Deep Reinforcement Learning for Intentional Controlled Islanding of Power Systems

Intentional controlled islanding (ICI) is often considered as a last resort to avoid power systems from a full blackout induced by large disturbances like cascading failures. This urgent task entails rapid and effective control in a few seconds, making most optimization algorithms intractable in efficiency due to the combinatorial and nonlinear nature of ICI. To overcome the problem, this article proposes a reinforcement learning (RL) method, named RL-Merge, to learn from the past and make real-time decisions, which can efficiently explore for quality solutions while guaranteeing the “N-island” constraint in ICI problems. Concretely, a “merge” strategy is proposed to modify existing solutions by “merging a bus into its opposite island” at a time and guarantee the constraint via a designed processing rule following a merging action. It allows the RL agent to be integrated into any warm-start solver for performance enhancement. An attention-based modeling is designed for the RL agent to structurally handle the features from both problem parameters and objective values of the current solution, leading to notably improved generalizability to unseen scenarios. Comparative case studies with various benchmarks are conducted, validating that RL-Merge is a more capable and efficient solver for both planning and real-time control tasks.