Strain prediction in a large-span arch bridge using the TimeXer model considering temperature and traffic loads

Strain is an important monitoring item in bridge structural health monitoring, providing a crucial basis for fatigue and safety assessments of structures. Under operational conditions, temperature and random traffic loads pose challenges for bridge strain prediction. To address this issue, this paper proposes a strain prediction framework over future forecasting horizons that explicitly considers both temperature and traffic loads. Historical traffic loads and temperatures are used as exogenous variables, and the TimeXer network is employed to predict the characteristics of temperature-related and traffic-induced strain in bridges, enabling the prediction of hourly strain characteristics over future horizons of 24, 48, and 96 h. Based on a year-long monitoring dataset from a large-span steel arch bridge, a strain dataset for typical locations was generated to validate the proposed method. The results demonstrate that TimeXer can accurately predict temperature-related strain and also effectively capture the trends of traffic-induced strain. Compared with traditional long short-term memory or other Transformer-based models, TimeXer, by incorporating exogenous variables, significantly improves prediction accuracy, achieving the smallest average error across all datasets. Based on the data from six strain measurement points on the in-service bridge, the proposed prediction method demonstrated the best overall performance. The findings demonstrate that incorporating physically relevant exogenous variables significantly enhances strain prediction accuracy and provides reliable support for bridge condition assessment and early warning applications.