Spatio-Temporal Power Flow Forecasting During Cascading Failure Propagation in Power Systems

Accurate power flow forecasting is crucial for enabling timely interventions to prevent cascading failure events from escalating into large-scale outages. In this paper, we propose a spatio-temporal learning model specifically designed for forecasting power flow redistribution during cascading failure events. The model combines a transformer-based temporal encoder with a graph attention network (GAT) to jointly capture the temporal evolution and spatial dependencies within the grid, thereby enhancing robustness under rapidly changing operating conditions. The proposed model is evaluated on the IEEE 118-bus system and compared against several state-of-the-art baselines, demonstrating superior forecasting accuracy across both short-and long-term horizons.