Airflow and temperature distribution optimization in data centers using artificial neural networks

To control energy usage in data center rooms, reduced order models are important in order to perform real-time assessment of the optimum operating conditions to reduce energy usage. Here computational fluid dynamics (CFD) simulation-based Artificial Neural Network (ANN) models were developed and applied to a basic hot aisle/cold aisle data center configuration in order to predict thermal operating conditions for a specified set of control variables. Once trained, the ANN-based model predictions were shown to agree well with the CFD results for arbitrary values of the input variables within the specified limits. In addition, the ANN model was combined with a cost function based multi-objective Genetic Algorithm (GA), which enabled the operating conditions to be inversely predicted for specified values of the output variable (e.g., server rack inlet temperatures). The ANN-GA optimization approach considerably reduces the total computation time compared to a fully CFD-based response surface optimization methodology. Consequently, operating conditions are capable of being reliably predicted in seconds, even for configurations outside of the original ANN training set. These results show that an ANN based model can yield an effective real-time thermal management design tool for data centers.