Sequential Prediction for Large-Scale Traffic Incident Duration: Application and Comparison of Survival Models

A quick and accurate traffic incident duration prediction could greatly facilitate traffic incident management. However, at the very early stage of an incident, limited information is available for prediction. Information gathering for large-scale traffic incidents is a chronological process when a multi-agency response is required. At the early stage, information such as incident start time and roadway and weather conditions may be available, but information about response agencies and incident management solutions (e.g., lane closures) remains unknown. The objective of this study is to develop a sequential prediction method to handle the chronological process of incident information gathering. The method is based upon parametric survival modeling, which is often utilized to predict incident duration. This study took advantage of a unique incident database and identified over 600 large-scale incidents in the East Tennessee area from 2015 to 2016. A five-stage prediction method is proposed according to the chronological process by which information becomes available during incident operations. Using the data, this study compared three survival models: frailty model, multilevel mixed-effects model, and finite mixture model. Generally, with more information becoming available for modeling from the first to the last stage, the models' performance improved according to the root mean square error and mean absolute percent error. The finite mixture model outperforms the other two models and its mean absolute percentage error is between 10% and 15%. Incident-associated factors at each stage are discussed and implications based on the study outcomes are also covered in the paper.