SUMMARY

Prior research has utilized data driven models to predict temperatures in data centers. This thesis considers and compares several data driven models for thermal predictions in data centers under steady state and transient scenarios. The steady state study compares six different data driven modeling techniques – linear regression, decision trees, artificial neural network (ANN), gaussian process regression (GPR), support vector regression (SVR), and relevance vector regression (RVR). While previous studies have looked at the application of some of these techniques to data center thermal predictions, this thesis compares several additional models. The effect of kernels on SVR and GPR models, and elastic net regularization – a methodology to select more important variables in given dataset, are also explored. It was found that GPR and ANN performed the best among the six data driven models studied. Furthermore, kernels were found to have a significant effect on the error for SVR, while kernels did not have a significant effect on error for GPR. The transient study compares long-short term memory (LSTM), and nonlinear autoregressive network with exogeneous inputs (NARX) in different scenarios. An ensemble was created to reduce the variance in LSTM and NARX predictions. This thesis also explores the possibility of utilizing relatively small datasets for transient temperature predictions in data centers, and how extrapolative input variables affect the prediction. Thermal prediction following computer room air conditioning unit (CRAH) failure and chiller failure scenarios are examined.

https://gatech.webex.com/gatech/j.php?MTID=m989319e3bc0af8b509e6ae4548c4b5f0