Abstract:As a new mode of urban public transportation, the emergence of shared bicycles has brought great convenience to people's daily travel; however, due to the dynamic and uneven nature of their usage, users often face the phenomenon of no bikes available for borrowing and no stakes available for returning, so an accurate demand prediction model is needed to solve the scheduling problem of shared bicycles; to this end, a new attention network model based on spatial-temporal fusion graph is proposed, which connects the same station nodes of different time slices to form a large spatial-temporal fusion graph; thanks to the self-attention mechanism under the spatial-temporal fusion graph, the model can capture local spatial-temporal correlation and global spatial-temporal correlation; experiments on the Divvy dataset show that the prediction accuracy of the model is better than previous spatial-temporal data prediction models, and the prediction results match the real result curves, which can provide an effective reference for bike-sharing demand prediction systems in practice.