Data Driven Modeling Advancements for Thermal Predictions in Data Center Applications

Thermal predictions in data centers have been utilized to reduce the electric consumption of thermal equipment in data centers. While most of the optimization of data center temperature has been performed through the utilization of Computational Fluid Dynamics (CFD) and heuristic methods, data driven modeling techniques are now also being used to optimize the data center temperatures. Some data driven models have been used on a static data set to obtain the steady state temperature predictions for given input variables while other data driven models have been trained to provide temperature predictions at live time. This paper aims to investigate the transient temperature prediction capabilities of two data driven models — Long-Short Term Memory (LSTM) and Nonlinear Autoregressive Neural Network with External Input (NARX). While these two methods have been previously studied on data center applications, they have not been compared with each other for transient temperature predictions for normal operations. The study also utilizes ensembles to provide better temperature prediction accuracy for smaller data sets. The study compared these two models based on an experimentally obtained data set and found that NARX outperforms LSTM for normal operations and that the data driven models are able to provide relatively good predictions even if the input variables are slightly outside the training domain.