Short-term Prediction of Bike-sharing Usage Considering Public Transport: A LSTM Approach

Bike-sharing has experienced rapid development worldwide in the last decade. Accurate usage prediction is necessary to support timely reposition and ensure availability. Existing prediction methods of bike-sharing usage are mostly based on its own history. They can capture temporal characteristics under normal state but cannot respond to sudden events. Bike-sharing is a part of public transport. The interrelation of bike-sharing and other public transport systems can be utilized to capture the impact of sudden events. This prediction approach considers both historical usage and real-time passengers of public transport. Then use neural networks to establish the connection among them. Long short-term memory (LSTM) is adopted because it shows outstanding performance to learn long-term temporal dependencies. Experiments show that the prediction through this approach is more accurate than baselines. The mean absolute error (MAE) reduces by 25.32% to 39.82%. With the input of real-time boarding passengers, the prediction better responds to sudden changes, and the MAE value further reduces by 21.81%. This approach also outperforms others in different prediction horizons. Moreover, the methodology is applicable to both traditional and station-free bike-sharing systems. Data for the prediction are available in many cities, and hence it is ready for practice.