Protecting privacy in microgrids using federated learning and deep reinforcement learning

This paper aims to improve the energy management efficiency of home microgrids while preserving privacy. The proposed microgrid model includes energy storage systems, PV panels, loads, and the connection to the main grid. A federated multi-objective deep reinforcement learning architecture with Pareto fronts is proposed for total carbon emission and electricity bills optimization. The privacy of data is protected by federated learning, by which the original data will not be uploaded to the server. Numerical results show that compared with the traditional single Deep-Q network, using the proposed method the accumulated carbon emission decreased by 3% and the electricity bills decreased by 21%.