A Reinforcement Learning Scheduling Method for Distribution Network Photovoltaic Carrying Capacity for Seasonal Time-series Assessment

Abstract

High-dimensional continuous controls introduced by photovoltaic (PV) inverters, SVCs, flexible loads, and energy storage make reinforcement-learning scheduling difficult in PV-rich distribution networks. This paper proposes an electrical-structure-guided reinforcement-learning method to improve PV hosting capacity. A soft assignment matrix constructed from electrical distance, common upstream paths, and voltage-response sensitivity compresses nodal states and actions into a physically interpretable cluster-level latent space. Executable nodal controls are recovered through base decoding, residual correction, and safety execution, reducing the exploration burden while preserving local regulation capability. A seasonal time-series assessment is further developed by applying confidence-quantile statistics to the maximum admissible PV output at identical intraday instants across multiple days and converting the resulting accommodation trajectories into installable capacity. Tests on the IEEE 33-bus system show that the proposed method improves training efficiency and hosting capacity relative to PSO and BRD-RL, while providing quantitative seasonal capacity estimates for PV interconnection planning.