Abstract Rapid and accurate fault localization and impact quantification are critical for preventing cascading failures in modern power grids. Traditional time-domain simulations suffer from high computational latency, while purely data-driven deep learning models lack physical interpretability and are prone to over-smoothing. This paper proposes a novel physics-informed dual-stage graph neural network (GNN) framework for real-time grid anomaly analysis. In the first stage, a degree-biased multi-head Graph Attention Network (GAT) is developed to pinpoint the fault source, effectively disentangling complex features under multiple concurrent faults. In the second stage, a Cascading Failure Propagation Model (CFPM) is explicitly embedded into the GNN’s message-passing mechanism as a physical regularization term. This physical prior forces the network to quantify the impact scope along actual power flow trajectories rather than mere topological hop counts. Extensive experiments on a hybrid dataset—comprising 15,000 simulated samples from the IEEE 118-bus system and 5,200 real-world SCADA/PMU records—demonstrate the framework’s superiority. The proposed method achieves a fault localization accuracy of 96.84%, outperforming traditional temporal models by 21%. Furthermore, the physics-informed regularization reduces the impact prediction root mean square error (RMSE) by 34% and delineates the impact boundary with an Intersection over Union (IoU) of 0.88. With an end-to-end inference time of merely 18 milliseconds on a provincial-level grid topology, the proposed framework provides a highly interpretable, ultra-low-latency solution for online grid dispatching and proactive defense.
Zhao et al. (Mon,) studied this question.