Physics-Informed Interpretable Graph Neural Networks for Real-Time Transient Stability Assessment and Preventive Control in Large-Scale Hybrid Power Systems

Abstract

As transmission systems incorporate gigawatt-scale renewable generation and inverter-based resources, the power system operating state space expands exponentially while system rotational inertia declines. Conventional time-domain simulation is too computationally slow for real-time transient stability assessment and emergency control. Conversely, standard deep learning models act as black-box classifiers that lack physical interpretability and fail to generalize when network topologies change due to unexpected transmission line outages. This paper introduces a physics-informed graph neural network architecture for real-time transient stability assessment and closed-loop preventive control in low-inertia hybrid power systems. The network embeds Kirchhoff current laws, synchronous generator swing dynamics, and nodal admittance matrices directly into the graph message-passing layers. We couple the stability classifier with a differentiable optimal power flow layer that computes preventive generator rescheduling and emergency energy storage active-power setpoints within sixteen milliseconds of fault clearance, with physical delivery of those setpoints then following the storage system's own ramp-rate limit. Tested on a modified one-hundred-eighteen-bus transmission network integrated with forty-five percent renewable penetration across fifty thousand topological fault scenarios, the proposed architecture achieves ninety-nine point four percent stability prediction accuracy, zero missed unstable alarms on the evaluated dataset, and reduces emergency control curtailment costs by thirty-four point two percent compared to a black-box convolutional neural network baseline.