Short-term Traffic Flow Forecasting Based On SVR In Hadoop Environment

An accurate forecasting of short-term traffic flow is a prerequisite of intelligent traffic guidance as well as an essential comportment of Intelligent Transportation System(ITS). Traditional forecasting method consist of regression analysis, trend expansion and time series. However, a traffic system is a time-varying, non-stationary stochastic system which tends to be influenced by climatic conditions, driver psychological status, emergencies and accidents, and many other factors. So the traditional linear forecasting methods are not able to forecast the short-term traffic flow with an ideal accuracy and the theory of artificial intelligence enjoys its popularity among an increasing number of researchers.Support Vector Regression machine(SVR), a machine learning method based on statistical learning theory, can solve the nonlinear problem with a small sample set provided and has proved superior to neural networks and other methods in short-term traffic flow forecasting. This thesis analyzes the importance of parameter selection for support vector regression and the advantages and disadvantages of different parameter optimization algorithms. Then it selects the simulated annealing algorithm to optimize the parameters of SVR, and proposes a short-term traffic flow forecasting model based on the SA-SVR algorithm. Since the model takes full advantage of the minimum structure risk of SVR and the globally optimizing ability of SA, it can meet the requirement of real-time and get excellent forecasting accuracy. The mean forecasting deviation of the model is4.96%, with the maximum is9.81%.With the increase of the training sample size, the time and space complexity of the traditional SVR training algorithms increase dramatically. So these algorithms implemented on a stand-alone server often fail to run. This thesis elaborates the existed algorithms used to solve this problem, they are based on the idea of "iteration" and "parallel" separately. Then we combine MapReduce, an excellent distributed processing model, and the ideas mentioned above to develop a large-scale SVR training algorithm running on Hadoop platform and build a short-term traffic flow forecasting model based on SA-SVR and Hadoop. The speedup of this model is up to16.03. The mean forecasting error is5.20%, with the maximum is10.70%.