Faulty feeder detection based on fully convolutional network and fault trust degree estimation in distribution networks

Faulty feeder detection is crucial to maintaining the safety and stable operation of power grids after single line to-ground (SLG) faults occur in distribution networks. Existing detection methods have achieved good performance based on multi-feature fusion, but the extracted features commonly lack completeness and the detection process is poor in interpretability. This paper proposes a detection method based on fully convolutional network and fault trust degree estimation, which seeks to enhance its completeness and interpretability by mining the clear physical meaning from raw fault signals. Firstly, the zero-sequence current waveforms of all feeders were superimposed in the same plot to acquire the overall evaluation of fault conditions. Subsequently, a fully convolutional network was established to segment the waveforms of faulty feeder, suspected faulty feeder, and healthy feeders from the superimposed waveforms. Secondly, the segmented waveform of faulty feeder was compared with raw current waveforms, and an estimation method was introduced to quantitatively describe the fault trust degree of each feeder on the waveform scale, which is clear, complete, and intuitive. Finally, the faulty feeder can be detected after fault trust degree estimation. Various tests using PSCAD simulation, RTDS, field test, and practical fault data validate the efficiency of the proposed method.

Automated summary

This paper proposes a detection method for faulty feeder based on fully convolutional network and fault trust degree estimation. By mining the clear physical meaning from raw fault signals, the completeness and interpretability of the detection method are enhanced. Various tests validate the efficiency of the proposed method.