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

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. When 10 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.