| Over the years,many countries in the world have been affected by the wind disaster,which has brought great losses to the safety of people’s lives and property.China’s coastal areas have also been greatly affected by the wind disaster.In 2018 alone,nearly 10 typhoons made landfall in China,with typhoon "Shanzhu" causing huge damage to the coastal areas of southern China.According to the survey statistics,the economic losses caused by building damage generally account for a larger proportion.Among the many types of buildings,low-rise buildings are more vulnerable to damage and more severe damage.Therefore,it is necessary to conduct relevant research onthe wind pressure distribution of low-rise buildings.The research methods of wind pressure prediction include Proper Orthogonal Decomposition(POD)method,interpolation function prediction.In recent years,with the rapid development of artificial intelligence technology,the core algorithms of artificial intelligence have been gradually applied to the study of engineering structure wind resistance.So far,many research institutions and universities around the world have established wind tunnel test databases for wind tunnels in some high-rise and low-rise buildings,and artificial intelligence has benefited from the large amount of data driven by the database in the development of engineering structure wind resistance.The aerodynamic wind loading database for the isolated low-rise building conducted by Tokyo Polytechnic University was adopted in this thesis for the basis data for the numerical prediction on wind pressure on low-rise building.The machine learning and deep learning methods were adopted to numerically predict wind pressure on low-rise building surface.The main research contents of this thesis are as follows:(1)Briefly review the research status of wind pressure on low-rise building surface,wind pressure distribution characteristics,the theory of machine learning and deep learning algorithms.(2)To determine the hyper-parameters of various prediction models,the Bayesian optimization algorithm is proposed in this thesis.The implement algorithm for this method is briefly introduced;Meanwhile the relevant evaluation indicators are proposed for the accuracy of the numerical predicted results based on the actual values obtained from experimental model.(3)The LSTM model in deep learning and the RFR,SVR,XGBoost,and Light GBM models in machine learning were adopted to predict the time series of wind pressure on the surface of low-rise buildings.The comparison results show that the prediction accuracy obtained from the LSTM model is generally higher than the accuracy from of the RFR,SVR,XGBoost,and Light GBM simualtion method.In addition,the LSTM model in deep learning can simultaneously predict the wind pressure time-series of all wind pressure taps on the same building surface simutaneously.Under the condition of guaranteed accuracy,the LSTM method can greatly save the time of wind pressure prediction.(4)The Proper orthogonal decomposition(POD)method was utilized to decompose and reconstruct the wind pressure field of low-rise buildings.Then,the combined POD and Deep Neural Network(DNN)in the deep learning algorithm(DNN-POD),along with the DRT-POD,RF-POD and GBRT-POD in the machine learning method to perform the numerical prediction on the time-series of wind pressure on un-measured taps on the surface of low-rise buildings.By comparing the predicted results,it is found that the accuracy obtained from DNN-POD method is generally higher than thosefrom the DRT-POD,RF-POD and GBRT-POD machine learning methods,and the accuracy of the former can be as high as 98.5%. |
