Comparison of Machine Learning-Based Methods for High Impedance Fault Detection in Distribution Systems

High Impedance Faults (HIFs) occur when an energized conductor contacts a high-impedance surface. They are critical events in Distribution Systems (DSs) as they may cause bushfires, endanger living beings, and there is no fully efficient HIF Detection Method (HIFDM) so far. Machine learning-based HIFDMs have become increasingly popular nowadays, so it is important to understand their strengths and weaknesses before a practical application. Under this perspective, this paper proposes a study to compare different Machine Learning Algorithms (MLAs) in the same conditions to detect and classify HIFs considering background noise, different fault locations, and the influence of distributed generators. The input metrics are the harmonics extracted from the current signals using Fourier Transform. Results demonstrate the significant applicability of the MLAs - Decision Tree, Ensemble Tree, Random Forest, Nearest Neighbor, Support Vector Machine, and Artificial Neural Networks in detecting HIFs and classifying DSs’ events. Also, the methods were trained by using a simulation HIF model, but tested with real HIF signals, resulting in high successful rates.