Rack level forecasting model of data center

Due to the nature of their operations, electronic data centers are highly dynamic in their computational load handling, and concomitantly in the heat generated by various servers inside the data center. In the absence of physics-based control algorithms, air-conditioning systems are typically operated through conservatively pre-determined set points, thereby resulting in sub-optimal electricity consumption. Computational fluid dynamics based models that have been developed are too computationally intensive to be used in control algorithms. Here, we develop a validated transient model for a data center that can be used as the basis for model predictive control algorithms. We consider a common raised floor plenum data center in which cold air is delivered to the server racks and hot air is exhausted through the ceiling. We test and demonstrate the performance of the data driven model on this data center. Thermal and flow dynamics in a data center are strongly affected by rack inlet and rack outlet temperatures, in addition to the computational load (and thereby the heat generated by the racks). Using these variables, the response data were collected at the top, center and bottom position of server simulator rack inside a raised floor plenum data center. The server simulator rack consists of four server simulators, each with independent control of heat load & fan speed. The simulator was first allowed to reach steady state (by holding conditions constant for over 3 hours). Subsequently, the heat load was ramped up in steps. To reach steady state, all simulators were held at 2.5 kW constant heat output with 0.15 m
³
/s fan speed. The computer room air-conditioning (CRAC) unit was modeled as a constant air supply at 17 °C. The response data were collected for various cases with heat load as input to the system. This input-output data was used for data driven modeling to predict the rack outlet air temperatures. The data analytics techniques that were used include partial correlation and total least square approach. Results show good agreement with experiments and have prediction capacity of 15 minutes. These results indicate that data driven modeling can potentially be used to form the basis for model predictive control.