Conference Agenda

Purpose:

This study proposes an explainable machine learning workflow for taxi-in time prediction at Frankfurt Airport to support transparent, data-driven ground operations. Accurate taxi-in time prediction is important for the scheduling of ground operations. Given the persistent labor shortages in this sector avoiding waiting times based on more accurate scheduling optimizes ground crew utilization but also avoids delays at the beginning of the aircraft turnaround.

Prior studies have investigated aircraft taxi times at various airports using ADS-B (Automatic Dependent Surveillance – Broadcast), A-CDM (Airport Collaborative Decision Making) and AODB (Airport Operational Databases) data. Most of the studies thereby focus on the taxi-out-times or the full turnaround from the perspective of an airline or airport with the objective of reducing delays or maximizing system efficiency. However few studies have predicted taxi-in-times specifically for optimized ground operations.

Methodology:

The study uses ADS-B-based aircraft movement data from Frankfurt Airport, covering approximately 67,000 arrival events between July 2025 and April 2026, combined with operational and contextual information. Compared with prior studies using broader feature sets, the workflow refines the input space to a compact set of 17 meaningful variables by retaining core operational predictors and adding a few relevant new features. This design improves interpretability, reduces redundancy, and relies on information that is practically available around landing, including landing runway, parking position, aircraft type, local traffic load, time-related variables, weather conditions, and airline-related information. Several supervised regression models are trained and compared, including Linear Regression, MLP, Random Forest, XGBoost, LightGBM, and CatBoost. CatBoost achieved the best overall performance and was selected for detailed evaluation. Model performance is assessed using mean absolute error and operationally interpretable accuracy thresholds. A bootstrap-based uncertainty estimation approach is used to estimate prediction confidence, while subgroup analyses identify operational conditions where taxi-in time is harder to predict.

Findings:

The CatBoost model achieves a mean absolute error of about 105 seconds (≈1.75 min). Around 86% of taxi-in events are predicted within 3 minutes and 96% within 5 minutes, showing useful accuracy under real airport operating conditions. However, subgroup analysis reveals substantial variation, suggesting that specific runway-stand combinations and traffic conditions are systematically more difficult to predict. SHAP-based explanation identifies landing runway and parking position as the most influential predictors. In addition to point predictions, the model provides a confidence score for each flight, indicating how reliable the prediction is expected to be.

Originality:

This study contributes to airport surface operations research in two main ways. First, it extends taxi-in time prediction beyond point estimates by introducing a confidence-aware workflow. Each prediction is accompanied by a confidence score that indicates its expected reliability. Second, the study examines prediction difficulty across operational subgroups, making it possible to identify the conditions under which larger errors are more likely. This provides a more operational view of model reliability than aggregate accuracy measures alone. Together with the explainable ADS-B-based modelling approach, these contributions can support practical planning decisions, such as dispatching ground crews to assigned stands under specific runway, parking, and traffic conditions.