Bike-Sharing Demand Prediction Based on Dynamic Time Warping and Spatio-Temporal Graph Attention Network

Bike-sharing demand prediction involves complex, dynamic spatio-temporal dependencies and various influencing factors, thus becomes one of technical challenges in intelligent transportation systems. Existing methods often rely on predefined adjacency matrices based on distance or road connectivity, and typically ignore multi-scale temporal features and external factors such as weather, holidays, social events, and so on. To address these limitations, we propose a model based on dynamic time warping (DTW) and spatio-temporal graph attention network (GAT) to improve the accuracy of bike-sharing demand prediction. In the proposed model, we use a data-driven approach to construct an adjacency matrix that effectively reflects the real dependencies between bike-sharing stations, and temporal attention mechanism is integrated with graph attention network to capture dynamic spatio-temporal correlations hidden in the data. Moreover, multi-scale temporal gated convolutions are applied to fuse short-term and long-term temporal features. The experimental results demonstrate that our proposed model significantly outperforms recent baseline methods in terms of MAE and RMSE evaluation metrics. Meanwhile, we find that the external factors of weather, public facilities and traffic accidents have different influence on results, and the weather has the greatest impact on bike-sharing demand.