分布式数据中心制冷控制的深度强化学习算法  

Deep reinforcement learning algorithm for cooling control in distributed data center

作　　者：张立 陈丁 万剑雄 李雷孝 ZHANG Li;CHEN Ding;WAN Jian-xiong;LI Lei-xiao(Inner Mongolia Meteorological Information Center,Inner Mongolia Meteorological Service,Hohhot 010051,China;School of Data Science and Application,Inner Mongolia Autonomous Region Engineering and Technology Research Center of Big Data Based Software Service,Hohhot 100080,China)

机构地区：[1]内蒙古气象局内蒙古自治区气象信息中心,内蒙古呼和浩特010051 [2]内蒙古自治区基于大数据的软件服务工程技术研究中心数据科学与应用学院,内蒙古呼和浩特100080

出　　处：《计算机工程与设计》2024年第5期1595-1600,F0003,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61862048);内蒙古自治区重点研发与成果转化计划基金项目(2021CG0033、2022YFSJ0013);内蒙古自治区高等学校青年科技英才支持计划基金项目(NJYT22084);内蒙古自治区直属高校基本科研业务费基金项目(JY20220078)。

摘　　要：为对数据中心空调送风温度与送风速度进行智能控制,降低数据中心制冷能耗,提出一种基于多智能体强化学习的数据中心空调制冷控制方法。所有空调智能体的制冷策略联合训练使其能合作制冷,设置温度约束保证设备在安全温度下运行。在不断与环境模型交互的过程中获得经验,通过经验优化控制方法,以更加节能的方式保证数据中心的正常工作。仿真结果表明,与单智能体强化学习的CCA方法相比,所提方法在分布式数据中心环境中,降低了14.1%的空调功耗与18.09%的温度违约惩罚。To reduce the cooling energy consumption of the data center,a multi-agent reinforcement learning based air conditio-ning and cooling control method was proposed.The cooling strategies of all air conditioning agents were jointly trained for efficient cooperative cooling,and temperature constraints were set to ensure that the equipment operated at a safe temperature.The experience was gained from continuous interactions with environmental models and the control policy was optimized through experience to ensure the normal operation of the data center in a more energy-saving manner.The simulation results show that the proposed method reduces the air-conditioning power consumption by 14.1%and the temperature penalty by 18.09%in the distributed data center cooling control environment compared with the single agent CCA method.

关 键 词：数据中心 送风速度 送风温度 强化学习 多智能体环境 合作制冷 温度约束 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP308[自动化与计算机技术—控制科学与工程]