Series DC Arc Fault Detection Based on Ensemble Machine Learning

Abstract

Series dc arc fault creates a fire hazard and negative impacts on the distribution bus if not detected and isolated quickly. However, the detection of a series arc fault is challenging due to the low fault current, lack of zero-crossing, and the erratic behavior of arc discharge based on different power electronic loads and controllers in modern power applications. This article presents a practical and versatile series dc arc fault detection method based on ensemble machine learning (EML) algorithms. A buck converter constant power load (CPL) and a boost converter CPL are designed and built to study the different arc fault behaviors and generate training data for the machine learning algorithms. A set of time domain features is extracted from the experimental data and analyzed using the feature importance attribute. An adaptive normalization function then processed the features to mitigate false positive classification caused by load changes. A two-step algorithm is proposed to recognize the arc fault in different load types. Various EML algorithms and the associated hyperparameters are benchmarked to select the most accurate hyperparameters for a detection algorithm for low-cost hardware implementation. Finally, the detection algorithm's effectiveness is verified with CPL testbed.