Forecasting public bicycle rental demand using an optimized eXtreme Gradient Boosting model

Abstract

The rational distribution of public bicycle rental fleets is crucial for improving the efficiency of public bicycle programs. The accurate prediction of the demand for public bicycles is critical to improve bicycle utilization. To overcome the shortcomings of traditional algorithms such as low prediction accuracy and poor stability, using the 2011–2012 hourly bicycle rental data provided by the Washington City Bicycle Rental System, this study aims to develop an optimized and innovative public bicycle demand forecasting model based on grid search and eXtreme Gradient Boosting (XGBoost) algorithm. First, the feature ranking method based on machine learning models is used to analyze feature importance on the original data. In addition, a public bicycle demand forecast model is established based on important factors affecting bicycle utilization. Finally, to predict bicycle demand accurately, this study optimizes the model parameters through a grid search (GS) algorithm and builds a new prediction model based on the optimal parameters. The results show that the optimized XGBoost model based on the grid search algorithm can predict the bicycle demand more accurately than other models. The optimized model has an R-Squared of 0.947, and a root mean squared logarithmic error of 0.495. The results can be used for the effective management and reasonable dispatch of public bicycles.