Research on fast prediction of multi plane airflow organization in data centers based on U-Net neural network

LIU Yuce; ZHOU Chaohui; HU Yue; GUAN Weiwei; HUA Rui; ZHANG Wenkai; ZHENG Zeyu; ZHAO Yang

doi:10.19301/j.cnki.zncs.2025.04.001

Abstract：Accurate airflow organization analysis is critical to the safe and energy-efficient operation of data centers.The traditional computational fluid dynamics(CFD)numerical simulation method is time-consuming,which restricts its application in the operation and maintenance stage of data centers.Data-driven methods have been preliminarily used to quickly predict airflow organization,but the generalization ability of the existing methods is poor,and it is difficult to accurately predict the airflow organization at different locations in the computer room.The paper proposes a fast prediction method of multi-plane airflow organization in data center based on U-Net neural network,which uses the ability of U-Net to process spatial features and retain high-resolution information,combined with the ability of artificial neural network to encode location information,to quickly predict the airflow organization of any plane of data center,and applies it in a data center in Hubei Province to verify its effect.The results show that compared with the conventional convolutional neural network(CNN)architecture,the proposed method reduces the mean absolute error(MAE)by 69.69%,the mean absolute percentage error(MAPE)by 69.85%,and the coefficient of determination(R2)by 7.02%.This method has good generalization ability and can meet the actual needs of airflow organization analysis in data centers.