Research On The Prediction Of Wind Load Based On Machine Learning Method

With the rapid development of economy,super high-rise buildings tend to be high-rise and complex,and the span of modern large-span spatial structure is also growing.Wind load has gradually become an important factor affecting the safety,comfort and economic performance of buildings.Therefore,it is of practical significance to obtain the wind pressure distribution characteristics on the surface of super high-rise buildings and large-span roof structures for structural wind resistance design,curtain wall design,wind-induced dynamic response,wind field characteristics analysis,etc.At present,it is difficult to obtain the complete wind pressure distribution characteristics of the structure by wind tunnel test or field measurement.In this paper,four kinds of network prediction models are established by using machine learning method,which are GA-BP network model formed by combining genetic algorithm and BP(error back propagation)neural network,general recurrent neural network(GRNN)model,PSO-BP network model formed by combining particle swarm optimization and BP neural network,and radial basis neural network(RBN)Function(RBF)model.In this paper,the network model is trained by using the wind pressure data collected from field measurement and wind tunnel test,and the wind pressure distribution of unknown measuring points is predicted by using the trained network model,which mainly includes the following contents:As for the establishment of wind pressure distribution prediction model,this paper adopts the method of model optimization and reorganization.Compared with the steepest descent algorithm of the traditional artificial neural network method,the quasi Newton method and the LM(Levenberg Marquardt)learning algorithm in this paper have faster calculation efficiency and convergence speed.Group optimization algorithm is used to optimize the initial weight and threshold of BP neural network model,which embodies the idea of model optimization and reorganization.The problem that the initial weight of the network is sensitive and easy to fall into the local optimal solution is solved by the restructured wind load prediction model,and the stability and prediction performance of the network model have been greatly improved.The performance verification of wind pressure distribution prediction model.In this paper,GA-BP and GRNN network models are trained by using the wind load data collected in the field of super high-rise buildings,and then the wind pressure time history and fluctuating wind pressure time history data of typical measuring points in the windward,crosswind and leeward areas of buildings are predicted by using the trained network models.The prediction results of the network model are compared with the field data in numerical value,power spectrum,non-Gaussian and correlation to verify the prediction performance and applicability of the network model.In this paper,PSO-BP and RBF network models are trained by the wind pressure data collected in the wind tunnel test of long-span roof structure,and the wind pressure distribution characteristics of unknown measuring points in the corner and middle area of long-span roof are predicted by the trained network model.The prediction performance and applicability of the network model are verified by comparing the error between the prediction value of the network model and the wind pressure data of the wind tunnel test.