Detection of single-phase-to-ground faults in distribution networks based on Gramian Angular Field and Improved Convolutional Neural Networks

Abstract

Fault detection of transmission lines is a fundamental guarantee for the stable operation of the power grid. The characteristic information is not obvious when a single-phase-to-ground fault occurs in the complex distribution network, and the existing methods are vulnerable to the problems of fault conditions and environmental noise. To address this problem, a new method of distribution network fault diagnosis is proposed based on Gramian Angular Field (GAF) and Improved Convolutional Neural Networks (ICNN) in this paper. Firstly, the time-series signals are normalized and transformed to a polar coordinate system. The encoded signals are converted into visual virtual images through the inner product calculation of GAF algorithm. Secondly, multi-scale module and attention mechanism are added to the convolutional layer of the ICNN. Among them, the multi-scale module, which enhances the local feature extraction capability for the original image, is composed of multiple convolutional channels. The purpose of attention mechanism is to help the weight distribution and depth feature filtering of ICNN, focusing on the key features between fault data. Finally, the transfer learning method is further introduced to improve the fault diagnosis model performance for different online situations. The experimental results show that the overall accuracy of the proposed fault location method is close to 99% under different topological situations. Even with the influence of ambient noise, the accuracy of GAF-ICNN still approaches 97.5%.