Theory And Method Of The Incident Duration Prediction

Nowadays, traffic congestion has become more and more serious because the traffic volume continues to increase. Incidents (vehicle accidents and disablements) are a major cause of traffic congestion in urban areas. In the case of recurring congestion, traffic management centers can provide travelers with the road traffic information according to the analysis of history data. Effective traffic incident management requires a full understanding of the various corresponding properties to accurately estimate incident durations and to benefit decision makings of reducing the impact of non-recurring incident relevant congestions.In this paper, Multiple Linear Regression, data mining, discrete choice model and other aspects of knowledge are employed to build incident duration model. All 62491 incident records from Beijing Transportation Management Bureau are used for model establishment and another 8000 records for validation.Firstly, Multiple Linear Regression model is used to construct the Functional relation between the characteristics and duration of incident. When the new accident occurred, the regression model will calculate the duration of the incident. Another 8000 data are used to the error analysis and the average relative error of the CART model is 31.4%.Secondly, decision tree methods are applied to forecast the duration of traffic incident. CHAID tree and CART tree are employed to establish the incident duration models. Another 8000 data are used to the error analysis. The average relative error of the CHAID and CART model is 30.8%and 29.5%.Thirdly, this paper incorporates discrete choice theory in incident duration prediction. There are two different method of discrete choice model. This paper applies the Multinomial Logit (MNL) model and Ordered Probit Model to estimate the traffic incident duration. Another 8000 registrations are used to test the validation of the model. The average relative error of the Multinomial Logit and Ordered Probit model is 27.2% and 30.89%.Finally, since the prediction accuracy of models above is different in different loops and radiation, we use the AHP (Analytic Hierarchy Process) to incorporate the three methods described above into the forecast system. Though comparing the error of different forecasting methods, we establish the matrix, calculate the coefficient of each method and give the prediction value of the model. Another 8000 data are used to test the validation of the model. The average relative error of AHP model is 26.43%.The results show that this AHP model significantly improves the accuracy of the prediction model.