Lightweight model for power grid cascading failures risk evaluation based on graph physics-informed attention network

Abstract

In modern power grids, cascading failures pose an escalating threat to grid reliability, leading to the importance of predicting the likelihood of such failures. While existing power flow-based models rely on detailed physical dynamics, their computational latency hinders online applications. This study introduces a lightweight Graph Physics-Informed Attention Network (GPIAN), uniquely integrating power grid physical laws with graph neural network attention to address this gap. GPIAN replaces conventional attention mechanism with a complex network-based framework, where the Electric Functional Strength (EFS), a metric quantifying node interaction guided by power grid principles, drives adaptive information aggregation. This design not only reduces model parameters by 90.7% compared to standard graph attention network but also embeds physical interpretability, enabling the model to prioritize critical node-edge dependencies in cascading failure scenarios. Experimental validation across IEEE-39, IEEE-118, IEEE-300, and Italian power grids demonstrates that GPIAN achieves higher prediction accuracy than mainstream methods, while maintaining fast inference speeds suitable for real-time deployment. These results highlight how merging physical principles with data-driven learning can transform cascading failure prediction, offering a practical, interpretable tool for proactive grid management and significantly advancing the field’s capacity to mitigate blackout risks.