Data-driven seismic response prediction of concrete bridges based on neural network intelligent mapping model

Abstract

Traditional methods for predicting seismic responses of bridges rely on empirical equations and numerical simulations, encountering bottlenecks such as low modeling efficiency and high computational costs when dealing with complex nonlinear behaviors. This paper proposes a data-driven prediction framework for seismic responses of concrete bridges based on Long Short-Term Memory (LSTM)/Gated Recurrent Unit (GRU) neural networks, overcoming the limitations of traditional methods in temporal feature extraction and high-dimensional nonlinear mapping. A refined bridge numerical model is constructed using OpenSees, and seismic ground motion records are selected considering the site-specific characteristics of the bridge location. Multi-dimensional seismic input is generated through wave truncation and amplitude adjustment, and a multi-dimensional dataset covering main beam displacement, bearing shear deformation, and pier response is established by conducting nonlinear time-history analysis. After verifying the universality of the prediction framework based on the BWBN model, a dual-neural-network prediction model is established to achieve end-to-end mapping from seismic ground motion sequences to structural response sequences. Model validation indicates that the predicted samples’ correlation coefficients are concentrated in the high-correlation range of 0.9-1.0, with a mean determination coefficient exceeding 0.95, confirming the reliability of the method. A systematic comparison using a multi-dimensional evaluation system (including root-mean-square error, cumulative energy loss distribution, and probability density of correlation coefficients) reveals that GRU, with its simplified gating mechanism, shows a significant advantage in predicting beam end displacement (mean R2 increased by 7%), while the performance of the two models converges in predicting low-frequency pier responses (both mean R2 reaching 0.97). This framework enhances the efficiency of seismic response prediction by two orders of magnitude, providing a high-precision and high-efficiency intelligent tool for bridge seismic performance assessment, and supporting decision-making for engineering resilience enhancement.