Real-Time Scheduling of Renewable Power Systems Through Planning-Based Reinforcement Learning

The large-scale integration of renewable energy still encounters significant challenges. One critical bottleneck is the dilemma between dependable forecasting duration and optimization speed. In the context of difficulties in extending reliable forecasting periods, we resolve the decision speed issues by proposing a planning-capable reinforcement learning (RL) based solution. We model two of the most important and timeconsuming optimization problems in power systems, optimal power flow and unit commitment, as a unified Markov Decision Process. A tree search mechanism, including a multi-step lookahead planning process, is employed to incorporate short-term forecasts and provide stable policy improvements. Experimental results exhibit that our method significantly reduces renewable curtailment by 79%, achieves a comparable performance to physics-based solvers in constraint satisfaction, and demonstrates a certain level of robustness to topological changes.