Fault Location Method Based on CNN-BiLSTM-Attention for locating Single-phase Ground Fault in Active distribution

The increasing integration of distributed generation (DG) into the distribution network has led to greater complexity in power flow distribution and transient current characteristics during single-phase grounding faults. As a result, traditional fault location methods are no longer sufficient. Adapting existing fault location methods to accommodate changing permeability has become an urgent issue. In response, a fault location method utilizing a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) network is proposed. The CNN is used to extract detailed longitudinal features from fault zero sequence current data at a specific time, compressing the data length to reduce subsequent network training parameters. Additionally, a cascade network with BiLSTM as its core is constructed to capture historical horizontal features of fault data during the fault evolution process. An Attention mechanism is integrated to ensure that the model focuses on changes in fault time and location data, thereby enhancing fault location accuracy. The simulation results demonstrate that the proposed method is capable of accurately identifying single-phase grounding faults, offering high precision and robustness in positioning, and exhibiting strong adaptability across various permeability fault scenarios.