A Hybrid GWO-VMD-LSTM Surrogate Model for Vehicle-Track-Bridge Response Prediction Under Near-Fault Earthquakes

To enhance the predictive efficiency and accuracy of high-speed railway vehicle-track-bridge (VTB) system responses under near-field earthquakes, this study proposes a hybrid surrogate model (GWO-VMD-LSTM) integrating long short-term memory neural networks (LSTM), grey wolf optimizer (GWO), and variational mode decomposition (VMD). This model employs GWO to optimize LSTM hyperparameters and VMD decomposition parameters, while incorporating temporal feature extraction and signal decomposition techniques to establish an efficient prediction framework. Its reliability and performance were systematically validated through field-measured data from an actual bridge engineering application. Then numerical case studies based on a nonlinear finite element model of a 9-span 32 meters VTB system were simulated by the extended Open Sees platform to perform a comparative evaluation of normal vibration responses and near-field seismic excitation scenarios. Key findings reveal that the GWO-LSTM model demonstrates superior performance in bridge displacement prediction compared to conventional convolutional neural networks (CNN) and LSTM architectures, achieving sustained high determination coefficients ([Formula: see text]). The enhanced GWO-VMD-LSTM configuration achieves significant improvement in acceleration response prediction accuracy, showing a maximum of 70% reduction in RMSE and MAE compared to CNN and LSTM models. The proposed hybrid surrogate model effectively captures nonlinear coupling characteristics of the VTB system under near-field seismic actions while achieving a reasonable balance between computational precision and efficiency, demonstrating substantial potential for seismic safety rapid assessment of moving trains across bridges during near-fault earthquakes.