Uncertainty-Aware Pavement Roughness Forecasting Using Adaptive Conformal Prediction

Document Type

Article

Publication Date

1-1-2025

Abstract

Accurate forecasting of future pavement conditions is essential for effective asset management and enables data-driven maintenance planning. While recent studies have demonstrated the potential of Artificial Intelligence (AI)-based methods to predict future pavement roughness, their application remains limited due to challenges in quantifying uncertainty. Most machine learning models rely on point-based accuracy metrics, which fail to capture performance variability across the full range of pavement conditions—particularly underrepresented or extreme scenarios where prediction errors tend to increase. To address this limitation, this study applies an adaptive conformal prediction—a forecasting framework that integrates seamlessly with AI algorithms to generate prediction intervals with targeted confidence levels. Specifically, we enhance the conventional conformal framework by developing a quantile-based adaptive strategy that dynamically adjusts interval widths according to the difficulty of each prediction. The adaptive Jackknife + approach consistently outperforms other methods, delivering 90.3% empirical coverage and an efficiency score of 0.588 at the 90% confidence level, while reducing the average interval width by more than 20% compared to its regular counterpart. Conditional-coverage analysis across pavement, traffic, and climatic features further demonstrates that the adaptive framework maintains stable and well-calibrated uncertainty across diverse operating conditions.

Publication Title

International Journal of Pavement Research and Technology

Link to Full Text

Share

COinS