摘要
“双碳”目标加速新型电力系统的发展,因虚拟电厂(virtual power plants,VPP)能够整合分布式资源和储能设备,提升能源供应的灵活性和可靠性,成为新能源参与调频的重要手段。而虚拟电厂的电气特性不同于传统发电机组,导致控制区域间协作调频能力降低,使得控制性能(control performance standard,CPS)标准恶化。为此,本文将集成VPP的负荷频率控制(load frequency control,LFC)过程建模为部分可观测马尔可夫决策过程,并严格考虑了CPS标准下的LFC系统评价,设计了基于多智能体深度强化学习的LFC协作控制系统。同时,提出了改进价值分解网络的多智能体协作算法,解决懒惰智能体导致区域间无法协作控制的问题。通过对集成VPP的两区域LFC模型和某省区电网系统模型进行仿真,验证了价值分解网络在集成VPP的LFC系统的有效性,且与传统多智能体算法相比,具有更高的协同控制能力与性能指标。
The“dual carbon”goal accelerates the development of new power systems.Virtual power plants(VPPs)can integrate distributed resources and energy storage devices,enhance the flexibility and reliability of energy supply,so it becomes a crucial means for new energy to participate in frequency regulation.However,the electrical characteristics of VPPs differ from traditional generator sets,leading to reduced collaborative frequency regulation capabilities between control areas,which deteriorates the control performance standard(CPS).Therefore,this paper models the load frequency control(LFC)process of integrated VPPs as a partially observable Markov decision process,strictly considering the LFC system evaluation under CPS standards.It designs an LFC collaborative control system based on multi-agent deep reinforcement learning.Additionally,an improved multi-agent collaboration algorithm based on value decomposition networks is proposed to address the issue of lazy agents,which hampers collaborative control between regions.Simulations of a two-area LFC model integrated with VPPs and a provincial power grid system model verify the effectiveness of the value decomposition network in the LFC system integrated with VPPs.Compared to traditional multi-agent algorithms,it demonstrates higher collaborative control capabilities and performance indicators.
作者
刘蔚
张野
吴应双
杨子千
唐王倩云
刘明顺
王寅
LIU Wei;ZHANG Ye;WU Yingshuang;YANG Ziqian;TANG Wangqianyun;LIU Mingshun;WANG Yin(CSG Electric Power Research Institute,Guangzhou 510700,China;Guizhou Power Grid Co.,Ltd.,Guiyang 550002,China)
出处
《供用电》
北大核心
2024年第12期62-71,共10页
Distribution & Utilization
基金
国家自然科学基金青年科学基金项目(52207110)
中国南方电网有限责任公司科技项目(GZKJXM20222215)。
关键词
多智能体
自动发电控制
强化学习
协同控制
虚拟电厂
multi-agent
automatic generation control
reinforcement learning
collaborative control
virtual power plants
