数据中心中机柜出风温度的快速模拟

Rapid simulation of outlet air temperature of racks in data centers

为研究数据中心的气流组织,快速预测机柜出风温度,提出使用本征正交分解和代理模型相结合的快速预测方法。以某数据中心为例,首先使用CFD软件计算了大量工况下的温度场数据,并分析了空调对微模块温度场的影响,然后使用本征正交分解方法建立了机柜出风温度的降阶模型。将计算的温度场数据分为训练集和测试集,对比了回归树、提升树、支持向量回归、神经网络、高斯过程回归、三次样条插值、分段三次Hermite插值作为代理模型的预测精度。结果表明:高斯过程回归、三次样条插值和分段三次Hermite插值的预测精度高于回归树、提升树、支持向量回归和神经网络。通过对比代理模型和CFD软件计算温度场花费的时间,发现使用本征正交分解和机器学习方法以及插值方法建立的代理模型的计算速度是CFD软件的100多倍。

In order to study the airflow organization in the data center and quickly predict the outlet air temperature of the rack,a rapid prediction method that combines the proper orthogonal decomposition(POD)and the surrogate model was proposed.Taking a data center as an example,the temperature field data under a large number of operating conditions were calculated by using computational fluid dynamics(CFD)software.The effects of air conditioning on the temperature field of the micro-module was analyzed.Then,the reduced order model of the cabinet outlet air temperature was established via using the POD method.The calculated temperature field data was divided into training set and test set,and the prediction accuracy of regression tree,boosted tree,support vector regression,neural network,Gaussian process regression,cubic spline interpolation,and piecewise cubic Hermite interpolation as surrogate model was compared.The consuming time of calculating the temperature field via using surrogate model and CFD software was compared.It is found that the computational speed of surrogate model composed of POD,machine learning method,and interpolation method was more than 100 times that of CFD software.