Traffic State Prediction Based On Ensemble Learning

With the rapid increase in car ownership, the contradiction between supply and limited road resources and the growing number of cars become more and more acute, making it impossible balance between traffic supply and traffic demand. This leads directly to major cities worsening traffic environment, increasingly severe traffic congestion, so rational planning of residents daily travel admits of no delay. The key problems of traffic control and guidance is real-time and accurate traffic state forecasting. However, how to accurately detect traffic incidents, how to real-time traffic state identification and how to predict the traffic state accurately and efficiently, has been the priority among priorities of research in the field of intelligent transportation, which has important theoretical and practical significance to research on these issues.Analysis of the four traffic state pattern classification methods, including fundamental diagram method, three-phase traffic flow method, service level method and ensemble learning method. It focuses on cluster analysis method and distance measurement method in ensemble learning. Traffic state is embodied by traffic flow data, traffic flow data mainly includes parameters such as speed, flow and occupancy, according to traffic state similarity principle, traffic state can be determined by calculating the distance between traffic flow data.This dissertation relies on the subsection (Regional transportation intelligent analysis and multi-modal control technology for travel behavior) of National 863 Plan (Key technologies for transportation cooperative control in metropolitan area) to research "traffic state identification and prediction". The main contributions of this thesis include the following aspects:Essentially traffic incident detection problem is a binary classification problem, for the low detection rate and the high false alarm rate caused by the traffic incident, the anomaly traffic state detection model can not meet the problems of practical application, we proposed two traffic incident detection method: Traffic Incident Detection Based on Random Forest and Multiple Naive Bayes Classifiers Ensemble:the former base classifier is decision tree, with the change of the number of decision tree, to adjust the performance of incident detection; the latter base classifier is naive bayes, five different integration rules were applied for incident detection.Analysis of the essence of traffic state identification, which can be understood as the process of the classifier learning historical traffic flow data of different class labels and judging real-time traffic flow data of different traffic state level. According to the Performance Index system of urban road traffic in Beijing City, introducing dynamic classifier ensemble theory, we proposed a traffic state identification method based on the nearest neighbor rule integration. This method first calculated traffic state identification rate of each base classifier sample in the neighborhood of the test set, and then picked out of the highest local accuracy of base classifiers and judging of the traffic flow test data set, the output information is the traffic state level. In the performance evaluation, introducing the concept of traffic state confusion matrix, it carried out statistical analysis of the actual traffic state and the identification state respectively.From two aspects of traffic state variables and traffic state classification, respectively realized the city road traffic state prediciton. The traffic state level prediction problem transformed into multi-class classification problem. In the construction of data sets, future traffic state label uesd as historical traffic state label. Division of the training sample space using clustering technology, which can choose the better ability of regional for traffic state prediction. The effectiveness of this method validate by field data. Traffic state variables prediction are similar to other prediction problem, can be transformed into a machine learning problem, namely given one or several training data sets, based on the best fitting or other principles for training one or more models, and then using trained models for forecasting traffic state. On the basis of this theory, we proposed a traffic volume prediction methods based on data disturbance classifiers ensemble, The parameters of this method does not rely on the traditional default build neural network, which is based training results of the neural network model on actual traffic flow data.According to the practical application of identification and prediction of city road traffic state, based on this paper proposed traffic incident detection, traffic state identification and prediction method, we designed and development of the ensemble learning-based traffic state identification and prediction platform. Real time platform running stable and smoothly, data analysis is reasonable. To provide a strong information support and guarantee for the traffic managers and travelers, but also the practicability and validity of the research results of this paper proved from another aspect.