High Impedance Fault Feature Extraction and Detection Method for Distribution Network

Traditional high impedance fault (HIF) detection methods face significant technical challenges, including difficulties in feature extraction and limited flexibility in threshold selection, which lead to misjudgment in extreme fault scenarios. Therefore, an HIF detection method for the distribution network is proposed. Firstly, the time-frequency distribution differences of transient signals between the HIF and normal disturbance condition are analysed by Shannon entropy quantization of wavelet packet. On this basis, the transient signal time-frequency waveform block with the lowest similarity is selected as the input sample, and the Dropout in the traditional Vision Transformer (ViT) is replaced by a new regularization method, DropKey, so as to construct a DropKey-Vision Transformer (DVit) classification model, which is suitable for the small-sample scenario of HIF detection for the distribution network. Finally, simulation and experimental test results demonstrate that the proposed method achieves an average accuracy exceeding 99.5% for detecting 10 kΩ HIFs. This represents an improvement of at least 1.5% compared to previous methods and an enhancement of approximately 2% to 7% relative to other techniques. Additionally, the method is applicable to arc grounding, grassland grounding, and pond grounding fault detection, exhibiting high robustness. Results from Grad-CAM visualization further validate the effectiveness and superiority of the proposed method.