Forecasting the demand for bicycle-metro integration incorporating spatiotemporal characteristics

With the intelligent development of urban transportation systems, accurate prediction of bike-sharing flow has become crucial for optimizing urban traffic management. This study addresses the bike-sharing flow prediction at metro stations in Longhua District by proposing a hybrid forecasting framework that integrates deep neural networks with XGBoost. A multi-source feature set incorporating temporal characteristics, weather conditions, POI density, and spatial features was constructed. The improved deep neural network and XGBoost models were trained separately and then intelligently integrated through performance-based weighting on the validation set. The neural network architecture, enhanced with simplified residual blocks and a feature attention mechanism combined with the Adam optimizer, effectively mitigated gradient vanishing and improved feature selection capability. The XGBoost model contributed through its exceptional feature importance analysis and robustness to outliers. Experimental results demonstrate that the proposed hybrid framework outperforms LSTM models in short-term forecasting, achieving a final R2 score of 0.94, confirming its effectiveness and practical utility in predicting bike-sharing demand for metro station connectivity.