Data-driven algorithms can effectively reduce the influence of multiple uncertainties and noise interference on detection thresholds for high-impedance faults in distribution networks. However

the “black-box” nature of these models limits their application interpretability. Thus

a method for high-impedance ground fault detection and interpretability analysis in distribution networks based on temporal convolutional networks (TCNs) is proposed. First

an improved adaptive noise-complete ensemble empirical mode decomposition is employed to decompose and reconstruct the zero-sequence current

suppressing noise interference while enhancing fault feature expression. Then

a TCN is developed to extract temporal features from the processed waveforms

thereby improving the model’s ability to distinguish high-impedance faults from typical disturbance conditions. Subsequently

a fractional-weighted class activation mapping scheme is designed to analyze the model’s decision basis. By combining attribution indicators of key waveform regions

the method characterizes the correspondence between the distinctive “zero-off” features of high-impedance faults and the model’s decision-focused regions

thereby enhancing interpretability. Finally

based on MATLAB/Simulink simulation models and field test data

the effectiveness and reliability of the proposed method are validated.