Hybrid surrogate model for online temperature and pressure predictions in data centers

The increase in cloud computing and big data storage has led to significant growth in data center (DC) infrastructure that is now estimated to consume more than 1.5% of the world’s electricity. Due to suboptimal DC design and operation, a significant fraction of this energy is wasted because of the cooling systems inability to effectively distribute cold air to servers. Consequently, additional cooling air must be circulated inside a DC to prevent local hot spots, which leads to undercooling at other locations. Row-based cooling is an emerging architecture that provides more effective airflow distribution, which lowers energy consumption. Since available methods are unsuitable for accurate online predictions, a general thermal model is required to predict spatiotemporal temperature changes inside a DC and hence optimize airflow distribution for this architecture. Typical approaches include physical models, computational fluid dynamics (CFD) simulations, and black-box data-driven models (DDMs). All three approaches are limited because they do not encapsulate the entirety of relevant operational parameters, are time-consuming and can provide unacceptable errors during extrapolative predictions. We address these deficiencies by developing a fast, adaptive, and accurate hybrid surrogate model by combining a DDM and the thermofluid transport relations to predict temperatures in a DC. Training data for the DDM is obtained from CFD simulations. An artificial neural network (ANN) with the Rectified Linear Unit (ReLU) activation function is shown to predict pressure distributions accurately in a row-based cooling DC. These predicted pressures are inputs for thermofluid transport equations to determine the temperature distribution. The applicability of the model is demonstrated by comparing predictions with experimental measurements that characterize the influence of varying server workload distribution and cooling unit operational conditions, i.e., temperature set-point, airflow rate, and fan locations, on the temperature distribution. The model can be used to (1) improve cooling configuration design, (2) facilitate thermally aware workload management, and (3) test “what if” scenarios to characterize the influence of operating conditions on the temperature distribution.