Short-term Traffic Flow Prediction Based On SVR And Multivariable Phase Space Reconstruction

Real-time and accurate short-term traffic flow prediction is the premise and key ofintelligent traffic control and guidance system, and prediction accuracy is directlyrelated to the operating results of traffic control and traffic guidance. Due to the random,time-varying and strongly nonlinear characteristics of the traffic system, thereforeartificial intelligence methods attract more and more people’s attention. Support VectorMachine (SVM) is a machine learning method based on the structural risk minimizationand statistical learning theory, and it can effectively solve small sample, nonlinearity,high dimension and local minima problems. In addition, some studies show that trafficflow has chaotic characteristics. Therefore, appling the SVM and chaos theory toshort-term traffic flow prediction is particularly important. Firstly, this papersummarizes the research situation of short-term traffic flow prediction models; secondly,analyze chaotic characteristics of traffic flow data; finally, propose a short-term trafficflow prediction model based on support vector regression (SVR) and multivariate phasespace reconstruction (MPSR-SVR model). The main innovation of this model isreflected in the design principles, namely the model is based on multivariate time series.The main research work are as follows:①Based on PeMS12.3database, this paper analyzed basic parameters of thetraffic flow(traffic flow, occupancy and average speed), introduced the traffic flow datapreprocessing methods, and completed the preprocessing of the measured traffic flowdata: missing or erroneous data’ pre-processing, noise reduction.②On the basis of an overview of chaos theory, the paper described multivariablephase space reconstruction theory and completed the experiment which using trafficflow data after pretreatment. As a result, get the embedding dimensions and delay timesof traffic flow, occupancy, and average speed time series, and complete multivariablephase space reconstruction.③On the basis of analyzing chaotic characteristics of traffic flow and itsidentification method, this paper calculated the largest Lyapunov exponents of thetraffic flow, occupancy and average speed time series, and the results verified the threesequences have chaotic characteristics. ④Based on the above theorys, this paper used genetic algorithm to determineparameters of support vector regression, constructed the MPSR-SVR model, proposedtraffic flow forecasting process and the prediction evaluation index. Finally, this paperapplied the measured data to test and verify the model, and compared the results of themodel based on SVR and univariate phase space reconstruction (UPSR-SVR model).The experimental results show that: MAE, MRE and MSE of the MPSR-SVRmodel which is proposed in this paper are less than the UPSR-SVR model’ results. itcan be seen that the MPSR-LSSVM model is superior to the UPSR-SVR model,whichcan effectively carry out short-term traffic flow prediction.