Fault localization and identification method based on dual-stream multi-task network

• Proposes a dual-stream network fusing temporal and impedance-weighted spatial cues. • Training and evaluation includes high-impedance faults, yielding inherent robustness to such events. • Improves classification by 0.97% identification, and localization by 5.35% over classical model on IEEE-33. • Achieves high accuracy using measurement from only 16 buses. • Exhibits topology generalization and superior speed over metaheuristic methods. Traditional heuristic algorithms often struggle to distinguish fault types and locate fault positions when phasor measurement unit (PMU) coverage is sparse and data quality is poor; moreover, the volatility of distributed generation (DG) outputs further weakens their practicality. To overcome these issues, we propose a dual-stream multi-task network (DSMN) constructed using gated recurrent unit (GRU) and the improved graph neural network GraphSAGE names impedance-weighted GraphSAGE (IWSAGE). Firstly, the electrical signals obtained by the PMU are encoded as a spatio-temporal graph aligned with the physical topology of the distribution network. Then, the GRU and GraphSAGE are used to extract temporal and spatial features respectively, and a dual-stream scoring network (DSSN) that dynamically weights time- and frequency-domain features. Then, six experiments are conducted on the IEEE-33 system for validation. The results show that, DSMN improves localization and type identification by 43.39% and 41.08% over heuristic methods, and by 0.97% and 5.35% over a classical spatio-temporal graph convolutional network. By coupling impedance-weighted, physics-aligned aggregation with task-aware gating, it remains accurate under sparse/noisy PMU data, supports single-pass real-time inference, reduces reliance on dense instrumentation, and enables faster restoration in active distribution grids.