Artificial Neural Network Based Prediction of Temperature and Flow Profile in Data Centers

Abstract

Thermal management of data centers continues to be a challenge because of their ever-increasing power densities, due for example to shrinking server footprints. Computational fluid dynamics and heat transfer (CFD/HT) have been used extensively to model thermal transport and air flow in data centers. However, the significant computational costs and time associated with accurate room-level CFD/HT simulations for data centers makes such simulations impractical for real-time prediction and control. Nevertheless, developing an effective control framework for data centers to minimize power consumption requires such real-time prediction of server inlet temperatures and tile flow rates. This paper focuses on the development of artificial neural network (ANN)-based models trained on datasets generated from offline CFD/HT simulations for real-time prediction of temperature and flow distributions in a data center. Using CFD simulation results to train ANN transfers computational complexity from model execution (in CFD) to model setup and development. A physics-based and experimentally validated room-level CFD/HT model was developed using Future Facilities 6Sigma Room. The numerical model of the room housing the data center includes an under- floor supply and ceiling-return cooling configuration and consists of one cold aisle with 12 racks arranged on both sides and three CRAC units around the periphery of the room [1]. For steady state modeling, flow and temperature distributions were obtained for 300 representative cases using CFD/HT simulations and used to train the ANN model using the Levenberg-Marquardt backpropagation algorithm. The multidimensional input parameter space for the simulations was comprised of the computer room air conditioner (CRAC) blower speed (N CRAC ), the CRAC return air temperature set point (T a,ret ) and the IT load distribution for the racks ($\dot{Q}_{room}$), while rack inlet temperature (T rack inlet ) and tile flow rate ($\dot{V}_{tile}$) were the predicted variables. The trained ANN model was tested with 33 test cases obtained from the same multi- dimensional input space. The results show good agreement with CFD simulations with an average error of < 0.6°C in rack inlet temperatures and 0.7 % in tile flow rates. For transient modeling a scenario with cooling failure was considered where first 200 s of data after failure were used for training ANN and subsequent 300 s were used for testing the accuracy of extrapolative predictions. The ANN was compared with the Proper Orthogonal Decomposition (POD) modeling framework in terms of prediction accuracy and computational time for this transient scenario. The validated neural network model can be used to obtain rapid prediction of temperature and flow distributions, and when combined with appropriate control strategy, can be used for real-time control of data centers.