Predicting Origin-Destination Ride-Hailing Demand With A Novel Generative Adversarial Network

With the popularity of mobile internet and shared travel,ridehailing has gradually become one of the important components of urban transportation system.Accurate demand prediction of ride-hailing can improve the service of ride-hailing and balance the traffic supply and demand.Many existing studies have made great achievements in total demand prediction based on traffic analysis zones.However,origindestination(OD)demand attracts a little attention,even though it carries much information and is of great practical significance,e.g.,it facilitates the routing and matching of ride-hailing services.With the rapid development of deep learning technology,new deep learning models can more effectively predict complex and nonlinear demands in traffic systems.In addition,reasonable spatial-temporal dependences information can improve the accuracy of OD demand prediction.Therefore,this thesis considers the spatial-temporal dependences of OD demand and proposes a new generative adversarial network model(GAN)to conduct short-term OD demand prediction for online ride-hailing.The main contents of this thesis are as follows:(1)Establish a new combined GAN model to predict ride-hailing OD demand.In this thesis,a new GAN model,named CWGAN-div,which combines conditional GAN(CGAN)and GAN with Wasserstein divergence(WGAN-div),is established to predict short-term OD demand of online ride-hailing.The CWGAN-div model has the structure of CGAN,so the spatial-temporal dependences information is taken as the conditional information to guide the model to generate the prediction results more accurately.At the same time,the model adopts the objective function of WGAN-div,which improves the stability and convergence of the original GAN.(2)Propose four methods to measure spatial dependences,and construct a model structure which integrates spatial-temporal information.In order to improve the prediction accuracy,spatial-temporal dependences are added into the prediction model.Considering OD demand has trendency and periodicity,historical OD demand and Timestamp*Day time information are taken as input of CWGAN-div model.In addition,four spatial correlation matrices,namely distance matrix,adjacency matrix,historical trend correlation matrix and Moran’s I correlation matrix,are proposed as the input of CWGAN-div model from geography and demand data aspects.In addition,residual neural network is introduced to process high dimensional demand data and spatial information in the generator and discriminator of CWGAN-div model.(3)Discuss the influence of different temporal and spatial dependences and conduct comparative experiments of other prediction models based on empirical experiments.An empirical study on OD demand prediction is conducted based on the ride-hailing ride data of Manhattan,New York City.The optimal prediction model is established by discussing the influence of different temporal and spatial dependences.Then,the analysis of the prediction results and the convergence of the optimal model are carried out.In order to verify the performance of the proposed model comprehensively,three groups of comparative experiments are as follows: the prediction performance of CWGAN-div model and traditional prediction models,the prediction performance of CWGAN-div model and common combined deep learning models,and the prediction performance of CWGAN-div model and basic GAN models.The experimental results prove that the proposed model has the best performance and is competent for predicting OD ride-hailing demand.