A Short-term Demand of Bike-sharing Forecasting Model Based on Spatio-temporal Graph Data

The research aims to use deep learning to develop a site-level bike-sharing demand prediction model to address the uneven distribution of free-flowing vehicles due to the growth of bike-sharing into the market. In recent years, cycling has become an important form of supportive public transportation, especially for 'last mile' commuting. However, with the increase of bike-sharing activities in the market, some free-flowing vehicles are facing different spatial and temporal distribution problems. To overcome these challenges, we use a Graph Convolutional Neural Networks (GCN) to capture the spatial relationships between bike-sharing sites, a Gate Recurrent Unit (GRU) to capture the temporal proximity and periodicity of each site's historical data, and an Attention mechanism to dynamically capture the temporal dependencies and improve the model's performance. It is shown that the proposed approach has better performance compared to other models, as demonstrated by MAE and RMSE measurements, which have signals of 1.09 and 2.21 on this dataset, respectively. the error is reduced by at least 21.4% compared to other comparative models, showing strong predictive performance. Thus, this paper implements a deep learning model that can accurately predict the demand of bike-sharing stations, which provides a decision basis for solving the scheduling of unbalanced spatial and temporal distribution of bike-sharing.