数据中心冷热电联产系统的前摄式智能节能优化算法  

Proactive intelligent energy-saving optimization algorithm for data center CCHP system

作　　者：李庆华 冉泳屹 刘启晨 孙彤瑶 陈双武 雒江涛[1] LI Qinghua;RAN Yongyi;LIU Qichen;SUN Tongyao;CHEN Shuangwu;LUO Jiangtao(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Architecture,Planning and Landscape,Newcastle University,Newcastle NE17RU,England;School of Information Science and Technology,University of Science and Technology of China,Hefei 230027,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]纽卡斯尔大学建筑、规划与景观学院,英格兰纽卡斯尔NE17RU [3]中国科学技术大学信息科学技术学院,安徽合肥230027

出　　处：《智能系统学报》2025年第1期139-149,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(U23A20275,62101525,62171072);重庆市自然科学基金项目(cstc2021jcyj-msxmX0586).

摘　　要：现有的数据中心节能降碳优化方法没有综合考虑碳足迹涉及的能源输入、生产耗能以及废余利用等环节的耦合性,难以实现系统性节能降碳。为此,提出了一种基于深度强化学习的优化算法DeepCCHP(deep combined cooling,heating and power generation),针对数据中心冷热电联产系统,联合控制供电子系统和制冷子系统,优化用电成本、碳排放量和能效。DeepCCHP结合长、短期时间序列网络和深度强化学习方法对联合优化问题进行求解,实现前摄式的联合控制发电设备和制冷设备。在基于Trnsys软件的仿真环境中,通过阿里巴巴数据中心集群数据的训练和验证。实验结果表明,与基准算法相比,DeepCCHP算法可以节省最高40%的成本和28%的碳排放量,且能够在能源成本、碳排放和能效三者之间取得更好的折中与平衡。The existing methods for energy-saving and carbon reduction optimization in data centers lack a comprehensive consideration of the coupling of carbon footprint-related factors,including energy input,production consumption,and waste utilization.This limitation hinders the achievement of systematic energy-saving and carbon reduction.To address this issue,a deep reinforcement learning-based optimization algorithm,named DeepCCHP,is proposed.This algorithm focuses on the combined cooling,heating and power generation(CCHP)in data centers,employing coordinated control of the power supply and cooling systems to optimize electricity cost,carbon emissions,and PUE.Deep-CCHP integrates a long and short-term time-series network-attention(LSTNet-Attn)for multi-dimensional time series forecasting and a deep reinforcement learning approach to solve the joint optimization problem,achieving proactive joint control of power generation and cooling equipment.The algorithm is validated through training and verification using Alibaba data center cluster data in a simulation environment based on Trnsys software.Experimental results demonstrate that,compared with baseline algorithms,the DeepCCHP algorithm can achieve up to 40%cost savings and 28%reduction in carbon emissions.It also demonstrates a better trade-off and balance among energy cost,carbon emissions,and energy efficiency.

关 键 词：绿色数据中心 冷热电联产 智能节能 深度强化学习 碳排放优化 能效提升 联合控制 预测网络 

分 类 号：TP272[自动化与计算机技术—检测技术与自动化装置]