Rutting depth prediction model for asphalt pavements based on a dual branch spatiotemporal attention network

Existing asphalt pavement rutting prediction models suffer from large long-term prediction errors due to their reliance on laboratory parameters and simplified assumptions. To address this issue, a dual-branch spatio-temporal attention network model (DSAN) is proposed. The model is constructed by fusing temporal convolutional networks (TCN), long short-term memory networks (LSTM), and multi-head attention mechanisms to form parallel feature extraction branches for hierarchical spatio-temporal modeling. Validation is conducted based on full-scale pavement loop test results of eight typical asphalt pavement structures (AC layer thickness: 12–52 cm) under 80 million equivalent standard axle loads (ESALs) accumulated during 2017–2023. Results show that the DSAN model significantly outperforms comparative models in prediction accuracy, breaking through the generalization bottleneck of traditional models for different layer thickness structures. This study provides an efficient solution for long-term performance prediction of asphalt pavements.