Short-term Traffic Flow Prediction Based On Wavelet Neural Network

Wit h the fast-develop ing soc ial economy,the congest ion,frequent accidents and pollut io n problems in t he transport syste m become the fo cus o f t he society.As a n important part of inte lligent transportat ion syste m,traffic guidance systems and traffic contro l s ystem p lay a great role to improve t he reductio n of road congest ion,traffic accidents and energy consumpt ion.Rea l-time and accurate short-ter m traffic flow forecasting is the basis of t he traffic guidance systems and traffic control syste m.Therefore,the study of urban traffic system in short-term traffic flow forecasting has great value.In this thes is,the advantages and disadvantages of each methods using in short-term traffic flow are analyzed based on the traffic flow characterist ics.As traffic flow has the features of nonlinear,time-varying and uncerta int y character ist ics,this thesis select the wave let neura l network as traffic predict ion models combine the good characteristics of the neural network.A met hod to hand le the miss ing histor ica l traffic data and out lier has been proposed.By ut ilizing C-C met hod and cut-and-tria l to determine the struct ure of wavelet ne ural network,and do simulat ion experime nts by using rea l traffic data in US PeMS system.A wavelet ne ura l network forecasting mode l based on optimizat ion genet ic algor it hm has been proposed.As wavelet ne ural network is unstable and easy fall into local opt imum va lue,the improved genet ic a lgor it hm proposed can pre-optimize t he weights and parameter at wavelet neura l network.Genet ic algorit hm is good at global searching,so it was used to determine the global optimum range and comb ine wit h wavelet neura l network-specific local search speed to determine the global opt ima l solut io n.In order to prevent premat ur it y,this paper opt imize t he genet ic a lgor it hm fro m fit ness funct io n,crossover and mutat ion rate,and evo lut io nary strategy.The experiments show that the genet ic algor it hm opt imizat io n of wave let neura l networks is more stable.An error correction(EC-WNN)based short-term traffic flow predict ion model has been proposed.As the wave let neura l network could not mod ify a fter training and have the high error if it is fa ll into local optimizat io n.In order to solve this proble m,time-series mode le was used to establish real-time error predictio n model,and the wavelet neura l network was used to predict the real-t ime traffic flow.Besides,the predicted results are dyna mic adjusting by us ing the predict ion error.At last the experiments show that the predict ion model corrected by error is more stable and more accurate than the prediction based on wavelet neural network.