A fusion deep learning network for shared e-bike demand prediction with spatiotemporal dependencies

The growing popularity of shared transportation services relies heavily on their accessibility to users. Achieving a balanced supply and demand for dockless shared electric bicycles (e-bikes) is crucial for the widespread adoption of such services. In this study, we introduce a fusion model that incorporates a hierarchical structure integrating various features. Our model sequentially processes the spatial dependencies of demand using convolutional neural networks (CNN), followed by the temporal dependencies using long shortterm memory (LSTM). Additionally, we employ CNN to extract the spatial dependencies of points of interest (POI) and introduce additional layers to handle external features capturing their variability. Notably, compared to other transportation modes, shared e-bike trips typically involve shorter distances and require finer spatial grids, which pose a challenge that needs to be addressed effectively. Our proposed model demonstrates high accuracy and generalization capabilities, even when dealing with fine-grained gridding and sparse data generated at the finer granularity. The results indicate that POI and timestamp are crucial for demand forecasting, while weather variables are less significant.