Multi-Agent Safe Policy Learning for Power Management of Networked Microgrids

This article presents a supervised multi-agent safe policy learning (SMAS-PL) method for optimal power management of networked microgrids (MGs) in distribution systems. While unconstrained reinforcement (RL) algorithms are black-box decision models that could fail to satisfy grid operational constraints, our proposed considers AC flow equations and other limits. Accordingly, the training process employs gradient information constraints ensure control functions generate feasible decisions. Furthermore, we have developed distributed consensus-based optimization approach train agents' while maintaining MGs' privacy data ownership boundaries. After training, learned can be safely used by MGs dispatch their local resources, without need solve complex problem from scratch. Numerical experiments been devised verify performance method.