Comparison of data driven modeling approaches for temperature prediction in data centers

Abstract

Energy-efficient thermal management of data centers based on dynamic optimization and provisioning of cooling resources requires rapid (nearly real-time) predictions of temperatures within data centers. This work for the first time compares multiple Data-Driven Models (DDMs) to achieve such rapid temperature predictions. DDM typically employs statistical or machine learning-based tools, in combination with physics-based modeling and/or experimental data to predict system behavior. In general, DDM approaches are well-suited to systems that have multiple operational states based on interactions between the many electrical, mechanical and control parameters typical of data centers.This study compares the performance of three different DDM methods, namely Artificial Neural Networks (ANN), Support Vector Regression (SVR), Gaussian Process Regression (GPR) in predicting both steady-state and transient rack inlet air temperature distributions in data centers. Additionally, Proper Orthogonal Decomposition (POD) was considered for transient modeling. The data used for training and analysis were obtained by performing 300 offline numerical simulations with a room-level, experimentally validated computational fluid dynamics/heat transfer (CFD/HT) model.The performance of the four data-driven models was evaluated based on the absolute mean error for interpolation and extrapolation, and the adaptability of the models to changes in physical domain (data center room) configuration. Additionally, the impact of the size of the training data set on prediction accuracy is also compared for the four models.For the steady-state case study, the predictions for ANN, SVR and GPR models are in good agreement with CFD/HT simulations, with the GPR model having the smallest overall average prediction error of 0.6 °C in rack inlet air temperature, corresponding to a relative error of 2.7% with respect to rack inlet temperature measured in °C. It was found that for all the frameworks the prediction error increases when the size of training data set was less than 300 samples. The GPR model had the best accuracy for smaller training data sets compared with the other models, with an average prediction error for rack inlet temperatures <1 °C when trained on only 50 simulations. For the transient case study, the interpolative prediction error for all the models is very low (<0.3 °C); however, the extrapolative prediction errors are much greater, and appear to be directly proportional to the (here, temporal) “distance” from the interrogation point to the input parameter space.