Regional-privacy-preserving operation of networked microgrids: Edge-cloud cooperative learning with differentiated policies

Privacy preservation and coordination of networked microgrids (NMGs) are conventionally contradictory objectives. To address this, this paper proposes a regional-privacy-preserving operation method for NMGs that collaboratively learns differentiated policy (DP) of each microgrid (MG) at the edge by using a designed federated deep reinforcement learning (FDRL) algorithm. In the proposed method, a scalable edge-cloud cooperative framework is designed to integrate multiple independently controlled regional MGs into the existing distribution network (DN) without affecting its operation model. With the proposed framework, MGs can collaboratively optimize the local operation costs and global DN voltage by the respective regional control agent which controls local distributed energy resources power based on the decentralized partially observable Markov decision process. The proposed framework models differentiated private neural network (NN) models for each MG agent at the edge to efficiently address diverse regional tasks, and models a global NN at the cloud server to achieve collaborative training. The differentiated local policy of each MG control agent is learned via edge computing with the proposed DP-FDRL algorithm, which solves different regional tasks, achieves global coordination, and avoids exchanging the raw energy data among different agents simultaneously. By only transiting the global model parameters during the coordinated training process, the private NN models of each agent at the edge are also preserved to the MGs locally. Numerical studies validate that the proposed framework can effectively handle the complex privacy-preserving NMGs coordinated operation problem with collaborative learning through the DP-FDRL algorithm.