Two-stage Deep Reinforcement Learning Approach for Dynamic Equivalence of Active Distribution Network

Conventional load modeling approaches struggle to accurately model active distribution network (ADN) with increased complexity and uncertainty due to their reliance on fixed structures and parameters. To address this issue, an equivalent ADN model considering dynamic behavior of motor loads and the static voltage profiles of EV (Electric Vehicle) loads is formulated. A two-stage deep reinforcement learning algorithm combining Double Deep Q network (DDQN) and Proximal Policy Optimization (PPO) is proposed. The DDQN agent handles the discrete action for motor selection, while the PPO agent addresses the continuous action for EV identification. Simulation results indicate that the proposed method effectively addresses the hybrid discrete-continuous equivalence problem and achieves high accuracy in both active and reactive power calculations.