Demand prediction for a public bike sharing program based on spatio-temporal graph convolutional networks

The operation of public bike sharing (PBS) programs has attracted attention again due to numerous problems encountered by free-floating bike sharing programs. These problems include malicious damage, theft, chaotic parking, large-scale deficit and bankruptcy. The short-time demand prediction is a key issue for the successful operation of PBS programs. In this study, we use a novel spatio-temporal graph convolutional network (STGCN) to predict the picking up/returning demand by exploring potential information from multi-view data. We apply graph convolutional neural networks (CNNs) to represent the spatial dependency based on the geographic information system data denoting the location of docks. Moreover, we use gated CNNs to denote the temporal dependency according to the time-series data representing the demand for picking up/returning public bikes. The STGCN and three recurrent neural network (RNN)-based competitors are trained and validated using the multi-view data from the Wenling PBS program for one month. The RNN-based competitors consist of the SimpleRNN, long short term memory and gated recurrent unit. Results show that the STGCN achieves higher prediction accuracy compared with its competitors. Although the STGCN consumes a longer training time compared with the SimpleRNN, it requires a minimal number of epochs to achieve convergence precision. The complete CNN structure in the STGCN can effectively address the spatial and temporal dependencies for PBS demand prediction.

Automated summary

The operation of public bike sharing programs faces various issues, with accurate short-term demand prediction being crucial for successful operation. STGCN outperforms other competitors in predicting demand accuracy. Despite longer training time compared to SimpleRNN, STGCN requires minimal epochs for convergence precision.