综合智慧能源2026，Vol.48Issue(8)：29-43,15.DOI:10.3969/j.issn.2097-0706.2026.08.003
基于优化双层分解与差异化建模的数据中心CPU利用率预测
CPU utilization prediction for data centers based on optimized two-stage decomposition and differentiated modeling
摘要
Abstract
The central processing unit(CPU)utilization series in data centers has the characteristics of nonlinearity,non-stationarity,multi-timescale coupling,and significant high-frequency disturbance.In view of these characteristics,a combined prediction method based on optimized two-stage decomposition,feature selection,and differentiated modeling across frequency bands was proposed.The CPU utilization time-series data of the Google data center from 2021 were selected as the research object.A total of 28 meteorological and environmental features consistent with the timestamps were configured.Supervised learning samples were constructed using a 24 h sliding time window,and data from March,June,October,and December were divided chronologically into training,validation,and test sets.The dingo optimization algorithm(DOA)was used to optimize the parameters of improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN),thereby achieving the primary decomposition of the original series.Based on the distribution characteristics of the zero-crossing rates of the intrinsic mode function(IMF)components,these components were reconstructed into high-,medium-,and low-frequency components.Furthermore,the football team training algorithm(FTTA)was used to optimize the parameters of variational mode decomposition(VMD),and secondary decomposition was performed on the high-frequency components to extract fine-grained fluctuation features.Subsequently,a two-stage approach combining light gradient boosting machine(LightGBM)feature importance ranking with validation-set coefficient of determination curves was used to determine the optimal input feature set for each subsequence.An adaptive kernel extreme learning machine was used to predict the medium-and low-frequency components with relatively stable variations and strong trends,whereas a combined bidirectional temporal convolutional network-bidirectional long short-term memory network-multi-head self-attention model was used to predict the high-frequency components with intense fluctuations and complex temporal relationships.The prediction results of all subsequences were superimposed and reconstructed to obtain the final CPU utilization predicted values.Case study results showed that after two-stage feature selection,the mean absolute error and root mean square error for the June samples decreased from 0.026 0 and 0.032 5 to 0.019 9 and 0.025 8,respectively,while the coefficient of determination increased from 0.824 3 to 0.887 5.The proposed model exhibited high prediction accuracy and stability across different months,with the coefficients of determination reaching 0.963 1 and 0.903 3 in October and December,respectively.The results indicate that optimized two-stage decomposition can effectively reduce the complexity of the original series,while differentiated modeling can fully leverage the advantages of different prediction models across frequency bands,providing a reference for data center resource allocation,operational status assessment,and energy consumption analysis.关键词
数据中心/CPU利用率预测/优化双层分解/特征筛选/差异化建模/算电协同Key words
data centers/CPU utilization prediction/optimized two-stage decomposition/feature selection/differentiated modeling/computing-electricity coordination分类
能源科技引用本文复制引用
朱虹谕,宋凯宇,武晓栋,黄亚松,刘一宁,张冬冬,张杰..基于优化双层分解与差异化建模的数据中心CPU利用率预测[J].综合智慧能源,2026,48(8):29-43,15.基金项目
国家重点研发计划项目(2026YFE0211200) （2026YFE0211200）
国家自然科学基金项目(52677064) （52677064）
内蒙古自治区自然科学基金项目(2026QC0364,2026QC0321)National Key R&D Program Project(2026YFE0211200) （2026QC0364,2026QC0321）
National Natural Science Foundation of China(52677064) （52677064）
Inner Mongolia Natural Science Foundation(2026QC0364,2026QC0321) （2026QC0364,2026QC0321）
