Predicting demand for a bike-sharing system with station activity based on random forest

Public bike-sharing (PBS) systems have expanded to major cities around the world in efforts to mitigate air pollution, traffic congestion and traffic accidents. Users can pickup and drop-off bicycles at any station, and thus inventory imbalances can occur. To improve system efficiency, system operators should establish appropriate repositioning strategies based on accurate predictions of demand for bicycles. This study aims to predict station-level demand for pickup and drop-off of bicycles using station activity information. In addition to time and weather information, the number of pickups and drop-offs at a station 1–3 h before the prediction was used as a predictor. A random forest machine learning technique is adopted for the demand prediction. The PBS database in Seoul, South Korea was used for the case study. To compare prediction accuracy by station usage patterns, the stations are classified into four clusters. The analysis results show that prediction accuracy including lag information provides mprovements of up to 20%, and the forecast for drop-off is more accurate than the forecast for pickup. This study practically contributes to increasing operational efficiency and reducing operating costs by improving demand predictability in a PBS system.