Data-driven energy sharing for multi-microgrids with building prosumers: A hybrid learning approach

The real-time optimal scheduling of distributed energy resources (DERs) in interconnected multiple microgrids (MMGs) is facing great challenges due to the uncertainty of renewables, non-linear network constraints, the involvement of multi-level interest entities etc. Here, a data driven hybrid learning approach is proposed for real-time hierarchical energy sharing for MMGs with building prosumers. First, a data-driven XGBoost-based supervised learning model is established to characterize price-based demand response behaviours of prosumers for online P2P energy sharing results estimation among prosumers. Moreover, a multi-agent deep reinforcement learning (MADRL) method is developed for the energy sharing among MMGs and multi-agent deep deterministic policy gradient (MADDPG) algorithm is adopted to solve the optimization problem through centralized training and decentralized implementation. Particularly, the XGBoost-based demand response model of prosumers is embedded into the MADRL environment so that a balanced optimization strategy can be learned through the continuous interaction between MMGs agents and the environment. Finally, the effectiveness of the proposed method is demonstrated by a case study simulation with an artificial intelligence experimental platform.