Explainable incremental learning for high-impedance fault detection in distribution networks

To enhance the generalization and explainability of data-driven models in fault detection, this study introduces a cutting-edge detection approach anchored in explainable incremental learning for high impedance faults (HIFs). Leveraging the discrete wavelet transform, our method discerns the wavelet coefficients of the zero sequence current, offering a quantitative lens to view HIFs via the calculation of the standard deviation. In succession, an artificial neural network (ANN) is refined using a regularization-based principle. This guiding principle charts the model's evolutionary path, emphasizing weight regularization and incorporating penalty terms into the loss function. Such an approach dynamically optimizes model parameters, ensuring the assimilation of novel knowledge found in waveform data streams while safeguarding against the detrimental effects of forgetting prior knowledge. The robustness of the proposed methodology is corroborated using the PSCAD/EMTDC platform and real-world field data. In addition, this paper delve into a comprehensive explainable analysis using Shapley's additive attribution theory, with an objective to elucidate model explainability from both a holistic and granular viewpoint.