Temperature Prediction of Flat Steel Box Girders of Long-Span Bridges Utilizing In Situ Environmental Parameters and Machine Learning

Abstract

Design, construction, and maintenance of large-span bridges require an accurate assessment of the temperature field in flat steel box girders (FSBGs). While this field is controlled by various environmental (meteorological) factors, including temperature, solar radiation, humidity, wind speed, and wind direction, there is no comprehensive model for its prediction based on multiple environmental variables. Given this, two novel methods for calculating the cross-sectional effective temperature (ET) of the FSBG were proposed in this study. Based on the bridge’s environmental variables measured on-site, regression models for predicting ET and vertical temperature difference (VTD) in FSBG were introduced, including a random forest (RF) model and empirical formulas. The RF model’s hyperparameters were derived by the Bayesian optimization algorithm. The proposed approach was applied to the case study of the Sutong Bridge, China, using 2 years’ data samples collected via the bridge health monitoring system and Copernicus Climate Change Service. The model’s training and testing results proved that the predictive performance of the multifactor random forest model significantly exceeded that of the single-factor linear model by about 60%. The RF model’s accuracy in the ET/VTD prediction also outperformed the support vector regression model and back-propagation neural network model. Besides, the correlation analysis of environmental variables revealed a significant time-lag between ET/VTD and the surface solar radiation intensity (about 3 h). The predictive performance of the RF model considering the time-lag effect was further improved (by about 20%–30%).