Optimized Demand Forecasting for Bike-Sharing Stations Through Multi-Method Fusion and Gated Graph Convolutional Neural Networks

This study presents an innovative approach to hourly demand forecasting for bike-sharing systems using a multi-attribute, edge-weighted, Gated Graph Convolutional Network (GGCN). It addresses the challenge of imbalanced bike borrowing and returning demands across stations, aiming to enhance station utilization rates. The proposed model employs GGCNs to capture spatial relationships between stations and incorporates both current and historical data via a gating mechanism to account for temporal dependencies. By integrating three key edge-weight attributes—stations distance, travel duration, and correlation—into a multi-attribute graph framework, the model significantly improves predictive accuracy for user travel patterns. Additionally, user characteristics are included as node features, enabling a more comprehensive analysis. The study utilizes the 2020 dataset from Jersey City’s bike-sharing system, starting with the application of the Isolation Forest algorithm to detect and filter anomalous data points. A detailed analysis of the temporal, spatial, and user-related factors is conducted to examine how these features interact and shape demand patterns in the bike-sharing network. The dataset is segmented by season to account for seasonal variations. The results demonstrate that the multi-attribute, edge-weighted GGCN outperforms baseline and single-attribute models, achieving a Mean Absolute Error (MAE) of 0.521 and Mean Squared Error (MSE) of 0.918 for spring and autumn, and an MAE of 0.307 and MSE of 0.608 for summer and winter.