基于软演员-评论家算法与逻辑型Benders分解相结合的风电不确定性下月度安全约束机组组合

Monthly security constrained unit commitment (M-SCUC) ensures the reliability and flexibility of power system operations under high penetration of renewable energy from a long-term perspective. Based on deep reinforcement learning, this paper proposes a Soft Actor-Critic combined with Logic-based Benders decomposition (LBD-SAC) algorithm to efficiently solve the M-SCUC. It addresses the high-dimensional, non-convex, and complex uncertainties in the M-SCUC model without any compactness approximations. In the LBD-SAC, a task-decomposed and optimization-assisted training mechanism is developed to ensure the operational constraints and to enhance convergence performance. The M-SCUC is decomposed into a master problem with a constrained Markov decision process for commitment decisions, and a subproblem with relaxed optimal power flow. The constraint violations from the master problem and the relaxation bounds derived from the subproblem are used as safe exploration costs for the agent in the decision space. By integrating an accelerated primal-dual optimization approach to efficiently minimize these costs, the algorithm explores the optimal solution while satisfying the feasible region of the M-SCUC problem. The efficacy of the proposed method is validated using the modified IEEE 118-bus, 300-bus and 500-bus test systems.