基于机器学习的竖直管内超临界压力CO_(2)湍流换热替代模型研究

Surrogate Model of Supercritical Pressure CO_(2) Turbulent Heat Transfer in Vertical Tubes Based on Machine Learning

超临界压力流体在热力发电、航空航天热防护等工业领域具有广泛应用和广阔的应用前景。然而,由于超临界压力流体在准临界温度附近物性变化剧烈,对流换热性能异于常规流体,为系统热设计带来了很大挑战。本文以受浮升力与热加速耦合影响的超临界压力CO_(2)管内湍流换热问题为例,利用机器学习的高斯过程回归算法和协同克里金法建立两种计算模型,研究多精度数据融合方法在复杂传热问题中的可行性及其特点,以期建立快速、高精度的超临界压力流体湍流换热替代模型。结果表明,在本文数据下上述两种机器学习模型在测试集上预测努谢尔特数的误差不超过5%,优于现有准则关联式预测精度,训练及预测时长最多为分钟量级。高斯过程回归结合协同克里金法的模型只需要48个高精度数据结合1200个低精度数据点即可达到采用96个高精度数据点的高斯过程回归算法预测精度。此外,本文基于机器学习建立的替代模型方法也可以应用于其他换热性能预测问题与热系统设计。

Supercritical pressure(SCP)fluids have wide applications and broad prospects in industrial fields such as thermal power generation and thermal protection in aerospace.However,due to the dramatic changes in the physical properties of SCP fluids near the quasi-critical temperature,the heat transfer performance is different from conventional fluids,which brings great challenges to its prediction and system design.In this paper,taking turbulent heat transfer of SCP CO_(2) affected by buoyancy and thermal acceleration in vertical tubes as an example,Gaussian Process Regression(GPR)and co-Kriging method are used to build two models to study the feasibility of multi-fidelity method,aiming at establishing a fast and high-precision model for the prediction of turbulent heat transfer.The results show that the error of the two models in predicting the Nusselt number on the test set is less than 5%,and the training time is on the order of minutes.The model combining GPR and Co-Kriging method requires only 48 high-fidelity samples and 1200 low-fidelity samples to obtain similar accuracy as the GPR algorithm using 96 high-fidelity samples.The method is generalizable in principle and can be applied to other heat transfer prediction problems or thermal system design.