Answers are grounded in this paper and cite their sources.
Power grids, as critical cyber-physical systems, face increasing threats from adversarial attacks that can compromise their operational integrity. This article introduces a novel cascading policy learning framework that leverages a sequential application of three deep reinforcement learning algorithms to bolster the power grid’s resilience. The framework implements a three-stage cascading approach: initially employing Proximal Policy Optimization to establish stable policy foundations, subsequently applying Trust Region Policy Optimization to ensure mathematically rigorous policy updates while maintaining performance bounds, and finally utilizing Advantage Actor-Critic to minimize policy gradient variance and optimize convergence. This sequential integration creates a robust control policy that progressively refines decision-making capabilities at each stage. Experimental validation in a simulated power grid environment evidences superior performance, with the framework achieving 84% success in maintaining continuous grid functionality for 24 h under bus-tripping attacks, exceeding baseline approaches by a considerable margin. Results confirm that this multi-stage learning strategy effectively amplifies convergence speed, maximizes cumulative rewards, and strengthens power grid resilience against cyber-physical threats.
No takes yet. Share an insight, caveat, or question.
Bensalah et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: