Coordinated Operation of Multi-microgrids Using Multi-agent Deep Reinforcement Learning

Optimal operation coordination between multiple microgrids is difficult because of the multi-directional flow of energy. Through multi-head attention mechanism-based multi-agent deep reinforcement learning algorithm, centralized training of all agents to develop coordinated control strategies is proposed. During the training process, each agent can pay attention to the specific information that is most relevant to its reward by using the multi-head attention mechanism. Upon completion of the training process, the proposed approach makes decision on the coordinated operation of multi-microgrid（MMG） system and carries out the decision in a fully decentralized manner. The simulation results show that with the data of each micro-gird（MG） protected, the proposed approach can enhance operational economy of the MMG system by effectively coordinating internal energy allocations within individual MGs and external multilateral energy interactions among interconnected MGs.