Seismic Response Prediction of Bridges Based on SCINet‐LSTM‐Attention Model

Accurate and effective prediction of seismic response is crucial for seismic resistance of bridges and structural health monitoring (SHM). However, finite element analysis methods for bridge seismic response are time‐consuming. Furthermore, multiscale sequence information from nonstationary seismic time‐series is not effectively extracted by traditional machine learning approaches, and challenges remain in improving the accuracy of response predictions. In this paper, a hybrid deep learning model for predicting bridge responses under seismic actions is proposed to accurately capture dynamic characteristics in nonstationary seismic time series and effectively utilize critical information within the data. In the proposed framework, sample convolution and interaction network (SCINet) is employed to decompose and interactively extract hierarchical temporal features from seismic excitation and structural response sequences, the long short‐term memory (LSTM) is used to characterize sequential dependency, and the attention mechanism adaptively emphasizes informative temporal features associated with critical structural responses. The performance of the proposed SCINet‐LSTM‐attention model is validated through numerical studies of a continuous girder bridge and a cable‐stayed bridge, as well as shaking table test of a bridge tower. The accuracy of the proposed method is further compared to LSTM, attention‐based LSTM (AttLSTM), and CNN‐LSTM models. Additionally, a comparative study for varying noise levels, damping ratios, ground motion amplitudes, and hyperparameters is investigated. The numerical and experimental results demonstrate that the SCINet‐LSTM‐attention model exhibited superior performance in predicting bridge responses compared to the reference models, indicating potential as an efficient surrogate model for monitoring‐oriented seismic response prediction of bridge structures.