Application Of Deep Learning Super-resolution Reconstruction Method In Wind Field Reconstruction Of Classical Buildings

The refined and high-resolution building flow field data can provide technical guidance for structural design,auxiliary calculation and safe layout.The traditional high-resolution flow field data is mainly obtained based on interpolation method and reconstruction method,but there are some shortcomings such as inaccurate texture information,jagged and insufficient prior information.Super-resolution reconstruction method based on machine learning is an effective image refinement method developed in recent years,and has been widely used in various fields such as security monitoring,medical image and remote sensing image.However,according to the author’s investigation,there is no systematic research on the super-resolution reconstruction of complex wind field in buildings.Therefore,this paper constructs a Deep Learning(DL)model for super-resolution reconstruction of seriously under-resolution flow field data.The model was based on the Convolutional Neural Network(CNN)and combined with the mixed Downsampled Skip-Connection/Multi-Scale(DSC/MS)model.The constructed super-resolution reconstruction model is applied to the high-resolution wind field of typical buildings,and the effectiveness of deep learning-based super resolution reconstruction method in the reconstruction of wind field inside and outside buildings is explored.Specific work has been carried out as follows:1.Super-resolution reconstruction of the flow field around the CAARC(Common Wealth Advisory Aeronautical Research Council)standard building model was carried out.Based on large eddy simulation flow around the CAARC standard architectural model to carry out the numerical simulation of the flow field analysis of the distribution characteristics of flow field,flow around buildings constructed building surface wind pressure field and velocity field data sets.The reconstructed super-resolution model was applied to the reconstruction of surface wind pressure field and velocity field around building flow field of CAARC standard building model,and the reconstruction abilities of wind pressure field and velocity field reconstructed by different methods in different under-resolution flow fields were compared.The results show that the deep learning model has good accuracy in reconstructing the high-resolution outflow field,and the reconstruction effect is better than the original convolutional neural network model and the traditional bicubic interpolation method.In the wind pressure field,the maximum relative error reduced by Multi-scale CNN model is 54.68%compared with bicubic method.Due to the universality of the method,it can be extended to the super-resolution reconstruction of wind fields around buildings with complex turbulent flows.2.The super-resolution reconstruction of the hydrogen leakage flow field in the high pressure pipeline in the hydrogen building was carried out.Based on Realizable k-εturbulence model and component transport model,the hydrogen leakage flow field in high pressure pipeline in hydrogen building was numerically simulated.The velocity of the flow field and the distribution characteristics of hydrogen leakage were analyzed,and the data sets of hydrogen mole fraction field and velocity field were constructed.Based on the Multi-Scale CNN model,the super-resolution reconstruction of the flow field in the hydrogen plant was carried out,and the reconstruction effects of the hydrogen mole fraction field and velocity field reconstructed by different methods in the longitudinal section and three cross sections were compared in different under-resolution flow fields.The relative error and coherence analysis methods were used to evaluate the deep learning model.The results show that the reconstruction error of the Multi-scale CNN model in cross section 3 is higher than that of the CNN model,and the reconstruction accuracy of the Multi-scale CNN model in other sections is improved.In the hydrogen mole fraction field and velocity field,the maximum relative errors reduced by the Multi-scale CNN model are 89.01% and 89.92%,respectively,compared with the bicubic method.