Research And Implementation Of Dynamic Traffic Prediction Methods Based On Process Neural Network Ensemble

Traffic congestion is an open problem that affects the whole world. Intelligent Transport System (ITS) is an effective way to enhance traffic performance and ease congestion. Short-time traffic prediction is a critical part which provides primary information for the typical applications like Control and Guidance, Traffic Time Prediction. However, as the traffic system is a complicated system consists of 'user', 'vehicle' and 'road network', it has some particular characters like high complexity, non-linear and non-deterministic. The research on real-time traffic prediction with high accuracy and reliability is not only challenging but also of practical and academic value.The shortcomings of conventional traffic prediction methods can be summarized as: 1) not taking traffic flow's special characters like chaotic, procedural and multi-status into consideration; 2) higher generalization error and limited accuracy due to ignoring multi-status character; 3) lack of adaptation and dynamic handling abilities because of omitting the variance of traffic.To overcome the shortcomings of conventional methods, this paper provides traffic prediction methods driven by dynamic data using Process Neural Network (PNN) Ensemble, and proved the effectiveness of these methods by experiments on PeMS real traffic data. The research content includes:1. Chaotic traffic time series prediction based on phase space reconstructionWe proved that there is chaos in PeMS traffic data using chaotic time series analysis, recovered the chaotic attractor using C-C method and phase space reconstruction. By mapping the time series into a high dimension phase space of topological isomorphism and constructing a prediction model using Process Neural Network, we can predict the chaotic time series in the chaotic phase space. The experiment on PeMS data indicated that this model made good use of the chaotic information and enhanced the accuracy effectively.2. Dynamic traffic prediction based on Process Neural Network Ensemble In the purpose of enhancing generalization performance and increase predict accuracy, a Process Neural Network Ensemble model is constructed and thus local PNN can be trained accordingly to the diverse traffic status. In order to handle the variance of traffic data, we provide on-line learning and adaptive updating methods of the ensemble model, as well as dynamic result merging algorithm. The experiment on PeMS real traffic data indicated that Process Neural Network Ensemble model overcomes the low generalization shortcoming of single predict model and increase the prediction accuracy notably.This research is supported by "National Natural Science Foundation of China"(No.60703066) and "Project of China Transportation Alliance".