A data-driven high impedance fault detection framework for distribution network based on improved temporal convolutional network

High-impedance faults (HIFs) pose a great challenge for the protection of new-type distribution networks because of their weak characteristics and strong similarity to disturbance conditions such as capacitor switching and load switching. Therefore, an accurate and reliable HIF detection method is crucial for the safe power supply of distribution networks. A data-driven HIF fault detection framework for distribution networks is proposed, built upon an improved temporal convolutional network with confidence enhancement and interpretability. Firstly, the efficient channel attention temporal convolutional network (ECA-TCN) based HIF detection model is designed to exploit the temporal features in the transient zero-sequence current. Secondly, this paper proposes a HIF detection framework based on time-series sliding window and confidence level check, which can effectively achieve the balance between accuracy and rapidity and enhance the credibility of the HIF detection model in practical applications. Then, the score class activation map (Score-CAM) analysis method provides an explanation for the model’s decision-making process and enhances the model interpretability. Finally, the proposed framework is validated on both the MATLAB/Simulink platform and field tests, demonstrating superior performance compared with existing methods.