Abstract: Short-term demand of shared bikes forecasting plays an important role in optimizing the layout of free-floating bike sharing systems and bikes rebalancing. To improve the accuracy of demand forecasting methods for the emerging shared bikes business, this study establishes a combined forecasting model based on Stacking strategy and examines the impact of temporal variables and weather factors on shared bikes demand. In particular, the Granger causality test is used to identify the key weather indicators that cause demand fluctuations. We extract the set of similar samples for each period of the day for prediction based on the grey correlation index of weather variables. The Stacking strategy is then introduced to integrate random forest, support vector regression, and other machine learning algorithms for establishing a combined forecasting model to predict the short-term demand in different regions. Finally, using free-floating bike sharing data in Beijing, the proposed combined model is tested. The prediction results of combined model demonstrate that the mean absolute percentage error decreased by 9.1% compared with the single prediction model, thus improving the accuracy of the short-term demand forecast of shared bikes and providing useful information for bike relocation.

Key words: shared bikes, travel demand, causal analysis, grey correlation, similar day, machine learning, Stacking strategy