Multi-model optimized stacking ensemble framework for pavement performance: predicting key indicators and the Pavement Maintenance Quality Index

Accurate prediction of pavement performance is essential for effective highway management. Single-model approaches struggle to capture the complex deterioration mechanisms across highway networks. To address this limitation, this study develops an automated multi-model stacking ensemble framework for predicting PCI, RQI, and RDI, using data from the G22 highway between 2020 and 2023 across five regions. The methodology integrates six base ML models via stacking with four meta-learners and evaluates ensemble sizes ranging from two to six base models. Input features include historical indices, age, lane, direction, section length, historical distress and repair records, weather conditions, and traffic data. Stacking ensembles consistently outperforms individual base models across both test and validation datasets. To forecast PQI, predicted PCI and RQI are combined with historical index, age, lane, and section length. SHAP and LIME analyses of base-model predictions indicate that historical index, distress and repair records, pavement age, weather conditions, and heavy-traffic classes are the most influential features. The stacking models show reliable predictive performance, with over 93% of predictions within acceptable error bounds, indicating the potential utility of the ensemble framework in pavement management. Future work will apply the framework to additional highway networks to further assess generalizability and long-term predictive capability. An automated multi-model stacking ensemble is developed to predict PCI, RQI, and RDI, integrating six base models with four meta-learners.The stacking ensembles consistently outperform all individual base models, with systematic evaluation identifying optimal ensemble sizes.SHAP interpretability analysis reveals historical index, distress/repair records, age, and weather as the most influential features for PCI and RQI.The ensemble framework demonstrates robust performance, with 93% of predictions falling within acceptable error.The predicted PCI and RQI successfully contribute to forecasting the Pavement Maintenance Quality Index (PQI), showing enhanced stability and performance. An automated multi-model stacking ensemble is developed to predict PCI, RQI, and RDI, integrating six base models with four meta-learners. The stacking ensembles consistently outperform all individual base models, with systematic evaluation identifying optimal ensemble sizes. SHAP interpretability analysis reveals historical index, distress/repair records, age, and weather as the most influential features for PCI and RQI. The ensemble framework demonstrates robust performance, with 93% of predictions falling within acceptable error. The predicted PCI and RQI successfully contribute to forecasting the Pavement Maintenance Quality Index (PQI), showing enhanced stability and performance.