Fault Classification in Power Distribution Systems using PMU Data and Machine Learning

Abstract

This work presents the analysis of machine learning methods for fault (short-circuit) classification in electrical distribution networks using data from PMUs (Phasor Measurement Units) installed along the network. The Alternative Transient Program was used to simulate 26,928 different instances distributed into 33 types of faults – single and multi-phase, including or not the ground and different wire breakages – and one normal condition of the system. The IEEE 123-bus distribution system was used as the test system. We compared five machine learning methods for classification: Linear Discriminant Analysis (LDA), Artificial Neural Networks (ANN), Support Vector Machines (SVM), k-Nearest Neighbors (kNN), and Decision Trees (DTs). The best result was achieved by the SVM with Gaussian kernel and ANN. The input data (feature extraction) was also varied, testing data from one or several PMUs, ABC sequence phasors and symmetrical sequence phasors. We obtained slightly better results for symmetrical components and multiple PMUs in the network. Finally, classes of the same short-circuit with different wire breakages were grouped, raising the overall classification accuracy, showing the feasibility of this approach for fault classification using PMU-data in a distribution network.