Prediction Of Wind Loads On Large-span Roof Structures Based On Data Mining Technology

The large-span structures have various forms of construction,and the distribution and fluctuating of wind load on their surface are very complicated.When design the structure,the determination of wind load often becomes the key and difficult point in the design process.So far,a large number of wind tunnel test data have been accumulated wind tunnel laboratories and many wind load databases have been established to facilitate the storage and application of these test data.However,the actual engineering structure does not always exactly match the model in the database.In order to fully utilize the wind load data in the database,forecasting the wind load based on data mining technology is an important measure to solve the problem.By comparing and analyzing the basic theories and characteristics of different data mining methods and combining with the characteristics of wind load data,the basic process of wind load forecasting is confirmed.Based on the wind tunnel test data in wind load database of Harbin Institute of Technology,wind load forecasting models are established for three kinds of typical large-span space roof forms,namely plane,spherical and three-center cylindrical surfaces,respectively,to predict the average wind pressure and fluctuating wind pressure on the surface of the structure,and compared with the wind tunnel test results to verify the feasibility and accuracy of the method.Finally,the wind load forecasting platform of large-span space structure is built based on MATLAB,including model training module and wind load forecasting module,which is convenient for research and engineering application.The main contents include:1.The primary wind load data is preprocessed by kernel principal component analysis and hierarchical cluster method in cluster analysis.The GRNN neural network algorithm is used to establish the wind load forecasting model.In order to improve the efficiency,the smoothing factor is determined based on double subgroups fruit fly optimization algorithm with characteristics of Levy flight.2.For the spherical roof,wind load forecasting model is established without using the cluster analysis to preprocess the sample data in consideration of that there is no need to consider the effect of the wind direction and the number of working conditions is small.The forecasting results of average wind pressure and fluctuating wind pressure under different conditions are obtained and the best smoothing factor values are determined.3.For the flat roof,classify the primary data according to the result of hierarchical clustering analysis of the original data based on wind direction.Due to the small number of working conditions,there is no need of cluster analysis based on the conditions;Then establish wind load forecasting model with the clustering analysis results based on wind direction,Obtain the prediction result and the best parameter value.4.For the three-center cylindrical roof,a series of wind tunnel tests is completed firstly.All the primary data is classified by hierarchical clustering analysis based on the conditions on account of the complicacy of the wind tunnel tests.The wind load forecasting model is established to predict the average wind pressure and fluctuating wind pressure on the roof surface according to the result of classification and parameterization of the open form.Determine the best parameter values of these forecasting models under different conditions.5.Build the wind load forecasting program platform based on MATLAB which can update the wind load forecasting model when the input data changes as well as realize the wind load forecasting function for the above-mentioned flat roof,spherical roof and threecenter cylindrical roof.