基于GRU的数据中心温度预测  

Temperature Prediction Model for Power Data Centers Based on GRU

作　　者：王旭 托娅 李鑫[3] 王芳 李晨霞 丁博[4] WANG Xu;TUO Ya;LI Xin;WANG Fang;LI Chenxia;DING Bo(Inner Mongolia Electric Power Economic and Technical Research Institute,Hohhot 010010,China;Inner Mongolia Electric Power(Group)Co.,Ltd.,Hohhot 010010,China;College of Energy and Power Engineering,Inner Mongolia University of Technology,Hohhot 010051,China;School of Computer Science and Technology,Harbin University of Science and Technology Harbin 150080,China)

机构地区：[1]内蒙古电力经济技术研究院,呼和浩特010010 [2]内蒙古电力(集团)有限责任公司,呼和浩特010010 [3]内蒙古工业大学能源与动力工程学院,呼和浩特010051 [4]哈尔滨理工大学计算机科学与技术学院,哈尔滨150080

出　　处：《哈尔滨理工大学学报》2024年第6期101-111,共11页Journal of Harbin University of Science and Technology

基　　金：内蒙古电力(集团)有限责任公司科技项目(2023-5-40);国家自然科学基金(61673142).

摘　　要：随着数据中心规模的不断扩大,电力的消耗也急剧增加,降低数据中心的能耗变得越来越重要,因此预测机房实时的温度变化,提前调整制冷系统控制策略,对减少数据中心的总能耗、提高能源效率具有重要意义。针对此问题提出了一种基于门控循环单元(GRU)的电力数据中心温度预测模型,并与卷积神经网络(CNN)和注意力机制(Attention)相结合,该方法以内蒙古某数据中心机房的历史温度数据作为输入,使用卷积神经网络提取高维特征,并加入了数据空间位置上的特征,将提取到的特征向量以时序特征集输入到GRU中进行特征提取和分析。同时利用注意力机制计算GRU隐藏层的注意力得分,将传给Attention的数据分为多个块,分别对块内和块间应用注意力机制,并分配注意力权重,突出高相关特征的影响,加快计算速度,最后完成温度预测。以内蒙古某数据中心提供的温度数据集开展实验,并与其他神经网络模型进行对比,实验结果表明本文所提方法具有更好的收敛性和更高的准确性。With the rapid expansion of data center infrastructure,the energy consumption has become a major concern.Reducing data center power consumption and improving energy efficiency are crucial.This paper proposes a power-efficient data center temperature prediction model based on Gated Recurrent Unit(GRU)integrated with Convolutional Neural Network(CNN)and Attention Mechanism.The model leverages historical temperature data from a data center in Inner Mongolia as input features.High-dimensional features are extracted using a one-dimensional CNN,and the extracted feature vectors are fed into the GRU for analysis and feature extraction.Additionally,the Attention Mechanism is employed to compute attention scores for the hidden layers of GRU and allocate attention weights,highlighting the influence of highly correlated features.The model ultimately achieves temperature prediction.Experimental evaluations are conducted using a temperature dataset provided by the Inner Mongolia data center,and comparisons with other neural network models are made.The results demonstrate that the proposed model exhibits better convergence and higher accuracy.

关 键 词：数据中心 温度预测 门控循环单元 卷积神经网络 注意力机制 

分 类 号：TP308[自动化与计算机技术—计算机系统结构]