Multivariate Prediction of Airflow and Temperature Distributions Using Artificial Neural Networks

Thermal Management optimization for data centers, including prediction of airflow and temperature distributions, is generally an extremely time-consuming process using full-scale CFD analysis. Reduced order models are necessary in order to provide real-time assessment of cooling requirements for data centers. The use of a simulation-based Artificial Neural Network (ANN) is being investigated as a predictive tool. A model for a basic hot aisle/cold aisle data center configuration was built and analyzed using the commercial software FloTHERM. The flow field and temperature distributions were obtained for 100 representative sets of operating conditions using the CFD package. The Latin Hypercube Sampling technique was employed to select values for three design variables: plenum height, percentage open area of the perforated tiles and air leakage fraction. The FloTHERM results were used to generate a database for the ANN training. The CFD results from 85 cases were used for training and 16 cases were used for validation. A multivariate mapping between the input design variables and output variables (individual tile flow rates and maximum rack temperatures) was obtained. Good agreement (0.5% average relative error) was obtained between the ANN model predictions and the CFD results. These preliminary results are promising and show that an ANN based model may yield an effective real-time thermal management design tool for data centers.