Demand Prediction with Gravity Graph Attention Networks for Shared Bike Systems

Shared bicycles, as an important component of green transportation, have experienced rapid development worldwide, making accurate demand forecasting crucial for urban transportation planning and operational management. However, existing deep learning approaches primarily focus on time series data and often fail to capture the complex spatial characteristics among stations. To address this issue, we propose a demand prediction method that integrates Graph Attention Network (GAT) with Gravity model. By constructing a geographic adjacency graph and incorporating multi-source information— such as station flow, built environment features, and geographic coordinates—into the node representations, the proposed method effectively combines the prior knowledge of the gravity model with the deep feature extraction capabilities of GAT, thus providing a comprehensive characterization of spatial interactions among stations. Experiments on a public dataset from Xiamen demonstrate that the fusion model outperforms standalone GAT, Gravity model, and GCN (Graph Convolutional Network) in terms of MAE, RMSE and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, validating its effectiveness and superiority in predicting shared bicycle demand.