Decoupling Key Factors of Asphalt Pavement Rutting and Predicting Rutting Evolution Based on a Deep Temporal Model

Abstract Accurate rutting prediction is essential for rational pavement design and proactive maintenance, yet mechanistic–empirical models require intensive calibration with limited generalizability, and purely data-driven models offer little mechanistic insight. This study proposes a hybrid deep learning framework that integrates long short-term memory and Transformer architectures with a two-stage residual-based decoupling strategy to jointly analyze and predict asphalt pavement rutting evolution. Using 7 years of multisource monitoring data from the Beijing full-scale circular track, the framework quantitatively isolates the independent contributions of cumulative traffic load and air temperature: load governs the long-term growth trend, whereas temperature modulates seasonal fluctuations. A t -test confirmed that the load-only model residuals were significantly correlated with temperature ( t = 5.361 and p < 0.001 ); after temperature compensation, this correlation became statistically nonsignificant ( t = − 1.480 and p = 0.156 ), confirming effective factor separation. Prediction experiments showed that rut depth paired with temperature yielded the best accuracy [root-mean square error ( RMSE ) = 0.551 mm ], and load paired with temperature provided comparable performance when historical rutting data were unavailable. The proposed framework offers a physically interpretable approach to rutting prediction validated under controlled full-scale test conditions.