Research And Application On The Traffic Flow Forecasting Technology Based On Machine Learning

With the speeding up of urbanization,there is an urgent need for the intelligent transportation system.As an essential part of ITS,short-term traffic flow forecasting has also made a rapid advancement.The problem of how to improve the forecast accuracy and ensure the efficient operation of ITS is still the emphases on which scholars focused.Machine learning algorithm provides a new idea for short-term traffic flow forecasting.Useful information hide in the complicated nonlinear traffic flow data can be unearthed based on the idea of data driven.The characteristics of traffic flow data have been analyzed in this dissertation.In view of the traditional traffic flow forecasting model's shortcoming of not being able to cope with the complex changes in traffic flow.Machine learning methods were applied into traffic flow forecasting and a traffic flow forecasting system was designed and implemented.The main contribution of this dissertation is as follows:1.Feature Selection: this dissertation focuses on the short-term traffic flow forecasting of multiple sections,considering the adjacent sections and impact factors such as the weather conditions,so there will be more features than the traditional one.Nevertheless,the conventional artificial feature selection method is inefficient and errorprone.This dissertation proposes an improved adaptive feature selection model based on random forest(LOO-RF),which can automatically selection importance features of traffic flow.Experimental results show that both the accuracy and efficiency of traffic flow forecasting are improved after the adaptive feature selection process.2.Short-term traffic flow forecasting model: this dissertation applies the support vector regression model of machine learning into the short-term traffic flow forecasting of multiple sections.The impact of white noise in the traffic data has been reduced by using the Gauss kernel function.Meanwhile,in view of the difficulty of SVR parameter optimization,the dissertation proposed an improved genetic algorithm based on chaotic theory(CGA).This algorithm can quickly find the optimal parameter combination of SVR.An improved traffic flow forecasting model named CGA-SVR is proposed by combining the two algorithms above.Experimental results show that the forecasting accuracy of this model is better than the traditional models.3.Design and implementation of the traffic flow forecasting system: this dissertation design and implement the traffic flow forecasting system based on the LOO-RF feature selection model and CGA-SVR traffic flow forecasting model,using the highway toll data in a province.The system can display the prediction results and traffic data,the traffic flow forecasting work can be done through a simple parameter setting.