Reduced Order Thermal Modeling of Data Centers via Distributed Sensor Data

Abstract

In this paper, an effective and computationally efficient proper orthogonal decomposition (POD) based reduced order modeling approach is presented, which utilizes selected sets of observed thermal sensor data inside the data centers to help predict the data center temperature field as a function of the air flow rates of computer room air conditioning (CRAC) units. The approach is demonstrated through application to an operational data center of 102.2 m2 (1100 square feet) with a hot and cold aisle arrangement of racks cooled by one CRAC unit. While the thermal data throughout the facility can be collected in about 30 min using a 3D temperature mapping tool, the POD method is able to generate temperature field throughout the data center in less than 2 s on a high end desktop personal computer (PC). Comparing the obtained POD temperature fields with the experimentally measured data for two different values of CRAC flow rates shows that the method can predict the temperature field with the average error of 0.68 °C or 3.2%. The maximum local error is around 8 °C, but the total number of points where the local error is larger than 1 °C, is only ∼6% of the total domain points.