Abstract: Empirical mode decomposition (EMD) is used to decompose the non-stationary wind pressure of a long-span membrane structure into a series of relatively stationary intrinsic mode functions and a residual component. In order to eliminate the effect of noise on the intrinsic mode function in actual wind pressure measurement, each intrinsic mode function is denoised by using wavelet transform. The residual components and the intrinsic mode functions after denoising are input as samples. Radial basis kernel, Hermite kernel and Hermite combination kernel are combined with least square support vector machine (LSSVM) respectively. Subsequently, optimizations for penalty parameters and kernel parameters are conducted using particle swarm optimization (PSO) and thus three algorithms based on PSO-LSSVM are proposed for wind pressure prediction. In addition, the robustness of the combined model is verified based on the measured wind pressure on the super high-rise building. The single point predicting indexes show that the prediction algorithm using Hermite combination kernel based on PSO-LSSVM has higher prediction accuracy and generalization ability compared to the other two algorithms. The results of spatial point prediction further prove the validity of this method for non-stationary non-Gaussian wind pressure prediction.

Key words: wind pressure prediction; empirical mode decomposition (EMD); wavelet transform; least square su-pport vector machine (LSSVM); particle swarm optimization (PSO)