Evolutionary Multi-Agent Deep Meta Reinforcement Learning Method for Swarm Intelligence Energy Management of Isolated Multi-Area Microgrid With Internet of Things

Abstract

In an isolated multi-area microgrid, a conventional centralized active control policy relies on excessive communication and therefore is incapable of coordinating the interests of multiple operators. For this reason, this paper proposes a swarm intelligence load frequency control (SI-LFC) method. Based on the swarm intelligence method, the proposed method equates the units in each area as independent agents and adopts the swarm intelligence centralized offline learning policy to achieve the balance of interests of different operators. In online application, each unit only needs to collect the frequency locally to achieve global optimal control, thereby reducing the communication burden across the network. In addition, this paper proposes an evolutionary multi-agent deep meta-actor critic (EMA-DMAC) algorithm, which introduces meta-reinforcement learning and evolutionary learning to achieve fast collaborative learning of swarm agents, thereby improving the robustness and quality of the obtained SI-LFC strategies. The effectiveness of the proposed method is demonstrated in a simulation of the four-area LFC model for Sansha island in the China Southern Grid (CSG).