Multi-agent safe reinforcement learning based distributed optimal dispatch for active distribution network with incomplete information

This article proposes a multi-agent safe reinforcement learning method for distributed optimal dispatch of the active distribution network (ADN). First, a multi-agent optimization framework has been constructed based on coalition game, which ensures algorithm convergence while protecting privacy by uploading non-critical information to the aggregator. Furthermore, a safe reinforcement learning based distributed optimization method has been proposed to train agents' policy functions. The proposed method constructs feasible regions for agents taking into account ADN constraints such as AC power flow equations, which improves convergence speed and ensures security of the real-time decisions. What's more, the game theory provides a theoretical basis for multi-agent collaborative optimization taking into account the conflicts of interest, which enhances the generality and interpretability of the proposed method. Finally, simulations have been devised on a modified IEEE-123 bus system to show the effectiveness of the proposed method.