Physics-Informed Deep Reinforcement Learning with Adaptive Action Masking for Secure Second-Level Emergency Scheduling in Energy Networks

Abstract

With the integration of a high proportion of volatile renewable energy into modern power grids, systems frequently face severe supply-demand imbalances and escalated situational risks under extreme events. Addressing the high-dimensional risks and stringent response timing requirements in real-time markets of new-type power systems, this paper proposes a Physics-Informed Deep Reinforcement Learning (PI-DRL) framework with adaptive action masking for secure energy network scheduling and emergency control. First, to mitigate the “curse of dimensionality” inherent in massive heterogeneous resources, a situational risk-driven dynamic feature selection and action masking technique is established to exponentially compress the agent’s exploration space, solving the convergence bottleneck of large-scale resource dispatch. Second, targeting low-probability, high-risk scenarios, a knowledge-driven physical security shield is constructed, utilizing local Jacobian sensitivity matrices to achieve near-real-time physical correction of actions. This overcomes the lack of boundary awareness in traditional black-box models. Furthermore, a Security Reward Shaping mechanism is introduced to foster endogenous security awareness within the agent through closed-loop training, ensuring millisecond-level decision safety during online deployment. Validation on a modified IEEE 118-bus system using high-fidelity meteorological big data from China demonstrates that the proposed method accelerates decision-making by 400-fold compared to traditional MILP algorithms, enabling second-level response. Compared to pure data-driven models, it reduces operational costs by 12.4% and ensures absolute compliance with physical constraints while suppressing risk spikes within 3 minutes during an extreme 450 MW power drop event.