关键词: 配电网故障辨识, 多源特征融合, 傅里叶卷积, 自适应去噪, 时序-通道注意力机制

Abstract: In response to the problems that fault features are easily masked by noise in the process of fault identification in distribution networks and it is difficult for a single feature to comprehensively reflect the fault information, this paper proposes a multi-source feature fusion denoising network method for fault identification in distribution networks. Firstly, the discrete wavelet transform (DWT) is used to decompose the normalized zero-sequence voltage and three-phase current signals in multiple scales, so that the original signals can be reconstructed into high-frequency and low-frequency components, and the effect of feature overlapping can be reduced; secondly, in the high-frequency branch, the temporal -frequency residual adaptive denoising module is constructed, and the high-frequency features with rich frequency-domain information are embedded in the Fourier convolution to extract the high-frequency features. rich frequency domain information, and combined with the channel attention mechanism to realize the accurate denoising of the signal; in the low-frequency branch, a lightweight convolutional network is designed to extract the temporal features of the low-frequency components to enhance the complementarity of the time-frequency information; finally, a temporal-channel attention mechanism is introduced for the adaptive feature fusion, which enhances the feature extraction capability while suppressing the redundant features, so as to realize the accurate diagnosis of the fault types under the complex working conditions. The proposed method is used to accurately diagnose the fault types under complex working conditions. The experimental results show that the proposed method achieves 99.12% accuracy in the simulation data set, which significantly improves the accuracy of fault identification compared with the existing methods.

Key words: distribution network fault identification, multi-source feature fusion, fourier convolution, adaptive denoising, temporal-channel attention mechanism