the Machine Learning Classifiers (MLC) as an intelligent fault diagnosis (IFD) for fault classification in distribution networks under different operation condition whether these cases are normal or fault cases. IEEE 16-bus system was studied as a case of distribution systems. The approach utilizes the Discrete Wavelet Transform (DWT) for transient current signal processing, using the MATLAB Wavelet Toolbox, specifically applying the maximum (D8) from the Sym3 wavelet at the 8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> level as input for (MLC). The study evaluates nine classification algorithms including Decision Tree, Discriminant Analysis, Logistic Regression, Efficient Linear SVM, Naïve Bayes, SVM, K-Nearest Neighbors Classifiers (KNN), Ensemble methods, Neural Network and Kernel through 34 different classifiers to assess their performance based on accuracy, prediction speed, and training time for phases A, B, C, and G. The results show that both Gaussian Naïve Bayes and Kernel Naïve Bayes achieved perfect validation accuracy (100%) in detecting faults in the distribution network for all phases. However, Gaussian Naïve Bayes outperformed Kernel Naïve Bayes in terms of prediction speed and training time, making it the most suitable classifier for this application. The classifiers effectively distinguished different fault types, including Line-to-Ground Fault (L-GF), Line-to-Line Fault (LL-F), Line-to-Line-to-Ground Fault (LL-GF), Three-Phase Fault (3L-F) with nearly 100% validation accuracy.