FAULT IDENTIFICATION AND LOCALIZATION IN DISTRIBUTION GRIDS BASED ON AN ATTENTION-HYBRID GRAPH NEURAL NETWORK

This thesis proposes a multi-task fault diagnosis framework for distribution systems based on Graph Convolutional Networks (GCN) and an enhanced Graph Attention Network (GATv2). By representing the power grid as a graph with electrical features and topological connections, the model simultaneously performs fault type classification and fault location prediction. The architecture incorporates residual connections, multi-head attention, and a Jumping Knowledge module to capture multi-scale structural patterns, while dynamic loss weighting ensures balanced task optimization under noise and sparsity. Experimental results on the IEEE 123-node test feeder demonstrate a fault classification accuracy of 96.43%, and fault localization accuracies of 84.64% (strict), 96.83% (1-hop), and 98.49% (2-hop). The model exhibits strong robustness and offers a potential solution for intelligent fault diagnosis in real-world power systems.