Research On Urban Short-term Traffic Flow Forecasting

Intelligent Transportation Systems(ITS)play an increasingly significant role in alleviating urban traffic problems and improving quality of experience for passengers.Depending on the prediction of futuristic traffic situations,ITS try to provide corresponding solutions for traffic problems and prevent traffic accidents.Particularly,traffic flow as the important parameter for analyzing the corresponding traffic situations has attracted a lot of attentions.Thus,during the recent three decades years,much relevant literature has focused on applying intelligent algorithms to accurately predict traffic flow in a short term.It is useful for promoting intelligent development for traffic systems.However,there are plenty of factors impacting the changes of traffic flow,such as weather,air quality and even the mood of users.So these uncertainties often make the traffic flow changing randomly and make it difficult for prediction works.In this thesis,we try to analyze these data features of traffic flow at first,and then prediction models are amended based on the corresponding results to more accurately capture the dynamic flow in the future.Our work is divided into two parts:At the first part,based on the data preprocessing to refine input information,the accuracy of prediction models can be improved.The main approach for short-term forecasting is exploring the relationship between historical traffic flow data and current or future traffic flow data.It is also useful for related applications in ITS.Traditionally,there are linear and nonlinear prediction models.For example,the history average model,ARIMA and SARIMA are the main linear models.As for these famous intelligent algorithms,such as Artificial Neural Networks,Support Vector Regression and Bayesian networks belong to nonlinear models.These prediction model are almost all trained by the historical data samples to predict changes of traffic flow under corresponding situations.Besides for as training sets,these data samples can be applied by researchers for exploring features hidden in traffic flow to amend the prediction models and improve the prediction accuracy.In this thesis,a datapreprocessing frame work is proposed under the stationarity and autocorrelation features existing in traffic flow.This framework can be employed to extract and choose effective features of input data,which can improve the accuracy of short-term traffic flow forecasting models.At the second part,a hybrid method is proposed in this thesis to construct the hybrid prediction model and decrease the generalization error for prediction models.Considering the complex features of traffic flow data result from the various factors,such as weather,environment and human behaviors,it is controversial to say which single model prototype can globally provide the best forecast performance among alternatives.In fact,the corresponding phenomenon can be observed,where these prediction models trained under specific traffic conditions are often inaccurate when they are applied in other situations.This phenomenon severely limits the application for most prediction models.In this thesis,with the Bayesian decision theory,a hybrid method is proposed to construct hybrid prediction models and decrease the impaction from changes of traffic situations.In this thesis three traditional prediction models SARIMA,Support Vector Regression and Artificial Neural Networks are amended based on these methods mentioned above.In order to verify the improvement for amended models,real passenger flow collected from the subway stations are used for simulation.From the simulation results,these amended prediction models can be more accurate than the original one respectively.Note that in all the simulation results,the hybrid prediction model based on the SARIMA can is the most accurate in most situations.