Enhancing multistep-ahead bike-sharing demand prediction with a two-stage online learning-based time-series model: insight from Seoul

Abstract

Bike-sharing is a powerful solution to urban challenges (e.g., expanding bike communities, lowering transportation costs, alleviating traffic congestion, reducing emissions, and enhancing health). Accurately predicting bike-sharing demand not only ensures the system meets community needs but also optimizes resource allocation, reduces operational costs, and enhances the user experience, thereby increasing the system's sustainability and city-wide benefits. However, prediction is complicated in low-computing environments with insufficient data due to privacy regulations or policy constraints. This study proposes an online learning-based two-stage forecasting model based on a low-computing environment with insufficient data for robust, fast multistep-ahead prediction for bike-sharing demand in Seoul. The model was applied with exploratory data analysis (EDA) to the Seoul Bike-sharing Demand dataset, split into insufficient training and sufficient testing sets. First, we generated prediction values for the random forest, extreme gradient boosting, and Cubist methods in training and testing. Second, we used the Ranger package trained with external factors and prediction values using time-series cross-validation for multistep-ahead prediction 1 h to 1 day later. We compared the model performance with that of 23 machine and deep learning models to verify its superiority. Using interpretability methods and EDA, we reported the relationships between external factors and bike-sharing demand.