Time-domain buffeting response prediction of a long-span bridge: A hybrid machine learning framework

As bridge spans continue to increase, wind-induced vibrations become a major concern for structural integrity and serviceability. Buffeting, caused by the impinging turbulence, significantly impacts fatigue life and serviceability of long-span bridges. Consequently, accurate and rapid assessment of buffeting-induced responses is crucial for various applications, including real-time monitoring and risk assessment. This study introduces a novel hybrid machine learning framework designed to simulate the buffeting-induced response of long-span bridges over time, addressing key limitations in existing approaches. Unlike previous studies, which often focused on localized predictions, limited wind scenarios, frequency-domain analysis, and suffered from error accumulation over time, the proposed framework captures the complete time-history response across multiple degrees of freedom, providing a more comprehensive understanding of the bridge's dynamic behavior. The framework combines autoencoders and Long Short-Term Memory (LSTM) networks to enhance the efficiency and accuracy of time-series prediction. Initially, autoencoder networks compress the high-dimensional wind speed and bridge displacement data into lower-dimensional latent spaces, capturing essential features while reducing computational cost. Subsequently, an LSTM network leverages these compressed representations to model temporal dependencies within the buffeting response, predicting the bridge's response based on encoded wind speed. The final predictive model integrates both autoencoders and the trained LSTM: the first autoencoder encodes raw wind speed, the LSTM predicts the latent bridge response from this encoding, and the second autoencoder reconstructs the final predicted bridge response vector. The model's effectiveness is evaluated through a simplified representation of the Lysefjord Bridge, rigorously assessing both interpolation and extrapolation performances. The proposed model achieves a good simulation accuracy on both training and testing sets, making it a compact and computationally efficient tool for real-time monitoring and assessment of bridges under various wind conditions.