Machine learning-based incident duration prediction integrated with microscopic traffic simulation for expressway...

Traffic incidents are a primary contributor to non-recurrent congestion on urban expressway networks, generating economic, safety, and operational impacts. Although machine learning has shown promise for incident duration prediction, a gap remains between predictive modelling and practical integration into traffic management workflows. This study proposes an integrated resilience-based decision-support framework that couples machine learning-based incident duration prediction with microscopic traffic simulation to support evidence-based incident management. A dataset of 2,772 incident records from the Expressway Authority of Thailand (EXAT) was used to train and evaluate four tree-based ensemble models: Random Forest, XGBoost, CatBoost, and LightGBM. LightGBM achieved the best test performance, with a test R² of 0.506 and a mean absolute error (MAE) of 8.182 minutes for short-duration incidents of 0-40 minutes. The predicted incident duration was then used as an input to Aimsun Next simulation on the Chalong Rat Expressway, where four incident management strategies were assessed using the speed deficit area metric. To account for prediction uncertainty, three incident-duration cases were evaluated: 22, 30, and 38 minutes. The 30-minute case represents the LightGBM-predicted duration, while the lower- and upper-bound cases were defined using the short-duration MAE of approximately 8 minutes. The results show that reversible lane activation produced the strongest resilience performance across all tested duration cases, with the lowest deficit areas of 588, 809, and 1,038 km·min/hr. In contrast, shoulder-only restriction produced the poorest performance, with deficit areas of 1,376, 2,281, and 3,233 km·min/hr. Overall, the proposed framework demonstrates how incident duration prediction can be translated into simulation-based resilience assessment, providing traffic management centres with a structured, quantitative, and data-driven basis for comparing incident response strategies.