Long-Term Demand Prediction for Public Bicycle Sharing System: A Spatio-Temporal Attentional Graph Convolution Networks Approach

Accurately predicting the long-term demand for public bicycle systems (PBS) is crucial for policy implementations such as operator rebalancing. With the continuous advancement of deep learning techniques, including self-attention mechanisms and graph convolution networks (GCNs), we can better understand the nonlinear relationship of spatio-temporal data. Despite increasing interest in long-series transportation forecasting, there are very few long-term spatio-temporal predictions focused on PBS data. In this paper, we introduce a novel spatio-temporal attentional graph convolutional network (ST-AGCN) designed to predict demand accurately over a 24-hour horizon, addressing the limitations of existing models and the characteristics of PBS data. The core innovation of our model is the integration of a probabilistic self-attention mechanism (Probsparse self-attention) for predicting long-term temporal dependencies and Chebyshev GCN for extracting complex spatial dependencies within the encoder-decoder framework. This integration significantly reduces computational time. Additionally, we devise a PBS data-based spatial weighting calculation method to generate the spatio-temporal graph, which facilitates capturing more inter-station relationships to improve computational efficiency and accuracy. Experimental results compared with seven baseline models on real-world datasets demonstrate that ST-AGCN outperforms state-of-the-art baselines in both accuracy and computational efficiency.