Real-time Energy Management of a Low-Carbon Micro-Grid Based on Deep Reinforcement Learning

To minimize total operation cost considering carbon emission, the energy management problem of a micro-grid is formulated by scheduling conventional generators (CGs), energy storage systems (ESSs) and power trading with the main grid. It is proposed in this paper to convert the energy management problem to a Markov Decision Process (MDP) and solve it with soft actor-critic (SAC), a cutting-edge deep reinforcement learning (DRL) algorithm. The agent is able to capture uncertainty of load and renewable generation (RG) from large volumes of historical data in training and able to provide scheduling plan from real-time measurement of uncertainty once training is completed. Simulation results demonstrate the effectiveness of the proposed method and its advantage over deep deterministic policy gradient (DDPG) in both operation cost and constraint violation.