Short-term traffic flow prediction in bike-sharing networks

Abstract

For station-based bike-sharing systems, the balance between user demand and bike allocation is critical for the operation. As a basic operational index, the short-term prediction of bike numbers (flow) plays an important role in demand forecasting and rebalancing resources of bike-sharing networks. Many different methods have been proposed for bike forecast in recent years, and the deep learning (DL)-based models have dominated this area because of their competitive performance. However, there still exist challenges in such approaches including: (i) how to appropriately select the training input for the DL model, and (ii) how to effectively utilize both the temporal and spatial features in the data for prediction. In particular, the arbitrary input may limit the model optimization, and the separate consideration of temporal and spatial features could change the original data representation. This paper uses a simple autocorrelation function to select the best input candidates and develops a three-dimensional (3 D) residual neural network to learn spatiotemporal features simultaneously. The proposed DL model is trained and validated using two separate bike-sharing datasets from New York and Suzhou cities. The learned features by two-dimensional (2 D) and 3 D CNN kernels under different input methods are compared. Results show that 3 D CNN outperforms other models and that the proposed input selection method yields better learning results for both datasets. The proposed methods help with a comprehensive DL model workflow and better forecasting accuracy.

Keywords:

• 3D Residual Neural Networks
• short-term traffic forecasting
• spatiotemporal data
• spatiotemporal features
• spatiotemporal model
• station-based bike-sharing system

Acknowledgment

Bo Wang was supported by CSIRO's Data61 and Monash University. Inhi Kim was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2021R1C1C1006405). The authors are thankful to all anonymous reviewers for their insightful comments and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

Bo Wang was supported by CSIRO's Data61 and Monash University. Inhi Kim was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2021R1C1C1006405). The authors are thankful to all anonymous reviewers for their insightful comments and suggestions.