深 度 强 化 学 习 在 电 网 实 时 计 划 编 排 中 的 应 用

In the face of the strong uncertainty of the new power system, the rapid growth of the control scale, the low-carbon operation target, and other changes, the real-time plan scheduling will present high-dimensional, nonlinear, and non-convex and complex characteristics. The data-driven algorithm represented by reinforcement learning brings new ideas to explore the fast optimization of real-time plan scheduling. In this paper, deep reinforcement learning is introduced into the real-time plan scheduling model, and a real-time plan scheduling simulation environment for reinforcement learning is constructed. Then, a real-time plan scheduling method based on double-layer multi-objective multi-agent deep reinforcement learning is proposed. Based on the idea of hybrid model-data-driven reinforcement learning, this method adopts double-layer architecture and multi-agent design to achieve the parallel and rapid scheduling of real-time plans. Finally, the effectiveness and feasibility of the proposed method are verified through an example.