Deep learning-based extension of wind pressure time series

Abstract

The spatio-temporal variation of the wind pressure field is crucial for understanding structural loads and their effect on design. However, obtaining long-duration wind pressure time series around bluff bodies through wind tunnel tests or stochastic and computational simulations is both costly and time-consuming. To address this challenge, this study develops a deep learning (DL) model called WPTSE-Net for extending non-Gaussian wind pressure time series, thereby eliminating the need for the characterization of their nonlinear features and providing an end-to-end flexible framework for extending pressure coefficient time series. The key innovation of WPTSE-Net lies in the reconstruction of the encoder, utilizing prior knowledge to eliminate complex steps in searching for the latent space. This improvement not only enhances computational efficiency and model performance but also substantially reduces the amount of training data that is required for the DL generative model. Comparative results indicate that the proposed WPTSE-Net model outperforms traditional methods in terms of statistical characteristics, i.e., spectra, and peak value distributions. Thus, WPTSE-Net is highly suitable for practical engineering applications as it provides an efficient means of generating long-time series of wind pressure on bluff bodies in wind resistance design.