Research

Data centers account for 1% of global energy consumption, and optimizing their energy use is a priority in the industry. Since cooling contributes several tens of percent of the total energy consumption in many data centers, the industry is moving towards very energy-efficient data center buildings. An effective cooling strategy is the containment strategy, which consists of separating the cold air going into the servers (supply air) and the hot air going out of the servers (return air), creating physically secluded cold and hot aisles. One important source of parasitic heat in data halls is the hot-air recirculation, flowing back from the hot to the cold aisle at the rack level and having multiple negative consequences, such as the creation of hot spots damaging the hardware and increasing the data center cooling needs. To prevent this phenomenon and improve the data centers' energy efficiency, the ability to predictively model conditions and events, even those not seen before, in data centers is increasingly important. We have identified three main reasons leading to recirculation in a Data Hall: baseline Recirculation characterized by design Hot Aisle Containment and racks leakage, recirculation caused by negative differential pressure locally at the rack, and flow deficit at the end of an aisle when there is not enough supply air to cool the racks. We developed physical-law-based models to predict thermal conditions in a data hall accurately and quantify the three types of recirculation airflow. This paper presents an application of these models in an actual data center in the United States with a successful prediction of the thermal behavior within 1°F of MAE (mean absolute error) for cold aisles in normal conditions. It achieves more accurate predictions than a data-science-only-based model in under-provisioned conditions.