Voltage Control Based on Multi-Agent Safe Deep Reinforcement Learning

Abstract

To address issues of voltage limit violations and fluctuations caused by the high penetration of distributed photovoltaic (PV) systems in the distribution network, a voltage control method based on multi-agent safe deep reinforcement learning is proposed. The voltage control with PV is modeled as a decentralized partially observable Markov decision process. A safety layer is introduced in the deep policy network for agent design, while the voltage barrier function based on traditional optimization model voltage constraints is used in defining the agent reward function. Testing results on the IEEE 33-bus system demonstrate that the proposed method can generate voltage control strategies that meet safety constraints under high photovoltaic penetration scenarios, and it can be used to assist dispatchers in making real-time decisions online.