Abstract
Improving the cooling efficiency of a data center becomes important due to the cooling power reaches one-half of total power budget in operating the data center. The major cause of inefficient cooling is the poor airflow management resulting in cold air bypass and hot air recirculation. Basically the architecture of airflow management includes the separation of cold aisle and hot aisle and raised floor with perforated tile. The cold air is generated by the Computer Room Air Conditioning (CRAC) unit and is delivered to the server rack through the perforated floor. To avoid hot air recirculation, an active tile is used to enhance the airflow in a specific location where the flow resistance is larger than the others. The priority of the airflow control is to eliminate hot spot owing to the hot air recirculation such that the control strategy is conservative. Recently many studies pay attention to real-time temperature prediction by utilizing airflow model to improve the control performance of the airflow. However the derivation of airflow model suffers the difficulty from the airflow measurement. Therefore this research aims at the development of deep learning model for predicting the inlet temperature distributions in real time for use in control strategies in order to improve energy usage dynamically. A rack tested system was established for the validation of the temperature prediction by using deep learning. In this paper we define the workload schedule of the rack to collect the training data and the GRU recurrent neural network model is used to develop the temperature prediction model. The experimental results show that the prediction model can predict the inlet temperature at a root mean square error of 0.29°C.