In this paper, we investigate energy-efficient control of data center air conditioning systems with outside air intakes. Taking account of the slow time constants inherent in the air conditioning systems, the present algorithm determines the current control actions, temperature set-points and intakes of the outside air, based on the predicted future values of physical quantities. To this end, we present an online modeling algorithm based on so-called just-in-time (JIT) modeling techniques with novel key variable selections based on both physical knowledge and data-based method, stepwise method. It is then shown that the future quantities are successfully estimated with an accuracy of correlation coefficient 0.97 or more. We then implement the present algorithm on a large-scale real data center and demonstrate that it successfully reduces the power consumption while meeting several operational constraints. Comparing the data over 2 months, the power consumption is shown to be reduced by 28.9% relative to a conventional method.
