The prediction on non-Gaussian characteristics of wind pressure for the long-span roof in the mountainous area using proper orthogonal decomposition–deep learning framework

Abstract

To make an accurate prediction of the non-Gaussian characteristics of wind pressure for the long-span roof, this study combines the proper orthogonal decomposition (POD) technique, convolutional neural network (CNN), and long short-term memory (LSTM) network to propose a novel POD-CNN-LSTM framework. Then, the proposed framework was well validated based on the wind tunnel testing of a long-span roof structure, and some error criteria, such as mean square root error and correlation coefficient, were adopted to evaluate the prediction accuracy of the non-Gaussian characteristics. Furthermore, two other methods, POD-CNN and POD-LSTM, were also used to conduct a comparative study. The obtained results illustrate that compared to POD-CNN and POD-LSTM, the proposed framework can achieve better performance on the pulsating wind pressure coefficient. For predictions of non-Gaussian characteristics, the output results of the proposed POD-CNN-LSTM show fewer errors, which means the predictions are close to the measured results, including skewness, kurtosis, and wind pressure probability density distributions. To summarize, the proposed POD-CNN-LSTM framework shows superiority over others, which means the proposed framework has good potential for the practical application of non-Gaussian prediction of the engineering structure.