Multi-timescale Deep Reinforcement Learning for Reactive Power Optimization of Distribution Network

With the access of high proportion distributed generation, distributed network faces tremendous challenges in dealing with uncertainty and coordinating a variety of reactive power compensation equipment. This paper presented a multi-timescale voltage regulation strategy based on a mathematical optimization model and data-driven method. First, for the online tap changer and switching capacitor with slow-timescale regulation, Aiming to minimize active power loss, the day ahead reactive power and voltage optimization model were proposed based on mixed-integer second-order cone programming. Second, to meet the real-time requirements on the fast timescale stage, an intraday real-time scheduling method based on multi-agent reinforcement learning was proposed, transforming the real-time reactive power optimization problem into a Markov game process and adopting a centralized training and decentralized execution framework. Compared with traditional methods, this method has low communication cost, better real-time performance, and does not rely on an accurate power flow model. Finally, the effectiveness of the proposed strategy is verified by an IEEE 33-bus example. © 2022 Chin.Soc.for Elec.Eng.