Series Arc Fault Detection Based on Random Forest and Deep Neural Network

Series arc is prone to cause fire accidents, but its occurrences induced by different load types and connections make the detection challengeable. This paper proposes a series arc fault detection and location algorithm for multi-load circuit topology, especially for branch arc faults and nonlinear power loads. Several typical loads of paralleling connected are considered to measure the current changes caused by the arcing phenomenon at different positions. Different aspects of arc features are extracted by time-domain, frequency-domain, and wavelet packet energy analysis. A feature selection method based on random forest (RF) is adopted to determine the specific feature sets according to the reduction of Gini impurity. The integrated top ten features with a high correlation to the arc were selected for different combinations of loads and are input into the deep neural network (DNN) for calculating and training. Eventually, a comprehensive arc detection model with the function of detection and location determination for different load types is obtained. It proves that the proposed RF-DNN-based arc detection algorithm can identify and protect the series arc faults in multi-load scenarios efficiently and accurately to meet the practical needs.