Safe deep reinforcement learning for Q/V droop curve of PV inverters towards distribution network voltage regulation

Abstract

The increasing penetration level of renewable energy sources in distribution networks can lead to rapid and severe voltage fluctuations and violations, which poses great challenges to traditional voltage regulation methods, particularly in real-time scenarios. Although PV inverters are advocated due to their high flexibility, the conventionally applied local Q-V droop strategy cannot achieve a system-wide optimal solution. To address these limitations, this report firstly establishes the PV inverters based decentralized voltage regulation framework where the whole network is divided into several sub-zones and each zone is regulated by an intelligent agent. The key innovation lies in the proposed Multi-Agent Safe Deep Reinforcement Learning (MASDRL) method, which optimizes the Q-V droop curves of PV inverters to simultaneously minimize network energy loss and voltage deviations across the entire system. A distinctive feature of the proposed approach is the integration of a safety layer, which ensures constraint compliance and enables safe exploration during the learning process, addressing critical gaps in existing methods. Finally, simulations on the IEEE 33-bus distribution system validate the effectiveness of the proposed method in enhancing the voltage regulation performance while improving constraint satisfaction.