Classifying and Forecasting Traffic Incident Duration Using Various Machine Learning Techniques

Abstract

Non-recurring events, such as crashes, construction zones, adverse weather conditions, are considered the primary cause of traffic congestion. The duration of those incidents is significantly affected by multiple factors, including their type, severity level, and other conditions. Compiling incident-related data can provide valuable information for municipal planners and road users to aid in traffic management. Recently, machine learning models have proven their efficiency in predicting incident duration. This study develops machine-learning models to predict incident-duration classes, including artificial neural networks (ANN) and random forest (RF). Incident records data were retrieved from TranStar Houston's Transportation Management Center, and after cleansing, about 48,200 records remained with over 50 independent variables. The collected response data are divided into four classes, where events that lasted from 5 to 15, 15 to 30, 30 to 60, and 60 to 120 minutes are classified as minor, intermediate, major, and severe incidents, respectively. The findings of the predictive classification models revealed that RF scored the highest overall accuracy of 71%. The accuracy of predicting minor incidents using RF was the highest of 80%, followed by major (76%), intermediate (64%), and severe (43%) incidents. On the other hand, the accuracy of the ANN model followed the descending order of severe, intermediate, minor, and major incidents scoring 81, 75, 53, and 43%, respectively. The mean square error of ANN was reported as 28.4%, which is 10.4% higher than that of RFs. The assessment results showed that the most significant independent parameters on incident duration prediction are type and number of lanes blocked, type and number of emergency department responses, type of vehicles involved, and temporal factors. The outcomes proved the higher performance efficiency of classification predictive models using RF compared to ANN. Overall, the findings showed satisfactory performance compared to similar studies found in the literature.