Predicting freeway incident duration using machine learning

Accurate prediction of freeway incident duration is important for traffic management and emergency resource allocation. This study evaluates 32 machine-learning model configurations using 828 incident records from the G2504 Hangzhou Ring Freeway in China. Four fixed feature sets containing the top 6, 11, 16, and 21 predictors were defined from a Minimum Redundancy Maximum Relevance (MRMR) ranking. Each feature setting was evaluated using a 15% hold-out test set, seven repeated random splits, and sevenfold cross-validation within the training subset. All 32 machine-learning models were tuned within the training subset using sevenfold cross-validation. To provide a concise description of the tuning procedure, the selected Medium Gaussian support vector machine, which provided the most balanced overall performance, is reported as a representative example. Performance was assessed using the mean absolute error (MAE) and the mean absolute percentage error (MAPE). The Medium Gaussian support vector machine provided the most balanced overall performance, and prediction accuracy generally improved from 6 to 16 predictors, with little additional benefit from 21 predictors. Compared with classical linear regression, the Medium Gaussian support vector machine reduced MAE by 5.7%–9.7% and MAPE by 21.2%–31.9% across the four feature settings. A sensitivity comparison showed substantially higher errors when incidents longer than 60 minutes were retained, reinforcing that the main findings apply to routine incidents. Interpretation analyses identified Response Time and severity-related variables as important predictors, while residual errors remained larger for longer and more complex incidents. The model is intended primarily for post-arrival updating and clearance-support decisions for routine freeway incidents, rather than prediction at the initial alarm stage.