An incremental high impedance fault detection method under non-stationary environments in distribution networks

In the non-stationary environments of distribution networks, where operating conditions continually evolve, maintaining reliable high impedance faults (HIF) detection is a significant challenge due to the frequent changes in data distribution caused by environmental variations. In this paper, we propose a novel HIF detection method based on incremental learning to handle non-stationary data stream with changing distributions. The proposed method utilizes stationary wavelet transform (SWT) to extract fault characteristics in different frequency do-mains from zero-sequence current data. Subsequently, a complex mapping from signal features to operational conditions is established using backpropagation neural network (BPNN) to achieve online detection of HIF. Additionally, signal features are analyzed using density-based spatial clustering of applications with noise (DBSCAN) to monitor the distribution of data. After encountering multiple distribution changes, an incremental learning process based on data replay is initiated to evolve the BPNN model for adapting to the changing data distribution. It is worth noting that the data replay mechanism ensures that the model retains previously acquired knowledge while learning from newly encountered data distributions. The proposed method was implemented in a prototype of a designed edge intelligent terminal and validated using a 10 kV testing system data. The experimental results indicate that the proposed method is capable of identifying and learning new distribution data information within non-stationary data stream. This enables the classifier model to maintain a high level of detection accuracy for the current cycle data, effectively enhancing the reliability of HIF detection.