一种基于能耗预测的IT设备自动配置方法  

IT Equipment Automatic Configuration Method Based on Energy Consumption Prediction

作　　者：陈晓江 黄宏聪 蔡学龙 丁博[2] CHEN Xiaojiang;HUANG Hongcong;CAI Xuelong;DING Bo(China Southern Power Grid Digital Grid Group Information and Telecommunication Technology Co.,Ltd.,Guangzhou 510663,China;School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]南方电网数字电网集团信息通信科技有限公司,广州510663 [2]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5期56-64,共9页Journal of Harbin University of Science and Technology

基　　金：南方电网公司2021年资源池建设项目(JY-JF-03-ZY-21-006-TQ-011);国家自然科学基金面上项目(61673142)。

摘　　要：IT设备是数据中心的最大耗电设施,现有的IT设备能耗预测方法一方面仅能获取特征间的时序依赖关系,无法挖掘特征间的空间依赖性;另一方面无法根据任务的类型,动态的预测能耗,导致对IT设备能耗预测的不准确。针对以上问题,提出了一种基于长短时记忆网络(long short-term memory network,LSTM)和图卷积神经网络(graph convolutional neural network,GCN)的IT设备能耗预测方法,首先使用LSTM捕捉IT设备能耗特征自身的时序依赖性,然后构建图结构,在通过GCN在图结构上挖掘特征间的空间依赖关系,并且阶段性地捕捉IT设备的动态能耗模式,接下来采用注意力模块根据特征重要性不同对特征加权学习,最终得出能耗预测结果。实验结果表明,本文提出的能耗预测模型的平均绝对百分比误差为1.48%,均方根误差为1.55,均优于现有方法。通过能耗预测结果可以有效的对IT设备进行配置和调度,实现了数据中心的节能减排。IT equipment is the largest electricity consumer in data centers.However,existing IT equipment energy consumption prediction methods can only capture temporal dependencies between features and cannot uncover spatial dependencies between features.Furthermore,these methods cannot dynamically predict energy consumption based on the type of task,which leads to inaccurate predictions.To address these problems,this paper proposes an IT equipment energy consumption prediction method based on Long Short-Term Memory Network(LSTM)and Graph Convolutional Neural Network(GCN).In this method,the LSTM is first used to capture the temporal dependencies of IT equipment energy consumption features.And then a graph structure is constructed and the spatial dependencies between features are uncovered through GCN,and periodically captures the dynamic energy consumption patterns of IT equipment.Next,an attention module is used to weight the features for different importance levels,and the final energy consumption prediction is obtained.Experimental results show that the proposed energy consumption prediction method achieves MAPE of 1.48%and RMSE of 1.55,which are much better than other existing methods.IT equipment can be configured and scheduled based on energy consumption prediction results,which can achieve energy saving and emission reduction in the data center.

关 键 词：能耗预测 IT设备自动配置 长短时记忆网络 图卷积网络 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论]