Abstract

Detecting faulty feeder for single-phase-to-ground (SPG) faults in distribution networks is challenging due to weak fault currents and complex fault transients. The existing detection methods either possess insufficient and incomplete feature extraction capabilities or lack comprehensive credibility evaluation, thus leading to low reliability of detection results. This paper introduces a highly reliable fault detection method based on image recognition using fully convolutional generative adversarial network (FCGAN), which enhances both feature extraction and credibility evaluation capabilities. Firstly, both sampled data of zero-sequence voltage (ZSV) and zero-sequence currents (ZSC) are utilized to generate the ZSV-ZSCs images. Secondly, a FCGAN with overall evaluation ability is established to recognize the ZSV-ZSCs images across the whole waveform scale and segment the entire faulty-feeder ZSC. Thirdly, the segmented faulty-feeder ZSC is evaluated for waveform similarity and continuity, and the comprehensive evaluation metrics are calculated. Finally, a faulty-feeder detection criterion with high reliability is constructed. The performance of the proposed method has been verified using extensive simulation data and recorded data collected from various distribution systems. Experimental results demonstrate excellent detection performance in PSCAD simulation, achieving over 99.9% detection accuracy even under strong noise with 1 dB, and 100% detection accuracy in recorded data tests.