Reinforcement Learning Based Double-layer Optimization Method for Active Distribution Network with Soft Open Point

Due to the integration of a large number of distributed genreators to the distribution network, the insufficient regulation and control capacity of the distribution network has led to serious problems such as large voltage fluctuations, low consumption level of renewable energy, and poor economic indicators. Therefore, a collaborative optimization method of soft open point (SOP) and dynamic reconfiguration of the distribution network is proposed to fully mobilize the flexibility resources on the network side, and a double-layer optimization model based on reinforcement learning is constructed. In the upper layer of the model, the generated radial topological structure set is used as the action space of reinforcement learning, and the topological structure is dynamically selected. The lower layer optimizes the operation of controllable active equipment with SOP based on the network structure of the upper layer, excavates the flexible power flow regulation ability of SOP, and continuously modifies the topology selection of the upper layer according to the optimization results, which can effectively solve the multi-time-scale mixed-integer nonlinear programming problem. The convergence speed of the algorithm is accelerated by deleting invalid networks and taking priority sampling of historical information. The example analysis shows that the optimization strategy given by the double-layer collaborative optimization method based on reinforcement learning effectively improves the operation level of the active distribution network, and can adapt to the source-load uncertainty scenarios. © 2023 Automation of Electric Power Systems Press. All rights reserved.