Using Machine Learning to Assess the Relevance of Weather Conditions for Short-Term Demand Predictions in Bike-Sharing Systems

Abstract

Bike-sharing systems (BSS) are being studied for their potential to enhance urban accessibility and sustainable mobility. Despite its popularity, effectively managing BSS encounters challenges due to demand-supply imbalances. The significance of BSS lies in accurately predicting bike demand at various stations, a task we approach using regression methods such as Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Regularized Linear Regression (Ridge), and Least Absolute Shrinkage and Selection Operator (LASSO), for short-term predictions. The study utilizes data from Los Angeles on bicycle-sharing and weather conditions, alongside the p-median method for station clustering. The results demonstrate that combining both datasets yields accurate predictions at the city level, with an error rate of 0.1%, and at the station level at 18%. Notably, RF emerges as the most accurate method among the regression models examined.