基于CNN与GAN深度学习模型近壁面湍流场超分辨率重构的精细化研究

REFINED STUDY OF SUPER-RESOLUTION RECONSTRUCTION OF NEAR-WALL TURBULENCE FIELD BASED ON CNN AND GAN DEEP LEARNING MODEL

由城市抗风减灾的目标出发,城市边界层的高保真再现是工程界亟待解决的关键问题.基于高精度的近地风场,有望实现真实环境下城市建筑风致效应的准确预测.传统的基于气象模型的城市风场模拟方法存在预测耗时长、成本昂贵、求解尺度过高等缺陷.为更准确、高效地预测边界层的空间变化,研究利用超精度卷积神经网络(SRCNN)与生成对抗神经网络(SRGAN),在空间上将低精度的近壁面湍流场超精度重构成高精度的风场.利用近壁面湍流直接数值模拟的公共数据库训练模型并评价模型的重构性能.为寻求合适的超精度模型生成方式,研究围绕训练样本量及网络深度,开展详细的敏感性分析,确定合适的训练网络及其较优的训练参数设置.同时,基于经不同下采样因子处理的低精度流场输入,分析模型在近壁面湍流重构中的适用范围.研究发现,对比于SRCNN模型,SRGAN模型对近壁面湍流内小尺度结构的重现效果更佳.当基于4层卷积残差块、300样本量开展训练时,所生成的SRGAN模型可在较低的训练代价下实现较优的重构效果.当进行10倍超精度重构时,SRGAN模型可保证较理想的预测精度.研究成果为边界层风场的准确重构提供技术支撑,为城区建筑物风致效应的高效预测提供精确的入流条件.

For the purpose of urban wind resistance and disaster reduction,the reconstruction of urban boundary layer with high fidelity is a key issue to be solved in the wind engineering.Based on high-resolution near-ground wind field,it is expected to achieve the accurate prediction of the wind-induced effect of urban buildings in the real environment.Traditional simulation method based on meteorological models have shortcomings including long forecasting time,high computational cost and limited solution resolution.In order to predict the spatial variation of the near-wall turbulence more accurately and efficiently,the super-resolution convolutional neural network(SRCNN) and generative adversarial neural network(SRGAN) are applied to reconstruct super-resolution near-wall turbulence field from the low-resolution one.This study employs the public database,which is built up from direct numerical simulation of turbulent channel flow,to train the reconstruction models and evaluate their performance.In order to determine a suitable model generation strategy,this study searches for a proper training neural network and its optimal parameter setting based on detailed sensitivity analysis of the training sample size and network depth.What's more,the application scope of the model is explored for near-wall flow field reconstruction,based on low-resolution inputs obtained by different down scale ratios.It is found that the SRGAN model has a stronger capacity to reproduce the small-scale structures in the turbulent near-wall flow,compared with the SRCNN model.As the training is based on 300 sample sizes and 4 convolutional residual blocks,the generated SRGAN model can obtain a good reconstruction accuracy,at a relatively lower training cost.When10 times super-resolution reconstruction is carried out,the SRGAN model can still maintain ideal prediction performance.The research results offer reliable technical support for reconstructing near-wall turbulence based on artificial intelligence.Subsequently,it provides precise inflow conditions to efficiently predict the wind-induced effects on buildings in urban areas.