Long short-term memory with wavelet decomposition for wind speed predicting based on SHM data

The wind field environment surrounding long-span bridges is characterized by its complexity and variability, resulting in wind speed exhibiting random, nonlinear, and uncertain behavior. To enhance bridge safety and mitigate the impact of wind speed, it is crucial to establish a reliable wind speed prediction model. In this study, a structural health monitoring (SHM) system was deployed on a long-span bridge to collect extensive wind speed data, which was subsequently denoised using the wavelet decomposition (WD) method. Leveraging the long short-term memory (LSTM) approach, a wind speed prediction model (WD-LSTM) was developed. The study focuses on investigating the effects of three different thresholds (Bayesian threshold, SURE threshold, and Minmax threshold) in the WD method, the number of hidden units (2, 4, 8, 16, 32, 64, 128, 256, and 512) in the WD-LSTM model, and the number of inputs (one-step prediction, five-step prediction, ten-step prediction, and twenty-step prediction) in the WD-LSTM model on the prediction performance of wind speed. Evaluation metrics such as RMSE and R2 are employed for this analysis. Furthermore, the calculation time of the WD-LSTM prediction models with different hidden units and inputs is compared. Finally, an optimal WD-LSTM prediction model is proposed, taking into account both prediction accuracy and calculation time.