A convolutional neural network-long short-term memory (CNN-LSTM)-Attention model based on wavelet transform for predicting non-stationary wind pressure coefficients on the surface of terminal glass curtain wall.

Abstract

In this paper, a neural network model combining wavelet decomposition and attention mechanism is proposed for the accurate prediction of non-stationary wind pressure on the surface of the glass curtain wall of an airport terminal building under strong wind conditions. The traditional methods often prove difficult in capturing local features and time-frequency variations of non-smooth signals when dealing with them, resulting in limited prediction accuracy. The proposed methodology involves a two-step process. Initially, wavelet decomposition of the original wind pressure coefficient sequence is performed, resulting in the reconstruction of subsequences at high and low frequencies. Subsequently, a convolutional neural network-long short-term memory (CNN-LSTM) neural network model incorporating an attention mechanism is constructed, leading to the attainment of high-precision wind pressure predictions. The experimental results demonstrate that the model performs well in the task of predicting non-stationary wind pressure coefficient sequences, with significantly lower prediction errors compared to a single prediction model. Furthermore, in comparison with alternative models that do not incorporate wavelet decomposition, the wavelet transform-CNN-LSTM-Attention model proposed in this paper has the capacity to enhance the mean absolute error, root mean square error, and mean absolute percentage error metrics by 15% to 18%, 12% to 16%, and 26% to 46%, respectively. This study provides reliable technical support for the safety assessment of glass curtain wall structures of airport terminals under extreme weather conditions, and has important engineering application value.