Predicting the duration time of incidents is important for effective real-time Traffic Incident Management (TIM). In the current study, the k-Nearest Neighbor (kNN) algorithm is employed as a nonparametric regression approach to develop a traffic incident duration prediction model. Incident data from 2008 on the third ring expressway mainline in Beijing are collected from the local Incident Reporting and Dispatching System. The incident sites are randomly distributed along the mainline, which is 48.3 km long and has six two-way lanes with a single-lane daily volume of more than 10000 veh. The main incident type used is sideswipe and the average incident duration time is 32.69 min. The most recent one-fourth of the incident records are selected as testing set. Vivatrat method is employed to filter anomalous data for the training set. Incident duration time is set as the dependent variable in Kruskal–Wallis test, and six attributes are identified as the main factors that affect the length of duration time, which are ‘day first shift’, ‘weekday’, ‘incident type’, ‘congestion’, ‘incident grade’ and ‘distance’. Based on the characteristics of duration time distribution, log transformation of original data is tested and proven to improve model performance. Different distance metrics and prediction algorithms are carefully investigated. Results demonstrate that the kNN model has better prediction accuracy using weighted distance metric based on decision tree and weighted prediction algorithm. The developed prediction model is further compared with other models based on the same dataset. Results show that the developed model can obtain reasonable prediction results, except for samples with extremely short or long duration. Such a prediction model can help TIM teams estimate the incident duration and implement real-time incident management strategies. First published online 28 January 2015.
展开▼
