融合需求时空特征关联的共享单车骑行量预测OA

[Background]Owing the widespread use of shared bikes in urban transportation,the is-sue of supply-demand imbalance has become prominent.An accurate prediction of their spatiotempo-ral demand is essential to address this issue.Existing studies do not adequately capture the spatiotem-poral correlations of bike-sharing demand.[Objective]This study proposes a graph convolutional neural network(GCN)model that captures both spatial and temporal correlations to improve the pre-diction accuracy of bike-sharing demands and provide a scientific basis for operational scheduling.[Method]Utilizing the 2018 Mobike bike-sharing data from Shanghai,the model incorporates a time-series feature matrix and adjacency matrices spanning three dimensions:geospatial proximity,land use,and demand-sequence similarity.These matrices enable the model to represent the complex relationships among various locations within the study area.[Result]The GCN models,which incor-porate spatial features,significantly outperforms the long short-term memory model,which only con-siders temporal dynamics across all evaluated metrics,including R2,RMSE,MAE,and WMAPE.Among the three types of adjacency matrices,the GCN model that incorporates land use yields the most accurate predictions,with R2values of 0.86 and 0.85 for pick-up and drop-off demands,respec-tively.These results underscore the importance of land use in bike-sharing demands.[Application]This study offers valuable technical support for precise demand prediction in bike-sharing ridership,thereby assisting operators in implementing more efficient scheduling and operational strategies,as well as enhancing the service efficiency and user satisfaction of the bike-sharing system.