Abstract
Bikesharing is a sustainable and environmentally friendly transportation mode that offers bikes â??on-demandâ?? to help improve daily urban mobility. However, their operation suffers from the effects of the fluctuating demand in space and time that leads to severe system inefficienciesâ??having either empty or full stations for long periods of time. To resolve the inefficiencies, bikesharing operators are forced to reposition bikes dynamically to avoid the system from collapsing. The knowledge of future demand patterns can aid in repositioning tasks, reducing relocation costs and increasing system performance. In this paper the authors use data from the Hubway Bikesharing systemâ??to which they add weather characteristicsâ??and implement Gradient Boosting Machines (GBM) to make station level forecasts at 20, 40 and 60 minutes. The authors demonstrate the advantages of GBM compared to Neural Networks (NN) and Linear Regression (LR), namely: reduced data cleaning and preparation times, insensitivity towards irrelevant explanatory variables and better prediction accuracies. A total of 18 models for the 61 stations are run and errors and optimal calibration parameters are obtained. For calibration purposes a differential evolution algorithm is implemented. The system root mean squared error (RMSE) normalized by the station capacity obtained without calibrating the GBM model is lower than all other models for all time windows. When compared to the equivalent NN, it is 1.33, 8.7 and 13.27 % better for the 20, 40 and 60 minutes predictions, respectively.