H3‑Cell‑Based Arrival‑Time Prediction for Lufthansa Flights

LMU‑Munich Data Science Practical in cooperation with Lufthansa Group

---

Overview

This project builds a full pipeline that turns raw ADS‑B surveillance data into minute‑level ETA predictions for Lufthansa flights inbound to Frankfurt (EDDF). Key points:

• Uses H3 hexagonal spatial indexing to capture local traffic density along each trajectory
• Supports whole‑route and last‑100 km prediction modes
• Codebase: data download, feature engineering, model training, evaluation, and inference

Feature set & modelling

• Features:
  • Position (distance‑to‑runway, sine/cos‑encoded lat / lon and bearing)
  • Kinematics (altitude, vertical speed, ground speed)
  • H3 traffic density over the past 10 / 30 / 60 minutes
  • Calendar & cyclic time (weekday, holiday, sine/cos‑encoded time‑of‑day and day‑of‑year)
• Targets: seconds‑to‑touchdown (full) or seconds‑to‑touchdown within 100 km
• Models: polynomial regression, XGBoost, MLP, LSTM
• Interpretability: SHAP plots available in src/evaluations/

License

MIT - see LICENSE.

Acknowledgements

• Dr. Viktor Bengs (LMU Chair of Artificial Intelligence and Machine Learning) - academic supervision and guidance
• Dr. Sebastian Weber - industry mentor (Lufthansa Group)
• OpenSky Network for ADS‑B data