Integrating Spatial Self-Attention and Convolution Networks for Improved Car-Sharing Demand Forecasting

The core challenge facing the field of car-sharing demand forecasting lies in the innovative construction of models that effectively capture the intricate spatio-temporal variations in the data. Current methods face two particularly significant challenges: first, Current models struggle to capture the mutual influences and connections between nearby parking stations; second, when addressing data involving long time series, traditional methods often encounter the dilemma of gradient vanishing or exploding. In view of this, we proposed the SG-SCINet prediction model, which cleverly combines the advantages of the spatial self-attention mechanism and the sample convolution and interaction network (SCINet). By introducing the self-attention module, SG-SCINet effectively analyzes spatial and functional interactions, improving the prediction of car-sharing demand. This series of designs significantly improves the model’s adaptability in complex spatio-temporal environments. Experimental verification shows that the SG-SCINet model shows significant advantages over a single model in terms of prediction accuracy.