Power system corrective control considering topology adjustment: An evolution-enhanced reinforcement learning method

Effective corrective control is critical for maintaining secure power system operations. Conventional corrective control methods often struggle to balance computational speed with control optimality. In contrast, deep reinforcement learning facilitates faster control. However, significant uncertainties introduced by the high penetration can lead to repeated and irregular line overloads, impeding the accurate estimation of action values. To address these problems, this paper proposes an evolution-enhanced reinforcement learning method. The proposed method integrates the double dueling deep Q-network with the evolutionary algorithm, utilizing cumulative rewards to evaluate agents and reduce value estimation errors for corrective actions. Furthermore, we propose the top-K strategy to ensure that generated actions comply with complex operational constraints and the Monte Carlo evaluation method to enhance training stability. Case studies conducted on 36-node and 118-node power systems demonstrate that the proposed method can ensure stable system operations via topology adjustment, re-dispatching generators, and dispatching energy storage systems. It outperforms traditional corrective control methods and mainstream deep reinforcement learning algorithms. Results indicate that the proposed method extends the power system’s operating time by 14.94% to 22.81% and reduces computational costs. • Develops an evolution-enhanced RL method to correct renewable-induced line overloads. • A top-K strategy filters and ranks feasible actions to ensure operational security. • A framework with parallel simulations robustly evaluates actions under uncertainty.