Transformer-Based Temporal Modeling for Network-Level Pavement Condition Index Prediction: A Data-Driven Component for Pavement Asset Management

Accurate prediction of PCI is important for highway asset management and maintenance planning. Improved forecasts may support sustainable asset management, but this study does not quantify environmental or economic benefits. However, PCI time series exhibit complex temporal patterns—including nonlinear degradation trajectories and abrupt declines that may be associated with freeze–thaw cycles and heavy traffic loads; however, these factors are not measured in this study. To address this gap, this study proposes a customized Transformer-based framework for network-level annual PCI prediction. The model employs a fixed 3-year sliding window and integrates three key innovations: low-dimensional feature expansion, enhanced trend-weighted positional encoding, and lightweight encoder stacking, collectively balancing predictive accuracy with computational efficiency. On the 2024 test year and the 25,258 non-maintained segments from the 2020–2024 Xinjiang dataset, the proposed framework achieves R2 = 0.945 (95% CI: 0.941–0.949), MSE = 9.39 (95% CI: 9.12–9.68), and MAE = 2.30 (95% CI: 2.24–2.36). It reduces MSE by approximately 54.7% compared with the best-performing conventional baseline (Random Forest) and improves R2 by approximately 6.4 percentage points compared with Random Forest. This is a next-year forecast for previously observed segments under a fixed data window, not evidence of long-horizon temporal generalization. Ablation studies across multiple random seeds show small but consistent improvements, with lightweight encoder stacking yielding the largest average gain. Robustness analyses further demonstrate resilience to noise, missing data, and detection errors, underscoring its practical reliability for large-scale maintenance planning. By improving the accuracy of natural degradation prediction, the proposed framework offers a data-driven basis that could provide a prediction component for future maintenance decision models and life-cycle assessment. However, the direct environmental and economic benefits have not been evaluated in this study and remain to be validated through integration with maintenance decision models and life cycle assessment.