Research on Fault Detection and Localization Techniques for Distribution Networks Based on Edge Clustering

Abstract

With the intelligent transformation of the power grid, the number and type of various terminals, sensors, and new types of loads in the distribution network increase, and the huge amount of information and noise information accessing the power grid brings great challenges for fault detection and localization. In this paper, we propose to combine wavelet transform and LSTM unit to form a novel neural network basic unit TFM and examine the fault detection and localization capability of TFM by combining them with a typical example system, IEEE33. Spectral clustering and the K-Means algorithm are utilized to cluster the edge nodes, and the number of nodes in the three partitions after partition correction is 11, 10, and 12 nodes in order. The data shows that the TFM system performs the fault diagnosis task with better test accuracy than LSTM in every dimension, and its accuracy improvement for fault localization is the largest, at 5.39%. Introducing two reconfiguration scenarios for three edge node partitioning models and retrograde fault localization detection, compared with the no reconfiguration scenario, their localization accuracies all produce different degrees of decline, but always not less than 80%. The localization accuracy of the MⅡ model is still not less than 98% in the fault resistance range of 500~1000Ω, which proves that the TFM system can effectively extract high-resistance fault features.