Data center power consumption prediction based on principal component analysis and DeepAR

Abstract

The era of big data and cloud computing has driven the rapid expansion of the number and scale of data centers worldwide, and the ensuing huge power consumption has put pressure on resources and the environment. Accurate prediction of data center power consumption can provide an important basis for current power management techniques, while effectively improving the efficiency of intelligent operation and maintenance of modern data centers. To address this problem, a server power consumption prediction model based on a combination of principal component analysis (PCA) and DeepAR is proposed in the paper. The model uses the time series of server power consumption and performance index data from the Zhengzhou Inspur data center to predict future moment power consumption, performs principal component analysis on the performance index, and inputs the effective principal components and historical power consumption data into the DeepAR network for prediction. The model is experimentally validated on all three server datasets, and the results show that the model outperform the DeepAR network model as well as other comparison models in terms of prediction. When compared with the DeepAR network, the MAPE of this model is reduced by 0.23%, 0.12%, and 0.05% on the data1, data2, and data3 datasets, respectively.