Faulty feeder identification method for active distribution network based on depth feature extraction and semi-supervision domain adaptation

Faulty feeder identification is a key technology for active distribution network, and the deep learning-based method have attracted great attention in the field of fault diagnosis. However, many challenges are still exiting, including complex working conditions, insufficient valid data samples and practical scenario verification. Therefore, a deep transfer learning-based faulty feeder identification method is proposed in this work by fusing convolutional neural network, attention mechanism and semi-supervision domain adaptation. Firstly, a depth feature extraction model integrating with temporal-spatial attention mechanism is designed to achieve both local and global fault feature enhancement of zero-sequence current. Secondly, adaptive clustering loss is proposed to realize the alignment between the simulation data and the actual data. Furthermore, the pseudo-label loss is applied to the unlabeled samples, and the pseudo-label is retained with high confidence, so as to promote the cluster classification results. Finally, the proposed method is verified by building distribution networks on the hardware-in-loop and filed test platform. Its identification ability is demonstrated through different fault scenarios, and the performance compared with other methods.