An Interpretable Machine Learning Approach for Predicting Ground-Handling End Time Adherence

Abstract

To accommodate the rising demand for airport resources and airspace capacity, a more efficient and robust orchestration of the components that facilitate air travel is required than is the case today. Over recent years, camera systems for automatic identification of ground-handling events were developed, providing the opportunity to automatically detect ground-handling delays as they manifest. Detecting ground-handling delays early allows the Air Navigation Service Providers (ANSPs), ground handlers and aircraft operators to reallocate their resources, facilitating more stable planning and robustness of the air traffic network. This study proposes a novel interpretable machine learning approach to predict initial ground-handling end time adherence. This approach is based on a discrete Bayesian network composed of a central network structure and modular sub-networks, representing distinct turnaround processes often found on the critical path. The Bayesian network structure and discretization were found to be adequate to model most ground-handling operations for narrow-body aircraft. The model proved to be more accurate than the ground handler in predicting initial Target Off-Block Time (TOBT) adherence at all prediction moments. At the initial TOBT-0 [minutes], a maximum in model accuracy was obtained amounting to 65.76%. In addition, the model demonstrated the capability to detect ground-handling delays earlier. At the initial TOBT-20 [minutes], an average recall of 50.4% was obtained for delays within 5 and 15 minutes. In contrast, the ground handler identified only 4.3% of small ground-handling delays at this prediction moment. The discrete Bayesian network approach is also inherently interpretable, providing transparency into the decision-making process of the model.