Multi-objective cooling control optimization for air-liquid cooled data centers using TCN-BiGRU-Attention-based thermal prediction models  

作　　者：Jianpeng Lin Wenjun Lin Weiwei Lin Tianyi Liu Jiangtao Wang Hongliang Jiang 

机构地区：[1]Department of Computer Science and Engineering,South China University of Technology,Guangzhou 510000,China [2]Peng Cheng Laboratory,Shenzhen 518066,China [3]Huawei Technologies Co.,Ltd.,Shenzhen 518066,China

出　　处：《Building Simulation》2024年第12期2145-2161,共17页建筑模拟(英文)

基　　金：supported by the National Natural Science Foundation of China(62072187);Guangdong Major Project of Basic and Applied Basic Research(2019B030302002);The Major Key Project of PCL under Grant PCL2023A09;uangzhou Development Zone Science and Technology(2023GH02).

摘　　要：With the breakthroughs in generative artificial intelligence(GAI)models,the vast computational demands are placing an unprecedented burden on the data center(DC)energy supply.The cooling system is the second major energy consumer in the DC,maintaining the safe and efficient operation of computing equipment.However,time-varying temperature gradients and power distribution pose a considerable challenge for efficient cooling management in DCs.For this problem,this work proposes a multi-objective cooling control optimization(MCCO)method to minimize cooling energy consumption while maximizing the rack cooling index(RCI)to ensure energy efficiency and security of hybrid-cooled DCs.The proposed method relies on high-fidelity models that characterize the dynamic thermal evolution and cooling power.Therefore,a novel network model(TCN-BiGRU-Attention)combining temporal convolutional network(TCN),bidirectional gated recurrent unit(BiGRU),and attention mechanism is designed to capture the features of multivariate time-series to predict temperature changes in thermal environments and cooling loops.Moreover,considering the complex heat transfer and operational characteristics of hybrid cooling systems,a machine learning(ML)-based power model is constructed to evaluate the holistic cooling power.Subsequently,the NSGA-II algorithm formulates the optimal cooling decision based on the predicted thermal distribution and cooling power,realizing the trade-off between energy consumption and cooling effectiveness.The results of numerical experiments using Marconi 100 data traces suggest that the proposed MCCO significantly reduces cooling energy consumption in summer and winter while maintaining the RCI above 95%.

关 键 词：data center thermal prediction cooling control multi-objective optimization energy saving 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]