Research On Demand Forecast Of Shared Bicycle Based On Spatio-Temporal Dependency

As a new mode of transportation,bike-sharing provides a convenient and lowcarbon travel solution for people.Bike-sharing can be divided into two modes: StationBased Bike Sharing(SBBS)and Station-Free Bike Sharing(SFBS).With station users are required to rent or return their bikes at fixed sites.Without stake,there is no such restriction.With more and more bike-sharing being put into the society,these two user-led models begin to expose the disadvantages of hindering traffic and uneven distribution.Therefore,it is urgent to accurately predict the demand of bike sharing,so as to provide the basis for the management and redistribution strategy of bike-sharing.The demand of bike sharing has complex Spatio-Temporal dependency and uncertainty.At present,the traditional deep learning methods only focus on the Temporal-Dependency modeling of bike sharing demand,but ignore the Spatial-Dependency and external factors between sites or regions.Based on the above shortcomings,this study combines Multi-Graph Convolutional Network(MGCN)and Temporal Convolutional Network(TCN)to explore the Spatio-Temporal dependency and external factors of bike sharing,constructs the demand forecasting model of bike sharing,and verifies the effectiveness of the model on the data sets of SBBS and SFBS.The main research contents and conclusions are as follows(1)Mining the Spatio-Temporal dependency of bike sharing demand.Based on the borrowing and returning records of Bay Wheels bike sharing in San Francisco and Mobike in Shanghai,the data sets of SBBS and SFBS for model training are constructed respectively,and the Spatio-Temporal dependency and external factors of bike-sharing demand of SBBS and SFBS are analyzed.The results show that: in terms of time dependence,there are approaching trend and daily periodicity for bike sharing in both SBBS and SFBS,while there is more obvious weekly periodicity for bike sharing with pile;in terms of spatial dependence,bike sharing with pile and without pile show similar local spatial dependence and global spatial dependence,and the local spatial dependence is in the morning and evening peak hours In addition,holidays and precipitation will reduce the demand for bike sharing.(2)The Spatio-Temporal prediction model of bike-sharing demand is constructed.Based on the space-time dependence of bike sharing,MGCN-TCN model is proposed to predict the demand of bike sharing.Firstly,the interaction graph and time series similarity graph are constructed to represent the local and global spatial dependence between sites or regions respectively,and the spatial dependence is extracted and fused by two graph convolution networks;secondly,three temporal convolution modules are used to capture the recent trend,daily trend and weekly trend of demand respectively,so as to analyze the Temporal dependency.Finally,the characteristics of external factors(holidays and weather)that may affect the demand of bike sharing are embedded through the fully connected layer to model the uncertainty of demand.The results show that: the MGCN-TCN model constructed in this paper has lower error on the both SBBS and SFBS dataset,and maintains higher prediction accuracy in the more difficult periods of morning and evening peak.(3)The importance of each module to the spatiotemporal dependency modeling of bike sharing is analyzed through the separation experiment of each module of the model.The results show that: in terms of spatial dependence modeling,for SBBS,the daily trend time convolution network is the most important;for SFBS the near trend time convolution network is more important.In terms of time-dependent modeling,the convolution module of interaction graph can improve the accuracy of the model for SBBS;for SFBS,the convolution module of time series similarity graph can improve the accuracy of the model more obviously.