Safe Reinforcement Learning-Assisted Second-Order Cone Optimization for Coordinated Distribution Network Reconfiguration and Battery Energy Storage Dispatch in Renewable-Rich Distribution Systems

High photovoltaic (PV) penetration can produce voltage rise, reverse power flow, feeder congestion, and less effective use of battery energy storage in radial distribution networks. This study develops a hybrid safe reinforcement learning–second-order cone programming (Safe RL–SOCP) framework for coordinated distribution network reconfiguration (DNR) and BESS dispatch. The end-to-end workflow was implemented as in-house research code. MATPOWER and OpenDSS were used for network simulation and AC power-flow verification, while YALMIP/CVX with MOSEK supported the conic projection stage. A constrained proximal policy optimization policy generates candidate topology and BESS actions; branch-exchange repair restores radiality, the SOCP layer corrects the candidate operating point, and AC power flow provides the final acceptance check. The framework is assessed on modified IEEE 33-bus and IEEE 69-bus feeders at 30%, 60%, and 90% PV penetration. Across the implemented benchmarks, the proposed controller reports lower daily energy loss and reverse power flow, tighter voltage ranges, and shorter online decision times. Monte Carlo evaluation over 365 unseen operating days further shows lower mean loss and dispersion. Residual post-projection voltage-violation probabilities of 0.3–0.5%, however, confirm that conic feasibility must be followed by AC verification rather than treated as an unconditional AC-feasibility guarantee.