Faulty feeder detection for single phase-to-ground faults in distribution networks based on patch-to-patch CNN and feeder-to-feeder LSTM

Abstract

Faulty feeder detection has always been recognized as an essential issue in guaranteeing the stable and reliable operation of distribution networks. Existing methods fail to fully exploit the spatial correlation and temporal dependencies in the zero-sequence currents, which are exhibited as waveforms on the spatial domain and are displayed as time-series data on the time domain. In this paper, a novel detection method based on a hybrid model of convolutional neural network (CNN) and long short-term memory (LSTM) neural network is proposed, and the spatial–temporal information can be comprehensively explored. Deep spatial features are extracted by using CNN, and they are further processed by using LSTM network. Besides, to maintain the temporal dependencies of the extracted spatial features, the current waveforms are recognized on the patch scale by the proposed patch-to-patch CNN (PToP CNN). Moreover, a feeder-to-feeder LSTM (FToF LSTM) network is established to learn and compare spatial–temporal correlations between feeders. The joint PToP-CNN and FToF-LSTM can achieve collaborative superiority on mining the fault features in the zero-sequence currents. To verify the prospects in real installations, the hybrid model is implemented in an embedded device, called NVIDIA Jetson AGX Xavier, and the real-time detection demonstrates the efficiency of the proposed method.