Research On Bayesian Network Modeling For The Prediction Of Traffic Incident Duration

Traffic incident is one of the main factors that leading to traffic congestions. Through controlling methods such as real-time traffic guidance, its impacts on traffic operation can be reduced. Accurate prediction of traffic incident duration is the prerequisite for effective control.Based on the research status, a more scientific prediction model is constructed to explore the relationship between the factors that influence the incident duration. Using MIT scoring functions, a ACOS algorithm to build the Bayesian network is approached. The networks are generated from top to bottom with ant colony algorithm to search the optimal network from the feasible set. To increase robustness of a Bayesian network, a random node selecting mechanism and a probabilistic choice model for local structure are introduced.According to the parameters of Bayesian Networks, we construct an improved Bayesian decision tree model with BNDT algorithm as the core, which introduced the internal mechanism of nodes into the Bayesian decision tree. The classification mechanism is started by adding a node type determination rule, which is based on CART algorithm. It combines the naive Bayesian theory with the Bayesian network inference method considering the dependency between attributes, which is more close to the actual situation.In the empirical study, the Bayesian network prediction model is built based on the traffic duration data which are collected by transportation department in Dutch. The Bayesian network reveals the duration mechanism among the attribute nodes and provides structural and quantitative results. In terms of the complete and incomplete observation node properties, the precision of the network prediction model is up to 76.97% and 93.23%, and the average precision is up to 87.82%. Using 20 minutes’error as the acceptable precision of prediction, the accuracy of the Bayesian classification model is 64.85% and 63.64% respectively under the complete and incomplete information. While the modified Bayesian decision tree model can give the prediction of average duration of each branch. In the case of complete information, the accuracy is 77%. As for vibrational degrees of information missing, the minimal precision is higher than 70%. The result indicates that Bayesian network model and the modified Bayesian decision tree model are both applicable and effective approaches to forecast the duration of a traffic incident.