A periodical decomposition-based two-stage NARX model for demand prediction of bike-sharing travel in hotspot areas

Abstract

Future free-floating bike-sharing travel demand forecasting systems can mitigate dispatch failures. Typically, time series forecasts for traffic demand and flow are computed using a global study area, which does not account for spatial heterogeneity. To address this, a periodical decomposition-based two-stage NARX (Nonlinear Auto Regressive with Exogenous Inputs) model is developed to accurately predict free-floating bike-sharing travel demand (BSTD) for individual hotspot areas. Kernel density analysis-based hotspot detection is employed to divide the study area into basic predicting units, thereby enhancing efficiency and guidance quality. Weather factors with poor predictability or low correlation with BSTD are further eliminated using rescaled range and gray correlation methods. Based on periodic decomposition results, an improved two-stage NARX model is constructed for BSTD prediction in multiple important steps. A random selection of 50 hotspot areas in Beijing was performed for methodology verification, with hotspots numbered and selected using a random number generator. Results indicate that the periodical decomposition-based NARX model significantly improves BSTD prediction accuracy in hotspot areas compared to typical time series forecasting methods. The model demonstrates higher R-values (correlation between targets and outputs) and lower MSEs (Mean Squared Errors). For instance, the average MSE of the periodical decomposition-based two-stage NARX model is 20.225, compared to ARIMA (26.151), NARX (28.748), ARIMA (32.854), and NAR (41.666), highlighting superior robustness and effectiveness across different hotspot types and locations. These findings enhance understanding of the spatial-temporal variation of BSTD and provide a foundation for optimizing time series forecasting within specific areas.