Faulty feeder detection based on image recognition of voltage-current waveforms in non-effectively grounded distribution networks

Abstract

Faulty-feeder detection is essential for fault location and fault isolation after single line-to-ground (SLG) faults in distribution networks. However, the SLG fault currents are weak, and the fault conditions are complex, thus posing great challenges for faulty-feeder detection. To improve the detection accuracy and reliability, a novel detection method based on waveform recognition is proposed, and it only requires zero-sequence voltage (ZSV) on bus and local zero-sequence current (ZSC), not ZSCs of all feeders. Firstly, the ZSV and local ZSC are superimposed in the same plot, and the ZSV-ZSC image for each feeder is generated. Subsequently, each image is individually identified using the established convolutional neural network with spatial attention residual learning blocks, which has strong discriminative capability. Finally, the fault and non-fault states of each feeder can be distinguished based on the corresponding image identification result. Large amounts of experimental results show that the proposed method with significant simplification of device installation can improve the accuracy of faulty-feeder detection considerably, demonstrating its strong generalization capability and good application prospects.