Interpretable Taxi-Out Time Prediction of Departure Flights Using Stacking Ensemble Learning and SHAP Analysis

To address the limitations of weak interpretability and poor generalization in existing taxi-out time prediction models, this study proposes a novel prediction model for departing flights based on Stacking ensemble learning and Shapley additive explanations. Firstly, decomposing taxi-out time into unimpeded taxi-out time and dynamic taxi-out time, followed by separate correlation analysis with influencing factors. Then, constructing a Stacking-based prediction model with comparative evaluation between holistic and phased prediction approaches. Finally, implementing SHAP analysis to quantify feature importance, and validate the rationality of the model using actual operating data from Shenzhen Bao'an international airport of China. The results indicate that: (1) Unimpeded taxi-out time is mainly influenced by the configuration of the airport, while the dynamic taxi-out time is mainly influenced by surface traffic flow; (2) Phased prediction shows enhanced interpretability despite marginally inferior performance (MAPE:12.0%, MAE:113.6s, RMSE:156.7s) compared to holistic prediction; (3) The Stacking model achieves superior accuracy (± 60s/±180s/±300s prediction rates: 41.0%/86.3%/96.5%) and generalization capability over existing methods; (4) The dual feature selection mechanism based on Shapley analysis and correlation analysis can ensure high prediction accuracy of the model while effectively reducing feature dimensions. (5) SHAP analysis was employed to quantify feature impacts on taxi-out time and decode feature interactions, thereby demystifying the model's black-box nature and offering actionable insights for air traffic controllers' decision-making.