Presentation Information
[NpVBMPAo-11]IRI Prediction Modelling for Brun-Way: Statistical and Machine Learning Approaches for Forecasting Pavement Roughness
*Hunter Song1, Jean-Sébastien Lacombe1 (1. AtkinsRéalis (Canada))
Keywords:
IRI,rutting,asphalt pavement,deterioration modelling,Monte−Carlo,XGBoost,LTPP,lifecycle planning,cold−region highways
Background. Performance−based highway OMR programs require reliable forecasts of network condition to optimize lifecycle interventions, particularly in cold−region corridors with heavy truck traffic. This study develops and compares data−driven approaches to predict asphalt pavement performance, using a hybrid statistical and machine learning framework.
Methods. We analyze Brun−Way’s 275 km, four−lane highway network (New Brunswick, Canada) using two−yearly lane−level IRI and rut observations. The measurement data are taken at 100m resolution from 2012 to 2024, and further augmented through treatment labeling and engineered deltas (ΔIRI, ΔRut). We estimate natural deterioration with (i) regression families: constant, linear, quadratic, and (ii) empirical lookup and MonteCarlo models that map IRI→ΔIRI using historical distributions. In parallel, we train XGBoost regressors on (a) a cold−region subset of FHWA's LTPP flexible−pavement records and (b) Brun−Way’s own history (predicting ΔIRI), with features including multi−year IRI lags, traffic, climate (freeze index, temperature, precipitation), thickness and lane type terms. Models are back−tested using the 2012−2024 time period distributions, and sensitivity analyses explore traffic, climate, and thickness perturbations.
Results. Annual IRI growth depends strongly on current IRI and lane traffic but shows weak dependence on pavement age. Network−level forecasts indicate quadratic and empirical−lookup models best match future IRI distributions; MonteCarlo captures realistic uncertainty envelopes. ML models reproduce individual−segment histories with low error but underperform statistical models at predicting the full network distribution. Treatment modelling suggests a Mill and Fill rehabilitation treatment drives post−treatment IRI toward ∼0.8 and rut toward ∼3 mm when pre−treatment rut > 3mm, whereas microsurfacing treatment has limited IRI effect but reduces rutting.
Conclusions. For portfolio−level planning under a performance−based contract, calibrated statistical models is effective in providing reliable distributional forecasts, while ML models add value for individual segment−level trajectories and scenario sensitivity.
Methods. We analyze Brun−Way’s 275 km, four−lane highway network (New Brunswick, Canada) using two−yearly lane−level IRI and rut observations. The measurement data are taken at 100m resolution from 2012 to 2024, and further augmented through treatment labeling and engineered deltas (ΔIRI, ΔRut). We estimate natural deterioration with (i) regression families: constant, linear, quadratic, and (ii) empirical lookup and MonteCarlo models that map IRI→ΔIRI using historical distributions. In parallel, we train XGBoost regressors on (a) a cold−region subset of FHWA's LTPP flexible−pavement records and (b) Brun−Way’s own history (predicting ΔIRI), with features including multi−year IRI lags, traffic, climate (freeze index, temperature, precipitation), thickness and lane type terms. Models are back−tested using the 2012−2024 time period distributions, and sensitivity analyses explore traffic, climate, and thickness perturbations.
Results. Annual IRI growth depends strongly on current IRI and lane traffic but shows weak dependence on pavement age. Network−level forecasts indicate quadratic and empirical−lookup models best match future IRI distributions; MonteCarlo captures realistic uncertainty envelopes. ML models reproduce individual−segment histories with low error but underperform statistical models at predicting the full network distribution. Treatment modelling suggests a Mill and Fill rehabilitation treatment drives post−treatment IRI toward ∼0.8 and rut toward ∼3 mm when pre−treatment rut > 3mm, whereas microsurfacing treatment has limited IRI effect but reduces rutting.
Conclusions. For portfolio−level planning under a performance−based contract, calibrated statistical models is effective in providing reliable distributional forecasts, while ML models add value for individual segment−level trajectories and scenario sensitivity.
