Variable Taxi-Out Time Prediction Based on Machine Learning with Interpretable Attributes

Abstract

This paper presents a machine learning-based approach for predicting the taxi-out time, with the departure process decomposed into two components: the time taken to travel from the gate to the departure queue, and the time spent in the departure queue. Gradient-Boosted Decision Tree (GBDT) models are trained to predict the two components using different feature sets, and a comparison of both model shows that they can provide better prediction accuracy compared with conventional methods, with a Root Mean Squared Error (RMSE) of 1.79 minutes and 0.92 minutes when predicting the taxiing and queuing times respectively, and 78% and 96% of predictions falling within a ±2 minute error margin. Predictions from the GBDT model are analysed and interpreted using SHAP (SHapley Additive exPlanations) values, a well-recognised technique for providing interpretability to many different black-box models, and allowing feature importance to be evaluated at global (model) and local (individual prediction) levels. In particular, the most important feature groups for the taxiing and queuing models are respectively the route features and runway queuing features. The model explainability provides a pathway towards the certification of machine learning techniques in Air Traffic Controller (ATCO) decision support tools.