Traffic Flow Prediction In Free Charge Holidays Based On CEEMD-FNN-SVR And Enhanced KNN

In order to solve the increasingly serious traffic congestion problems,Intelligent Transportation Systems(ITS)has been widely used to alleviate traffic jams and improve the road efficiency in traffic management.Traffic flow prediction plays an important role in intelligent management and dynamic control,and it is the key to traffic guidance.A large amount of traffic data provides the data for analysis and prediction of traffic flow.Since the policy that free charge for cars under 7 seats during important holidays has been carried out from Oct.1st,2012.There has been a rapid growth in traffic flow during those holidays,which has made a serious traffic jam.Real-time and accurate traffic flow prediction contributes to analyze traffic condition,and plays an essentially important role in road network transportation planning and efficient control strategies designing.Based on the analysis of temporal and spatial characteristics for traffic flow in holidays,we propose an integrated forecasting model combined KNN with CEEMD-FNN-SVR.Firstly,we propose a modified KNN model with differential Euclidian distance to select related stations.Secondly we decompose the related traffic data into trend component and random component by CEEMD.Thirdly,we combine FNN model with support vector regression model to build FNN-SVR model for traffic flow prediction in holidays.The random component is predicted by FNN-SVR model,and the trend component is predicted by linear regression with GDP.We merge the trend prediction and random prediction to get final prediction result.Finally,the feasibility of the method is verified with real traffic data collected from 60 toll stations in National Day and May Day from 2011 to 2015.Through the simulation analysis,it is concluded that:(1)modified KNN could efficiently optimize training set for better prediction;(2)CEEMD is an efficient method for traffic data decomposition;(3)CEEMD-FNN-SVR model with KNN could make a believable prediction,and it has strong anti-interference ability.It can be found from the prediction results that,the prediction error of National Day is less than 10%,which reduced 10.2% and 1% compared with SVR and DFT-SVR model.The error of forecasting results of May Day reached 9.7%,which reduced 8.8% and 4.39% compared with SVR and DFT-SVR model.