Hot spot temperature prediction and operating parameter estimation of racks in data center using machine learning algorithms based on simulation data

In this paper, models to predict hot spot temperature and to estimate cooling air’s working parameters of racks in data centers were established using machine learning algorithms based on simulation data. First, simulation models of typical racks were established in computational fluid dynamics (CFD). The model was validated with field test results and results in literature, error of which was less than 3%. Then, the CFD model was used to simulate thermal environments of a typical rack considering different factors, such as servers’ power, which is from 3.3 kW to 20.1 kW, cooling air’s inlet velocity, which is from 1.0 m/s to 3.0 m/s, and cooling air’s inlet temperature, which is from 16 ℃ to 26 ℃. The highest temperature in the rack, also called hot spot temperature, was selected for each case. Next, a prediction model of hot spot temperature was built using machine learning algorithms, with servers’ power, cooling air’s inlet velocity and cooling air’s inlet temperature as inputs, and the hot spot temperatures as outputs. Finally, based on the prediction model, an operating parameters estimation model was established to recommend cooling air’s inlet temperatures and velocities, which can not only keep the hot spot temperature at the safety value, but are also energy saving.