A Two-Stage Reinforcement Learning Approach for Unit Commitment Considering Photovoltaic Power Forecasting Errors

This paper proposes a reinforcement learning solution for the unit commitment optimization problem in power systems with photovoltaic generation accounting for PV output prediction errors. First, we utilize the Markov Decision Process to transform unit commitment model into a dynamic optimization problem within a reinforcement learning framework. The approach thoroughly considers key features of PV prediction errors, transmission line power constraints, and thermal generator constraints to establish a reinforcement learning environment that minimizes total power supply costs. Second, regarding optimization algorithms, we employ the Actor-Critic algorithm to achieve rapid solutions for both discrete unit commitment decisions and continuous power output optimization. Finally, we verify our method using a 6-unit 24-hour case study, demonstrating that this approach significantly improves decision-making efficiency for unit commitment problems scenarios compared to traditional mathematical programming methods, while offering strong reusability.