Research On Short-Term Traffic Flow Prediction Of Expressway Based On Esemble Learning

As the data basis of active traffic management and control technology,short-term traffic flow prediction is a significant guarantee in the development of intelligent transportation system(ITS).To achieve the goal that a transportation with fine-grained perception,precision predition and sincerity service,the short-term traffic flow prediction with efficiency and accuracy could be a solid basis.Also,it is a guidance for travelers.This dissertation selected a basic section in G50 expressway as the research object.Modificating the error data and compeleting missing parts by data preprocessing technology.From the analysis of relationship in both time and spatial dimensions,the origin data set used for follow-up studies could be determined.Aiming at the non-linear realationship in original traffic flow sequence,this dissertation constrcted a feature selection model to select the historical data collected by detectors.The genetic algorithm and back propagation neural network algorithm were used for model sloving.Comprehensively considered the diversity of sample features and complexity of the model,regarding mean square error function as optimation goal of the selection model.Then the optimal plan was founded for short-term traffic flow prediction.This dissertation constructed three models by the aid of algorithms that back propagation neural network,support vector regression and decision tree.The mean square error and mean absolute error were used to evaluate the model performance in traffic flow prediction.Result that covers six days shows that the accuracy of each model has room for improvement still.Therefore,a new machine learning algorithm named Dendrite Net was applied to build an ensembled model which ensemebled the three models above refering to ensemble learning rule.A new model based on stacking ensemble rule was constructed and applied for short-term traffic flow prediction at different statistical intervals.Results shows that advantages of every basic model were remained.Furthermore,accuracy of the new ensembled model had an obviously raise compared with other models.In addition,the accuracy of new model and basic models reveals a close connection between them,when basic model got accurate results,ensembled model could also achieve that.With the expand of statistical interval,evaluation indexes became better gradually.Moreover,two treatment groups were applied to verify the validity of the feature selection model and short-term traffic flow prediction model which is based on stacking ensemble rule,particularly.Taking five-minute statistical interval as examples,the first test result indicates that the feature selection model for detectors’ historical data picked useful features efficiently.So that the prediction model performance improved.Another test result implicates that diversity of basic models,suitable integration method and a prominate algorithm may lead to a great promotion on predition accuracy.In other words,both models presented by this dissertation works well.At last,the conclusion was summarized and insufficient points in this research were proposed.Also,some ideas for further research were put forward.