Energy management system by deep reinforcement learning approach in a building microgrid

In this paper, we study the application of the deep reinforcement learning to train a real time energy management system using the DQN algorithm. We consider a building–scale microgrid with PV production, non-shiftable loads, a battery unit, and a unidirectional connection to the utility grid. The price of electricity follows peak / off-peak rates. The objective of the energy management system (EMS) is to minimize the operational cost of the microgrid without any forecaster, but based on past data. The EMS is designed to respond in real-time to the net energy demand of the microgrid and control the battery via a discrete set of actions. Numerical experiments are conducted and results show the efficiency of the training phase and the reliability and near-optimal performance of the trained agent