Multi-Agent Safe Reinforcement Learning Based Real-Time Volt/Var Optimization for Modern Distribution Networks

Increasing photovoltaic (PV) power generation may lead to fast voltage fluctuations, and the Volt/Var optimization (VVO) technique can be used to regulate the voltage profile. In this paper, a real-time data-driven safe VVO framework is proposed for modern distribution networks (MDNs), where a MDN is divided into multiple regions, and a decentralized deep reinforcement learning (DRL) is used to achieve the coordinated control of reactive power within each region for the voltage profile regulation and power loss minimization. The VVO is formulated as a safe partially observable Markov decision process. To provide a safe operational zone, a penalty-based reward function is formulated. A modified transition probability is then integrated into the proposed safe DRL-based VVO framework. This proposed framework can be used to regulate PV power generation within a 3-minute time resolution, and is validated through the IEEE 33-bus test system by using actual load and PV power generation data, showing superior performance comparing to three state-of-the-art DRL-based VVO frameworks.