Reducing PV Curtailment and Battery Degradation Using Multi-Agent Deep Reinforcement Learning-Based Volt-VAR Optimization

The integration of rooftop photovoltaic (PV) in unbalanced distribution feeders often triggers overvoltage and revenue loss due to PV curtailment. This paper presents a PV-curtailment-aware multi-agent deep reinforcement learning (MADRL), applied to Volt-VAR Optimization (VVO) framework with centralized training and decentralized execution (CTDE). The agents coordinate smart-inverter Volt-VAR/Volt-Watt modes, on-load tap changers, switchable capacitor banks, and battery energy storage systems (BESS), while an online feasibility layer projects joint actions onto device/network-feasible sets. The reward internalizes the monetary price of curtailed PV energy together with BESS degradation, feeder losses, and limited switching, directly aligning learning signals with curtailment reduction and secure voltages. Realistic PV/load scenarios on the modified IEEE 13-bus feeder was implemented in Python 3.10 with PyTorch 2.1 and OpenDSSDirect. It was found that the proposed method reduces PV curtailment cost more than half, BESS-degradation cost by one-quarter, and active-power losses by one-eight compared with a DRL-based VVO baseline, while eliminating aggregated voltage violations over 50 days. These results indicate that cost-aligned rewards and feasibility-aware execution enable robust, deployable voltage regulation with minimal PV curtailment.