Predicting Voltage Stability Using Graph Convolutional Networks

Real‐time voltage stability assessment is critical for secure power grid operation, yet traditional simulation‐based methods suffer from high computational latency. While machine learning offers a promising alternative, standard models often fail to leverage the inherent topology and physics of the power grid. This paper addresses this gap by proposing a novel graph deep learning framework to predict the fast voltage stability index (FVSI). The key technical contributions of this work are threefold: (1) the introduction of a novel physics‐informed graph attention network (PI‐GAT), where the attention mechanism is explicitly guided by physical line impedance to enhance model interpretability and accuracy; (2) the first formulation of FVSI prediction as a spatiotemporal graph learning problem, accompanied by a comprehensive benchmark of architectures, including temporal GCN (TGCN); and (3) the integration of a dropout‐based framework to provide practical, real‐time uncertainty quantification for predictions. Evaluations on the IEEE 14‐bus system demonstrate that the proposed graph‐based models are significantly more accurate and robust than classical baselines. This framework offers a scalable, interpretable, and uncertainty‐aware solution for real‐time voltage monitoring.