Research For Short-Term Traffic Flow Forecasting Method Based On Chaos And SVR

With the further research and development of Intelligent Transportation System (ITS) the major cities have undertaken the corresponding ITS strategic planning study for the city and even to inject new vitality into economic development.Traffic flow prediction is the key technology of the traffic information system for processing information in depth, and also is one of the core processing system of urban transport guidance which is to facilitate public travel, and also is one of the important application technology of traffic processing system and othe ITS subsystems. Because when the state of road traffic flow variability and complexity, it is difficult to give accurate analytical expression to describe the changes in law, so it is extremely significant for real-time and accurate traffic flow forecasting.In this paper, we use the method with a small dataset to calculate the chaotic featrues of Lyapunov exponent, which is a parameter to distinguish the chaos in traffic flow, the study of the short-term traffic flow time series phase space reconstruction is base on the chaotic time series analysis of traffic flow, the C-C method is used to calculate the reconstructed phase space embedding dimension m and time delay T,so the relationship of forecast data is established. To Simulate with the actual traffic flow, the results show that the short-term traffic flow is chaos and predictability.Researching the theory of the support vector regression (SVR), the applicability for limited samples and nonlinear traffic flow forecast with SVR is analysised. Researching the theory of the kernel function and parameters of the SVR model construction, the traffic flow time series phase space reconstruction and the SVR model are combined to construct the traffic flow model of single-point single-step prediction based on the phase space and SVR, respectively, the traditional grid search method (GS) and genetic algorithms (GA) presented in this paper are used to optimize the parameters of the SVR model. Using these two models to simulation with the observed data of working days and holidays from PeMS, the results show that the normalization of the source data processing can effectively improve the performance, and the prediction preformance of the SVR model with GA is better than the SVR model with GS.Optimization of SVR by genetic algorithm proves slow speed of operation of SVR training, it is not enough good to meet the real-time requirements of forecasting. Based on the ensemble learning theory, the thesis anaylyses the Bagging and Boosting method which both belong to ensemble learning. Then both of the methods are combined with SVR are used to the short-term traffic flow forecasting separately. Both of the models are used to simulate observed traffic flow, also compared and analysed with the GS-SVR and GA-SVR which both belong to single SVR. The results of simulation show the SVR ensemble method achieves better comprehensive prediction performance with less prediction time, which provides a thought and methods for the design of high-performance for real-time traffic flow prediction.