Abstract:In order to solve the problem that a large number of distributed photovoltaic (PV) are connected to the distribution network, which leads to the increase of PV abandoned power due to the overlimit of node voltage, a two-agent deep reinforcement learning active distribution network optimization strategy was proposed. On the power grid side, the Double Deep Q-Network (DDQN) is used to train three kinds of agents that adjust the operation strategy of the equipment, namely capacitor bank, reactive power compensation device and photovoltaic inverter, aiming at the minimum voltage deviation of the distribution network and the minimum PV power discarded. On the load side, considering the large regulating capacity of the power changing station, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm agent is adopted according to the historical power changing data and operation income data of the power changing station. Make it in the training to obtain the maximum operating profit of charge and discharge experience. In the training process, the two agents realized information exchange through the charging and discharging power of the changing station, and each realized the optimization of the objective function. After training, the two agents are deployed in the distribution network dispatching center and the control center of the changing station respectively. According to the real-time operation data of the power grid, the action strategy of the equipment and the charging and discharging strategy of the changing station are adjusted online. Finally, a 10kV system is taken as an example to verify the effectiveness of the model.