Real-time Power Dispatch Method for High-proportion Renewable Energy Power Systems Based on Two-stage Safe Reinforcement Learning

Modern power grids present characteristics of supply-demand uncertainties, increased dispatchable resources, and reduced safety margins. Deep reinforcement learning (DRL) methods, characterized by self-learning and self-optimization capabilities, offer a promising approach to optimizing and scheduling power systems. However, traditional RL often fail to adequately account for action safety and suffer from low training efficiency. To address these issues, this paper proposes a real-time power dispatch method tailored for high-penetration renewable energy resource (RES) systems. Firstly, a two-stage safe DRL (SDRL) framework is introduced, which comprises three components, an offline pre-training module, an offline SDRL module, and an online decision-making module. Next, the real-time power dispatch problem is modeled as a constrained Markov decision process (CMDP). The DDPG-Lagrangian algorithm is employed to derive the optimal strategy. Finally, the effectiveness of the proposed method is validated using an improved IEEE 39-Bus power system.