Abstract
Pavement management systems (PMS) require reliable forecasts of roughness progression to support maintenance prioritization. This study develops a deep learning–based workflow for forecasting Mean Roughness Index (MRI) deterioration and converting forecasts into maintenance-relevant urgency outputs. Using 4,756 observations from 374 maintenance-free Long-Term Pavement Performance (LTPP) sections, a simplified NBeatsX-inspired neural network (NBeatsX-NN) is benchmarked against classical machine-learning and statistical baselines. The model is used as a compact feedforward forecasting component rather than as an implementation of the canonical NBeatsX architecture. Performance is assessed with practical validation protocols, including grouped holdout splitting, LOSO-ZS testing on unseen sections, and operational backtesting that withholds the final two years of each section. The proposed model demonstrates strong predictive accuracy (RMSE = 0.1422, R² = 0.9415) and strong operational performance in backtesting (RMSE = 0.0594, R² = 0.9903), while maintaining transferability in LOSO-ZS evaluation (mean RMSE = 0.0983 ± 0.1158). Forecast trajectories are extended as scenario-based projections up to a planning horizon of 30 years and translated into PMS indicators, including remaining service life (RSL) and rule-based urgency classes using a serviceability trigger of MRI ≥ 2.0 m/km. Uncertainty is incorporated through quantile forecasts (P10–P90) to derive threshold-exceedance probabilities, enabling risk-aware prioritization without full optimization.