Research On Short-term Traffic Flow Forecasting Method Based On SVM

Since the20th century, the transportation is an important industry of driving the development of the economy which has a direct impact on the healthy operation of the national economy. For decades, the transport exports have advanced many different ways adapting to their own actual situation of ITS for different cities to attempt to use the traffic information system, dynamic route guidance system and other methods to deal with the problems of urban transport. At the same time, accurate traffic flow predicting is the premise of ITS implementation of dynamic route guidance, taffic assignment, incident detection. So it has an important significance.After analyzing the characteristics of the traffic flow, according to chaos theory, studying the use of largest Lyapunov exponent to distinguish the predictability of traffic flow. Then the author uses the traffic flow time series phase space reconstruction to show the inherent laws of traffic flow for the later build data relationships. Finally the author chooses a different traffic flow data from the PeMS system to deal with experimental simulation, which proves that the choice of traffic flow data is indeed predictable.In this thesis, on the base of the principle of support vector machine and solving the regression s-SVR, after analyzing the adaptability of the SVR model for traffic flow forecasting, different kernel functions (radial kernel, mixed kernel) and kernel parameters on the performance of SVR, the author constructs a traffic flow forecasting model based on the phase space reconstruction and SVR. Firstly the author uses the grid method (GS), particle swarm optimization (PSO) as well as improved particle swarm optimization (IPSO)for optimal selection of model parameters; secondly the author constructs GS-SVR, PSO-SVR, IPSO-SVR prediction model; thirdly the author constructs preediction model based on BP neture network. Finally, the author selects the working days and holidayes traffic flow measured data form the PeMS to do with the experimental simulation for GS-SVR, PSO-SVR, IPSO-SVR and BP models, then compare and analysis the resultes of the forecasting model performance. The resultes show that:the model based on BP network is weaker than the one baesd on SVR; in the same kernel function, the SVR model based on the IPSO relativing the SVR model baesd on PSO and GS has the better prediction performance; in the same parameter optimization algorithm the SVR model based on the mixed kernel function has the better prediction performance.However the SVR model baesd on the IPSO takes a long time, the real-time is not good. In this thesis, on the base of analysising the integrated learning principles and the SVR modle, the author focuses on the traffic flow forecasting model that integrates Bagging algorithm and Boosting algorithm and SVR model. For the fusion model I select holidays and working days traffic flow measured data for experimental simulation and compare with the single SVR. By experimental simulation, as opposed to a single SVR model, the integration of SVR can also get a relatively good prediction result, while its real-time is good.