Bike Usage Prediction Based On Station-based Bike-sharing System

In recent years,as a new mode of travel,shared bikes have gradually entered the lives of the public,effectively solving the ‘last-mile’ problem of residents’ travel,and have now become an essential mode of transportation for many people.However,in the actual operation process,there are inevitably some problems with the bike-sharing system that affect the user experience,one of the prominent problems is the imbalance between the supply and demand of bicycles at many shared bike stations.Accurately predicting the usage of shared bikes in bike-sharing systems can help solve this problem.Based on this motivation,this paper conducts research on how to effectively improve the accuracy of bike usage prediction in bike-sharing systems.The research work of the paper is mainly summarized as follows:1.A migration trend clustering-based shared bike distribution prediction model MT-DPM is proposed to predict the bike usage in the bike-sharing system in the future time.Firstly,a migration trend clustering algorithm MT-FCM based on fuzzy C-mean clustering is designed to cluster the shared bike stations by considering the geographic location information of all stations within the bike-sharing system and the transfer trend of shared bicycles among the stations,and then the distribution prediction idea is used to predict the number of borrowed and returned bicycles respectively.For the prediction of the number of borrowed bikes: first,MT-DPM use the gradient boosting decision tree to predict the overall number of borrowed bikes of the bike-sharing system based on the number of borrowed bikes at historical moments,and then use the multi-similarity reference model(MSI)to predict the proportion of borrowed bikes of each cluster generated after clustering,and in this way,the overall number of borrowed bikes is assigned to each specific cluster,so as to derive the number of shared bikes of each cluster for the prediction of the number of borrowed bikes;for the prediction of the number of returned bikes,firstly,MT-DPM use a multisimilarity reference model to learn and predict the flow trend matrix of shared bikes between each cluster,and then predict the borrowing time of users based on the normal distribution of the borrowing time,and finally,we consider the number of borrowed bikes of previous moments,the flow trend of shared bikes between each cluster,and the borrowing time of users to infer the The number of returned bikes in each cluster is estimated.To evaluate the performance of the MT-DPM model,it is experimentally validated on the New York bike-sharing system dataset.The comparative experimental results show that MT-DPM has good prediction performance compared with other models.2.A deep spatio-temporal residual network model(RST-Net)based on region reconstruction is proposed to predict the number of borrowed and returned bikes in the bike-sharing system in future time.First,we propose a regional reconstruction algorithm RCS,which uses the geographic location of shared bicycle stations and the ‘station→region’ transfer trend matrix of each station as the basis,and uses a hierarchical iterative Gaussian Mixture Model clustering algorithm to reconstruct all the shared bicycle stations in the city into different regions.RCS makes good use of the migration trend of shared bicycles between regions and effectively improves the accuracy of clustering.After the regional reconstruction by RCS,the matrix of the number of bikes borrowed and returned in each region is calculated based on the results of the regional reconstruction,and then the regional reconstruction algorithm is combined with a deep spatiotemporal residual network to model the key factors affecting shared bicycle usage to obtain the final prediction results.Comparative experiments on the New York bike-sharing system show that the prediction accuracy of the RST-Net model is significantly better than that of many existing deep learning-based bike-sharing traffic prediction models.