An interpretable framework for power grid dispatch risk identification integrating domain knowledge and reinforcement learning

Abstract

To address the limited interpretability and adaptability of conventional power grid dispatch risk identification, this paper proposes an interpretable framework integrating domain knowledge with reinforcement learning. Grid topology, operational constraints, dispatch rules, and risk propagation knowledge are embedded into state representation and policy evaluation to guide risk-sensitive learning. A knowledge-driven explanation mechanism traces risk sources, affected equipment, and potential consequences. Case studies show that the improved model reaches a mean Average Precision of ~0.72–0.73, while the optimal learning-rate configuration achieves 0.77 accuracy and 0.72 AUC (Area Under Curve), demonstrating effective and interpretable risk identification.