A DRL-based optimization method for microgrid operation

To address the challenges of source-load uncertainty and insufficient scheduling flexibility in microgrids， an optimization method for microgrid operation based on deep reinforcement learning （DRL） is proposed. First， a microgrid model comprising photovoltaic （PV）， energy storage， and generation equipment is constructed， along with its constraint conditions. Second， a multi-objective optimization framework is established to minimize operating costs and imbalance of the system， considering uncertainties such as PV generation， load demand， and electricity prices. The twin delayed deep deterministic policy gradient （TD3） algorithm is employed to derive microgrid scheduling strategies in a data-driven manner. Third， a penalty term for high-proportion erroneous actions is incorporated into the reward function to constrain the output of each device within a reasonable range， mitigating the risk of insufficient safety guarantees inherent in reinforcement learning methods. Finally， simulation results demonstrate that， compared to the deep deterministic policy gradient （DDPG） algorithm， the proposed method achieves superior economic efficiency and stability， with economic costs closer to those of ideal deterministic optimization methods.