AC Series Arc Fault Detection Based on RLC Arc Model and Convolutional Neural Network

Abstract

AC series arc faults in the power system can lead to electrical fires. However, the generalization performance of the determined detection method would be affected under unknown loads, as current features vary with loads. To address this issue, this paper presents a series arc fault detection method based on a high-frequency (HF) RLC arc model and one-dimensional convolutional neural network (1DCNN). By the current transformer used for receiving differential HF features (D-HFCT), current with complex features is firstly simplified and divided into different oscillation-signal types. Since the types of real D-HFCT data are limited, the RLC arc model is used to generate D-HFCT data with various types of oscillation features by adjusting load types, initial phase angles and Bernoulli-sequence frequencies. Then, the simulated data are adopted to train the 1DCNN model. Finally, the trained 1DCNN model can detect series arc faults under different types of real loads. Compared with the 1DCNN method driven by the limited types of real-current data, the presented method shows good generalization ability and achieves 99.33% average detection accuracy under nine types of unknown loads, which benefits from the training of simulated D-HFCT data with abundant HF oscillation features.