Metaheuristic algorithm-optimized framework for predicting bridge measurement data using TVFEMD and deep learning

Bridge wind speed and response prediction are crucial for early warning and abnormal condition detection in bridge health monitoring. However, the complexity and randomness of measurement data caused by long-term exposure to combined vehicle and environmental loads pose challenges in improving prediction accuracy. In addressing this challenge, this study proposes a novel hybrid framework combining time-varying filtering-based empirical mode decomposition (TVFEMD), Grey wolf optimization (GWO), gradient-based optimization (GBO), and long short-term memory (LSTM) network for predicting bridge measurement data. The GWO algorithm is employed to optimize decomposition parameters (i.e., bandwidth threshold and B-spline order) of the TVFEMD method, and the GWO-TVFEMD can adaptively decompose the measurement data into several stable subseries. Additionally, the GBO algorithm is employed to optimize the number of hidden layers, learning rate, and maximum iterations of LSTM to enhance the deep learning performance. Experimental results from bridge field measurements demonstrate that the proposed hybrid model outperforms the variational mode decomposition LSTM and TVFEMD-LSTM models. Moreover, the proposed framework exhibits good generalization capabilities in predicting bridge wind speed, displacement, and strain, providing reliable results for practical bridge engineering.