Abstract
This study proposes an hourly demand prediction method based on a multi-feature edge-weighted gated graph convolutional network to address the imbalance in station borrowing and returning demands, as well as low station utilization in bike-sharing systems. By employing graph convolutional neural networks to capture spatial relationships between stations and utilizing gating mechanisms to integrate current and historical information, it captures the long-term dependency of time series data. Creatively, it combines three single edge-weight features—station distance, time, and correlation value—into a multi-feature edge-weighted input model graph structure, enhancing the accuracy in reflecting traveler behavior characteristics. Additionally, the study considers not only temporal and spatial factors but also incorporates traveler features as node inputs to the model. Using bike-sharing trip data from Jersey City in 2020, the study employs isolation forest algorithm for outlier detection, followed by feature dependency analysis to reveal the impact of time, space, and traveler features on demand. Moreover, it accounts for the seasonal influence on bike-sharing trips by dividing the dataset into different seasons and conducting unified research on similar seasons. Results demonstrate that the multi-feature edge-weighted gated graph convolutional neural network achieves an MAE of 0.52 and MSE of 0.906 for the spring and autumn seasons, and an MAE of 0.296 and MSE of 0.594 for the summer and winter seasons, outperforming baseline models and single-feature edge-weighted predictive performance.