Research On Taxi Demand Forecasting Based On Spatio-temporal Convolutional Network

Recently,with the progress of the domestic economy and urbanization,a massive population flows to cities,increasing the daily transportation demands of the city residents.A new transportation mode,online ride-hailing,has emerged due to the evolution of Internet technology and the prevalence of smart devices.The massive use of online ride-hailing has facilitated people’s lives while bringing the problem of traffic congestion in the meantime.Order demand forecasting is an important task for online ride-hailing,and its goal is to predict the order demand in the future period based on historical and current order demand data in the urban traffic road network.A highly accurate order demand forecasting model can help the online ride-hailing platform distribute orders and dispatch vehicles more reasonably,improve vehicle utilization,reduce passenger waiting time,and improve consumers’ trip experience.This dissertation compiles and analyzes the current state of domestic and international research on traffic forecasting and related knowledge and proposes a Spatial-Temporal Diffusion Convolutional Network（ST-DCN）model for predicting the order demand of online ride-hailing by analyzing the existing works and their shortcomings.The model can effectively model the spatio-temporal factors which affect the order demand of online ride-hailing and improve the accuracy of order demand forecasting.The main research works of this thesis are listed below.（1）A virtual station graph construction method based on the Density Peak Clustering algorithm is researched.In order to be able to better use graph neural networks to model the order demand prediction task,the idea of clustering is introduced into the graph construction method.The virtual stations identified by the clustering algorithm are used as nodes of the graph.A virtual station graph construction method based on the Density Peak Clustering algorithm is researched to realize turning the city into a real graph and improve the efficiency of the graph neural network in capturing spatial dependencies.Moreover,the constructed virtual stations graph conforms to the traffic road network structure in real scenarios,which helps the engineering application of the model.（2）A spatial convolutional network module based on two-phase diffusion graph convolution is proposed.Due to the ability of graph convolutional neural networks to model complex road networks,in recent years,graph convolutional network-based methods have been extensively employed in the area of traffic forecasting.In order to enable graph convolutional neural networks to capture spatial dependencies effectively,this dissertation researches a solution that can effectively alleviate two limitations of graph convolutional neural networks: over-squashing and over-smoothing.A two-phase diffusion graph convolutional network is designed in the dissertation,and an adaptive adjacency matrix is introduced into the network to capture the hidden spatial dependencies automatically.（3）A hybrid dilation convolution-based temporal convolutional network module is proposed.It is particularly important to model the time series when performing order demand forecasting tasks.In order to effectively capture the long-term time series relationships,this thesis uses a temporal convolutional network to capture the long-term temporal dependencies on the one hand,and a hybrid dilation convolution is used to overcome the grid effect problem on the other hand.Finally,the temporal periodicity factor is taken into account in the model to obtain more accurate forecasting results.This dissertation conducts comparative experiments on the proposed ST-DCN model with two large-scale real-world datasets.The experimental results show that the ST-DCN model can be effectively used for order demand prediction,and the forecasting accuracy is significantly higher than other same-type deep learning methods,especially when the forecasting time increases.