强化学习在自动发电控制中的研究进展与展望  

Research Progress and Prospects of Reinforcement Learning in Automatic Generation Control

作　　者：解立辉[1] 席磊 XIE Lihui;XI Lei(College of Electrical Engineering&New Energy,China Three Gorges Univ.,Yichang 443002,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002

出　　处：《三峡大学学报（自然科学版）》2023年第5期133-141,共9页Journal of China Three Gorges University：Natural Sciences

基　　金：国家自然科学基金项目(52277108)。

摘　　要：自动发电控制(automatic generation control,AGC)是维持系统频率和功率动态稳定、提高电网控制性能的有效手段.随着以新能源为主体的新型电力系统的快速发展,高比例大容量新能源的接入将带来强随机扰动,传统的AGC控制策略不足以应对上述的严峻形势,而强化学习具有自主学习并积累和应用相关知识等特点,被逐渐应用于AGC系统中.本文分3个方面分析了强化学习在AGC中的重要研究进展:在集中式AGC策略方面,分别介绍了基于单agent RL“控制”与按各机组可调容量裕度的比例进行“分配”(PROP)的策略,基于PI算法“控制”与单agent RL“分配”策略;在分布式AGC策略方面,分别分析了多agent RL“控制”与PROP的策略,基于PI算法“控制”与多agent RL“分配”策略;在分层分布式智能发电控制策略方面,介绍了基于多agent RL“控制”与“分配”分层分布式智能发电策略.最后,结合新型电力系统发展需求和强化学习在AGC中存在的问题,对强化学习在AGC领域未来的发展方向进行了分析与展望.Automatic generation control(AGC)is an effective mean to maintain the dynamic stability of the frequency and the power in system and improve the performance of grid control.With the rapid development of the new power system dominated by new energy,the access of new energy with high proportion and large capacity will bring about strong stochastic perturbation.The control strategy of the traditional AGC is not enough to cope with the above severe situation,and reinforcement learning(RL)has gradually been applied to the AGC system because it has the characteristics of learning independently and accumulating and applying the relevant knowledge.Three aspects are divided in this paper to analyze the important research progress of RL in AGC.In terms of centralized AGC strategies,the strategies based on single-agent RL“control”and“allocation”in proportion to the adjustable capacity margin of each unit(“PROP”),and the strategies based on PI algorithm“control”and single-agent RL“allocation”are introduced respectively.In terms of distributed AGC strategies,the strategies based on multi-agent RL“control”and PROP,and the strategies based on PI algorithm“control”and multi-agent RL“allocation”are analyzed respectively.In terms of the control strategies of hierarchical distributed intelligent power generation,the strategies of multi-agent RL-based“control”and“allocation”hierarchical distributed intelligent power generation are introduced.Finally,the future development direction of RL in the field of AGC is analyzed and prospected in light of the development needs of new power systems and the problems of RL in AGC.

关 键 词：强化学习 自动发电控制(AGC) 多智能体 协同控制 研究展望 

分 类 号：TM734[电气工程—电力系统及自动化]