Multi-agent deep reinforcement learning-based autonomous decision-making framework for community virtual power plants

Modern grids are facing a reduction of system inertia and primary frequency regulation capability due to the high penetration of distributed energy resources. In this paper, a decision-making framework is proposed to facilitate community-based virtual power plants (cVPPs) to promptly provide ancillary services to the grid. A non-convex cVPP decision-making model is established to optimize the operational plans of a cVPP’s internal energy resources and the bids it puts in a local energy market to minimize the cVPP’s operation cost. Bidding and management strategies will be automatically executed by solving each participating cVPP’s decision-making problem. Due to its nature of high complexity and intractability, the problem is transformed into a partially observable Markov game model and solved by a multi-agent actor transformer-based critic method. A shared transformer encoder is used in the critic network to extract more robust features from the cVPPs’ observations and actions. Numerical simulation demonstrates that the proposed method can effectively support cVPPs to autonomously generate energy bidding and management strategies without acquiring other cVPPs’ private information.