化工进展2026，Vol.45Issue(7)：3897-3907,11.DOI:10.16085/j.issn.1000-6613.2025-1173
基于CNN-MLP-ThermoAttention的数据中心三维温度场预测
Prediction of three-dimensional temperature fields of data center based on CNN-MLP-ThermoAttention
摘要
Abstract
The temperature field distribution in data centers directly affects operational safety,equipment performance,and cooling energy consumption.While the mainstream computational fluid dynamics(CFD)simulation can obtain 3D temperature contour maps,it is constrained by computing power,with a single simulation taking hours to days,failing to meet the real-time requirements of actual data center operation scenarios.To address this,this paper used CFD data as training samples and proposed the thermodynamics-constrained attention(ThermoAttention)mechanism,which embedded thermodynamic laws into the multi-head attention mechanism and coupled it with convolutional neural networks and multi-layer perceptrons to construct a neural network framework capable of real-time prediction of 3D temperature fields.Combined with point cloud technology and particle systems,dynamic reconstruction and visualization of temperature field contour maps were realized in Unity 3D.Experimental results showed that the model's prediction error was≤2％and a single prediction was shortened to minutes or even seconds,providing efficient support for real-time operation and maintenance of data centers and offering a new approach to real-time prediction of high-dimensional physical fields.关键词
数据中心/计算流体力学/神经网络/算法/温度场预测Key words
data center/computational fluid dynamics(CFD)/neural networks/algorithm/temperature distribution prediction分类
化学化工引用本文复制引用
汪方舟,刘丰,杜忠选,吴莉莉,乐逸凡,胡姝凡,曹军..基于CNN-MLP-ThermoAttention的数据中心三维温度场预测[J].化工进展,2026,45(7):3897-3907,11.基金项目
国家重点研发计划(2025YFEO199100). （2025YFEO199100）
