A Distributed PV Voltage Control Method based on RL and Attention Mechanism

This paper investigates the limitations of traditional reactive voltage control methods in the context of distributed new energy integration, especially in scenarios involving reactive voltage control with photovoltaic inverters. Given the deficiencies in model dependency and solution speed, the paper adopts an interactive multi-agent deep reinforcement learning approach to learn reactive voltage control strategies for photovoltaics, making improvements and designs for different issues within the distribution network system and algorithm. Key contributions include: using a centralized training decentralized execution (CTDE) framework among multi-agents to optimize the balance between communication cost and control effectiveness; distinguishing between global rewards and local rewards to optimize the overall cost of the grid and voltage control at photovoltaic grid connection points, and enhancing the visualization and interpretability of the multi-agent algorithm through the introduction of an attention mechanism, which also visualizes the importance of contributions from different agents during the voltage optimization control process.