Analyzing Freeway Traffic Incident Clearance Time Using a Deep Survival Model

Abstract

Accurately predicting freeway incident clearance time and analyzing influential factors are essential for developing effective traffic incident management strategies. Existing approaches for analyzing traffic incident clearance time include statistical models and machine learning models. Whereas the statistical approach is able to quantify the impact of influential factors on incident clearance time, it often yields unsatisfactory levels of the prediction accuracy. Conversely, the machine learning approach lacks model interpretability but can generate accurate predictions. To combine the advantages of both approaches, a survival analysis model based on deep neural network (DeepSurv) is applied to predict the traffic incident clearance time. We used the SHapley Additive exPlanations (SHAP) method to interpret the modeling results of the DeepSurv model and analyze the impact of influential factors on traffic incident clearance time. Results show that the DeepSurv model outperforms statistical models (i.e., Cox proportional hazard, accelerated failure time and quantile regressions) and traditional machine learning models (i.e., support vector machine, random forest, and extreme gradient boosting algorithm) in terms of prediction performance. The analysis results indicate that response time, incident type (collision), lane closure type (all travel lanes blocked, total closure), involvement of fire and traffic control are significant influential factors affecting traffic incident clearance time. Our findings indicate that the proposed DeepSurv model is a more effective approach to predict traffic incident clearance time.