Station-level Demand Prediction for Bike-sharing Based on Multi-spatio-temporal Scale Features

Nowadays, the bike-sharing system has advanced to new levels, solving the issue of people riding short distances. The most critical challenge for the successful implementation of the bike-sharing system is resolving the rent/return imbalance. By improving the predictability of bicycle demand, this issue can be solved. Accurate predictions can be made by taking temporal and spatial into account. However, these features obtained from different temporal and spatial scales are distinct. And the disadvantage of sequential structure is that the extracted spatial and temporal features may affect each other due to the stacked structure. As a result, we propose a parallel architecture with multi-spatio-temporal scale (MST-RLGCN). The parallel structure consists of a RLSTM model construct around historical data at multi-temporal scale (hour-level, day-level, and week-level) and an MGCN model construct around multi-spatial (first-order, second-order, and third-order graph convolution operators). Our framework not only efficiently extracts the temporal and spatial dependencies between station, but also allows the temporal and spatial features to be effectively combined, which improve the accuracy of demand predicting. We compare MST-RLGCN with other baseline models using New York bike-sharing system data. Experimental results show that our method outperforms baseline prediction models.