计及隐私保护的多智能体深度强化学习有源配电网电压控制策略  

Multi-agent Deep Reinforcement Learning Voltage Control Strategy for Active Distribution Networks Considering Privacy Protection

作　　者：刘洋 伍双喜 朱誉 杨苹[2] 孙涛 LIU Yang;WU Shuangxi;ZHU Yu;YANG Ping;SUN Tao(Guangdong Power Dispatching Center of Guangdong Power Grid Company,Guangzhou 510000,China;Guangdong Key Laboratory of Clean Energy Technology,South China University of Technology,Guangzhou 510000,China)

机构地区：[1]广东电网有限责任公司电力调度控制中心,广东广州510000 [2]广东省绿色能源技术重点实验室(华南理工大学),广东广州510000

出　　处：《智慧电力》2024年第12期104-116,共13页Smart Power

基　　金：国家重点研发计划资助项目(2023YFB4203102);南方电网公司科技项目资助(GDKJXM20220335)。

摘　　要：随着“双碳”目标的推进和我国可再生能源规模的快速增长,配电网电压控制面临新的挑战。针对有源配电网分散式电压控制,提出了一种基于区域间隐私保护的多智能体深度强化学习算法,旨在解决集中训练阶段全局信息传播导致的隐私泄露问题并提升电压控制性能。首先,基于多智能体深度强化学习的特点,构建了隐私保护下的多智能体协同控制框架;然后,提出了一种结合局部观测与全局目标的分散式强化学习算法,用于优化电压调控设备的协调控制;最后,通过算例验证表明,所提出的方法能够有效提升配电网电压的稳定性与安全性,并在保证隐私的前提下实现高效的电压控制。With the implementation of China’s“dual carbon”goals and the rapid expansion of renewable energy,voltage control in distribution network faces new challenges.This study proposes a multi-agent deep reinforcement learning algorithm with inter-regional privacy protection to address decentralized voltage control in active distribution network,aiming to prevent privacy leakage from global information sharing during centralized training while improving voltage control performance.Firstly,a privacy-preserving collaborative control framework is developed,leveraging the principles of the multi-agent deep reinforcement learning.Then a decentralized reinforcement learning algorithm integrating local observation with global objective is established to optimize the coordinated control of voltage regulation device.Finally,case studies confirm that the proposed method improves voltage stability and security in the distribution network,achieving the efficient voltage control while preserving the privacy.

关 键 词：电压控制 隐私保护 分散式控制 多智能体深度强化学习 

分 类 号：TM731[电气工程—电力系统及自动化]