应用不同深度学习代理模型的灯笼型扰流柱通道换热性能分布预测方法比较

Comparison of Different Prediction Methods for Heat Transfer Distribution of Lantern-Shaped Pin-Fin Channel with Different Deep Learning Surrogate Models

为提高对灯笼型扰流柱通道端壁面的换热分布预测能力,构建并比较了几种深度学习代理模型的预测性能。应用数据驱动的思想,建立了多项式响应面模型(RSM)、径向基函数模型(RBF)、径向基神经网络模型(RBFNN)、Kriging模型和总体平均近似模型(Ensemble)共5种传统代理模型以及5种基于生成对抗网络的深度学习代理模型,包括有残差网络的pix2pix模型、pix2pixHD模型、CycleGAN模型、StarGAN模型以及单一无残差网络的pix2pix模型,以扰流柱截面参数为设计变量,通过拉丁超立方抽样分别得到了样本数为50、25、12、6和3的训练集,并根据扰流柱尾迹高换热区分布特点,将训练样本数为50的数据集分为宽样本数据集和窄样本数据集,比较了不同代理模型对端壁面换热性能的预测精度、计算成本和泛化能力。结果表明:有残差网络的pix2pix模型相比于无残差网络的pix2pix模型,预测精度得到有效提高,在样本数为50的情况下,面平均值预测误差从0.68%降低到0.32%,平均相对误差从6.89%降低到6.41%,而且当训练样本数减少时,有残差网络的模型预测能力更加突出;传统代理模型的时间成本可忽略不计,但深度学习模型的单卡训练时间较长,且增加残差网络后的模型计算成本更高;当训练样本数为50时,传统代理模型和深度学习模型之间预测精度差异不大;当训练样本数逐渐减少时,深度学习代理模型展示出更高的预测精度和泛化能力;Kriging模型虽然泛化能力强,但是预测结果趋同;RSM模型、RBF模型和Ensemble模型泛化能力最差,训练样本数较少时,预测结果严重失真。可见在换热性能预测方面,深度学习模型在预测精度与泛化能力上均有显著优势,尤其适合于小样本问题,对提高灯笼型扰流柱截面设计效率具有参考价值。

In order to improve the prediction ability of heat transfer performance on lantern-shaped pin-fin endwall,several deep learning surrogate models are constructed and their prediction performance compared.Based on data-driven thinking,five traditional surrogate models,namely polynomial response surface model(RSM),radial basis function model(RBF),radial basis function neural network(RBFNN),Kriging model and ensemble of surrogate model(Ensemble)are proposed.Additionally,five generative adversarial network surrogate models based on deep learning,namely pix2pix model,pix2pixHD model,CycleGAN model,StarGAN model all with residual network,and another pix2pix model without residual network are proposed.The training sets with various sample numbers of 50,25,12,6 and 3 were obtained by Latin hypercube sampling,respectively,using the pin-fin cross-sectional shape as the design variables.According to the distribution characteristics of high heat transfer zone,the dataset with a training sample number of 50 is divided into a wide sample dataset and a narrow sample dataset.On this basis,the various heat transfer performance of different surrogate models are compared in terms of prediction accuracy,computational cost and generalization ability.The results show that,compared with the pix2pix model without residual network,the prediction accuracy of the pix2pix model with residual network is effectively improved,and the prediction error of the surface mean is reduced from 0.68%to 0.32%,and the averaged relative error is reduced from 6.89%to 6.41%when the sample number is 50.When the number of training samples decreases,the prediction performance of the models with residual network becomes more prominent.The time cost of traditional surrogate models is negligible,but the training time of deep learning models is long especially for models with residual network.When the number of training samples is 50,there is little difference in prediction accuracy between the traditional surrogate models and deep learning models.With gradual decrease in the number of training samples,deep learning surrogate models show higher prediction accuracy and generalization capabilities.Although the Kriging model has strong generalization ability,the prediction results tend to be similar.RSM model,RBF model,and Ensemble model have the lowest generalization ability,and the prediction results are severely distorted when there are fewer training samples.For the heat transfer distribution prediction of the pin-fin endwall,the deep learning models possess significant advantages in terms of both prediction accuracy and generalization ability,which are more suitable for small batch sample problems.This study has important reference value for improving the cross-sectional shape design efficiency of lantern-shaped pin-fin.