VisGCL: Visibility Graph Convolutional Learning on Time Series Data for Arc Fault Detection in Low-Voltage Distribution Networks

Abstract

Arc faults in low-voltage distribution networks significantly threaten power system safety due to their randomness and concealment. Traditional arc fault detection methods, which rely on time-domain and frequency-domain features, often struggle with accuracy and robustness in variable load environments. To address these challenges, this paper introduces Visibility Graph Convolutional Learning (VisGCL), a novel approach that segments current signals into visibility graphs and employs hierarchical graph convolutional networks for analysis. This method directly learns arc failure modes from the graphical representation of current signals, simplifying the detection process and enhancing both accuracy and robustness. Experimental results demonstrate that the proposed method achieves an accuracy of 98.58 ± 0.14%, with precision, recall, and F1-score reaching 98.05 ± 0.25%, 98.36 ± 0.47%, and 98.16 ± 0.23%, respectively. Extensive experiments validate the effectiveness of VisGCL, confirming its superiority in detecting arc faults under diverse load conditions, and offering a new efficient and reliable solution for arc fault detection in low-voltage distribution networks.