Reinforcement Learning-Based Dynamic Event-Triggered Control of Markov Jumping Power Systems Under DoS Attacks

This work addresses the reinforcement learning (RL)-based dynamic event-triggered control problem for Markov jumping power systems (MJPS) subject to transient line faults and denial-of-service (DoS) attacks. First, to conserve communication resources and control costs, a novel RL-based dynamic event-triggered protocol (DETP) is established, which models the control policy and the event-triggered error as zero-sum game players, with the aim of obtaining the worst event-triggered error to maximize the triggering interval. Meanwhile, from the viewpoint of the defender, we incorporate the impact of DoS attacks into the performance value function, which enhances the robustness of the learned control policy and ensures the system’s reliable operation in hostile environments. Then, to obtain the Nash equilibrium solution of the secure optimal control problem (OCP), a data-driven improved policy iteration (IPI) algorithm is introduced without relying on the system dynamics. By incorporating a constant sequence, the algorithm not only avoids to seek the initial stabilizing control policies (ISCPs) and improves learning efficiency through faster convergence compared to traditional PI and value iteration (VI) methods. Finally, this paper provides the theoretical guarantees on convergence and stability of the proposed algorithm, and simulation results also validate the effectiveness and superiority of the proposed approach.