BikeCAP: Deep Spatial-temporal Capsule Network for Multi-step Bike Demand Prediction

Given the recent global development of bike-sharing systems, numerous methods have been proposed to predict their user demand. These methods work fine for single-step prediction (i.e., 10 mins) but are limited to predicting in a multi-step prediction (i.e., more than 60 mins), which is essential for applications such as bike re-balancing that requires long operation time. To address this limitation, we leverage the fact that the demand for upstream transportation, e.g., subways, can assist the future demand prediction of downstream transportation, e.g., bikes. Specifically, we design a deep spatial-temporal capsule network called BikeCAP with three components: (1) a historical capsule that learns the demand characteristics for both the upstream (i.e., subways) and downstream (i.e., bikes) transportation systems, where a pyramid convolutional layer explores the simultaneous spatial-temporal correlations; (2) a future capsule that actively captures the dynamic spatial-temporal propagation correlations from the upstream to the downstream system, in which a spatial-temporal routing technique benefits to reduce the accumulated prediction errors; (3) a 3D-deconvolution decoder that constructs future bike demand considering the similar downstream demand patterns in neighboring grids and adjacent time slots. Experimentally, we conduct comprehensive experiments on the data of 30, 000 bikes and 7 subway lines collected in Shenzhen City, China, The results show that BikeCAP outperforms several state-of-the-art methods, significantly increasing the performance by 38.6% in terms of accuracy in multi-step prediction. We also conduct ablation studies to show the significance of BikeCAP’s different designed components.