| Under the sharing economy,the more flexible ride-hailing can to some extent meet the personalized travel needs of the public.In the process of ride-hailing from inefficient competition to orderly development,accurately grasping the travel demand characteristics of users is the key foundation.However,due to the microscopic randomness and macroscopic regularity of personalized demand,the limited and spatio-temporal distribution of ride-hailing resources,ride-hailing scheduling becomes difficult.In the context of the internet of vehicles,this paper uses the trajectory data collected from the GPS of ride-hailing in different regions to analyse the demand and travel characteristics of ride-hailing,and proposes three deep learning models based on different data design methods.The main contributions of this paper are as follows.(1)A graphical spatio-temporal model for ride-hailing demand forecasting in functional areas is proposed.Firstly,the areas with more concentrated data in the datasets are divided according to the functional characteristics of the city.Secondly,the origindestination matrices are constructed using the demand between the functional areas.Then,the adjacency matrices based on distance and interaction type are constructed according to the topological characteristics of the road network.Fourthly,the temporal convolution module and the spatial convolution module are proposed for the long-term dependence of time and extraction of spatial correlation,respectively.Finally,the non-autocorrelation of the data is analyzed and predictions are made in terms of two temporal divisions: periodic and serial,respectively.(2)A graphical spatio-temporal meta-learning model(GST-Meta)for ride-hailing demand forecasting is proposed.Firstly,the collected temperature,weather,holiday,POI,traffic hubness,and COVID-19 data are pre-processed respectively.Secondly,GST-Meta is designed,which consists of a temporal convolution module,a meta-spatial convolution module,and a meta-temporal convolution module,where the meta-spatial and metatemporal convolution modules mainly consist of a meta gating fusion module and a meta learner,which are able to model the spatio-temporal correlation of traffic data by integrating external auxiliary information.Finally,the non-autocorrelation of the data is analyzed and the above data are directed to different processing modules of GST-Meta for prediction.(3)A tree spatio-temporal model for ride-hailing demand forecasting in functional areas is proposed.In the design of the tree matrix,firstly,the spatial relationships are abstracted into the connection relationships between nodes according to the divided functional areas.Secondly,each node is used as the root node to construct a planar tree matrix.Finally,each planar tree matrix is stitched together to form a final tree matrix with overall spatial relationships.In the design of the model,firstly,the temporal dependency strength between the demand data is obtained by the temporal attention convolution module.Then,the tree matrix is multiplied with the output of the first layer by the tree convolution module to obtain the result incorporating the spatial correlation.Finally,the above results are fed into the temporal convolution module to extract the higher order temporal correlations.Numerous experiments were conducted on multiple real-world datasets,including comparison experiments with various baselines,ablation experiments,and analysis of experimental results from both temporal and spatial perspectives,resulting in the lowest error metrics for the three methods proposed in this paper,which can more effectively and accurately predict the demand for ride-hailing in the future. |
