Fuel Burn Prediction from Open Flight Trajectory and Meteorological Data for the PRC Data Challenge 2025

Abstract

This paper describes a segment-level fuel burn prediction system developed for the PRC Data Challenge 2025, operating on open aviation data sources: ADS-B and ACARS trajectories, ERA5 pressure-level reanalysis, METAR surface observations, and FAA aircraft performance records. The dataset comprises 131,530 labeled segments derived from 11,037 European flights (April–September 2025). Rather than applying machine learning directly to raw trajectory statistics, the methodology constructs a 73-dimensional feature space that encodes physically motivated intermediate quantities: ERA5 wind-corrected true airspeed, OpenAP-derived drag and thrust coefficients, total energy model mass estimates for climb and descent phases, and phase-specific fuel-flow surrogates identified via the traffic library phase detector. Gradient boosting regressors (XGBoost, LightGBM) with Optuna-tuned hyperparameters are trained on this feature space using K-fold cross-validation. XGBoost achieves a cross-validation RMSE of 232.76 kg; LightGBM reaches 243.91 kg. The best submission to the challenge blind test set scores 227.32 kg RMSE. The primary methodological distinction from prior purely statistical approaches is the systematic substitution of latent physical variables — mass, airspeed, phase-resolved fuel flow — in place of raw kinematic observables, without requiring access to proprietary flight data recorder or BADA performance data.