| description abstract | The long-span bridge, characterized by slenderness and flexibility, is particularly sensitive to wind action. Extreme wind events, including typhoons and hurricanes, will threaten the safety and serviceability of long-span bridges. More specifically, long-span bridges usually experience significant vibrations under typhoon events, increasing the risk of serviceability failure and traffic accidents. An efficient way to mitigate the risk of such threats is to predict the typhoon-induced response (TIR). Traditionally, TIR prediction of long-span bridges is usually carried out based on finite-element (FE) simulations. Due to the assumptions in formulating the FE model and establishing the wind load (e.g., boundary conditions, simplified structural elements, stationary aerodynamic wind forces), prediction accuracy is inevitably undermined. In this work, as opposed to traditional FE-based analysis, TIR is predicted in real time from a data-driven perspective. Quantile random forest (QRF) with Bayesian optimization is presented as the data-driven method for probabilistic prediction. A long-span cable-stayed bridge with a main span of 1,088 m is used as a test bed to illustrate the effectiveness of the present method. Other optimization algorithms used with QRF (grid search and random search) and response surface methodology (RSM) are also implemented for comparison purposes. Results indicate that QRF with Bayesian optimization can provide reliable probabilistic estimations allowing for quantification of uncertainty in prediction. It shows superior performance compared with other optimization algorithms and models in terms of accuracy and computational expense. The analysis and predictive framework for TIR are expected to provide insight into data-driven structural wind engineering. | |
