Abstract The prediction of wind-induced buffeting response in long-span bridges is a fundamental scientific challenge in bridge wind engineering. Conventional buffeting analysis methods are constrained by simplifying theoretical assumptions, high experimental costs, and inherent model errors, making accurate full-scale prediction difficult. This study focuses on the Xihoumen Bridge, a representative long-span suspension bridge, and utilizes five-year multi-source heterogeneous monitoring data acquired from its structural health monitoring (SHM) system. By integrating deep learning techniques with classical buffeting theory, a frequency-domain, data-driven buffeting response prediction framework is proposed. Guided by the Davenport–Scanlan buffeting theory, the physical mapping relationship between wind field features and structural acceleration power spectral density (PSD) is first established, which motivates the design of an Attention-embedded Frequency Convolutional Network (Att-FCN) and a Residual Bidirectional Gated Recurrent Unit (Res-BiGRU) module. The prediction process is then physically decomposed into two sequential mapping stages—wind field to buffeting force spectrum and buffeting force spectrum to structural response spectrum—yielding a Cross-Attention Buffeting Network (CA-Buffeting Net). Experimental results demonstrate that CA-Buffeting Net reduces the mean squared error (MSE) by 29.7%, the mean absolute error (MAE) by 20.2%, and improves the cosine similarity (CS) by 1.2% points relative to the baseline model. Visualization of the channel and frequency attention coefficients further validates the physical interpretability of the proposed model, offering an effective technical tool for wind-resistance analysis and structural safety assessment of long-span bridges.
Wu et al. (Sun,) studied this question.