A Research On Travel Demand Prediction Based On Taxi Trajectories

Taxi is one of the most important traffic tools in the city.It can meet the demand of customized travel of citizens.However,there exist the micro randomness and macro regularity for customized travel,it is hard to dispatch for taxi company because the limitless of taxi resources and the spatial and temporal distribution of travel demand.Due to that,the dispatch problem of taxis for taxi companies has become the hot research issues in the area of Intelligent Transportation.With the development of technology,the invention of Internet of Vehicles provides technical means for the solving of this problem.On this background,this dissertation utilizes the data which is collected from the GPS equipped on the taxi to investigate the method to predict the destination of partial GPS trajectory data,which supports the evidence for the taxi dispatch.The GPS dataset used in this dissertation is from Porto in Portugal which is published on the Internet,the details of the work in this dissertation show below.First of all,this dissertation conducts an in-depth analysis for every feature for the format of the GPS data in Porto.Then the method of statistical analysis and visualization are used for cleaning the dataset.As a result,we get a number of 1700 thousand of cleaned taxi trajectories.Secondly,this dissertation introduces two methods of feature extraction,include transfer the trajectories into images and unify the length of trajectories.With these methods,this dissertation solves the problem of comparison problem which is caused by the varied length of trajectories.Based on these methods,this dissertation transfers the original trajectories.The result shows that the unified method can get higher performance than the image method.Thirdly,four kinds of GPS cluster algorithms and the mean-shift algorithm is finally chosen for the clustering of the destination points of trajectories.Based on this method,1700 thousand destination points are clustered,and finally get 3392 cluster points.This dissertation evaluates the quality of cluster points by visualization.In order to solve the problem of data sparseness,this dissertation changes the labels of trajectories with cluster points.Based on this method,the number of destination points drop from thousands to 3392.The sparseness problem is solved by this method.Finally,this dissertation designs a method to build classifiers which use the data from feature extraction and the classical MLP algorithm and KNN algorithm to predict the destination of trajectories.The result shows that the gap between the prediction value and the label value is 2.4,which proves that this method is effective for the prediction of destinations.This dissertation utilizes two kinds of feature extraction and two classifiers to evaluate the effectiveness on the real taxi GPS dataset from Porto,the result of the method in this dissertation has high prediction precision and can meet the demand of taxi company for the dispatch of taxis.