Estimating fuel consumption over flight segments using gradient boosting methods

Abstract

Accurately estimating aircraft fuel consumption from surveillance data is essential for assessing environmental impact and improving the efficiency of air traffic operations. However, many variables that determine fuel burn, such as aircraft mass, engine performance, and atmospheric conditions, are not directly observable from ADS-B. This study investigates the extent to which fuel consumption can be inferred statistically from ADS-B trajectories that are enriched with weather information, inferred mass, and physics-informed features. We construct a multi-stage preprocessing pipeline that includes ADS-B cleaning, interpolation, resampling, wind-data augmentation using ERA5, estimation of take-off weight, and derivation of fuel flow and mass profiles based on the OpenAP performance model. Several feature groups are engineered, including statistical embeddings of ADS-B and ACARS variables, detailed vertical-rate phase decomposition, mass-related descriptors, and meteorological corrections. Using these features, we train a LightGBM regression model to predict fuel flow, which is then integrated to obtain segment-level fuel consumption. Experimental results on the challenge ranking dataset show that mass estimation, especially when using a heuristic interpolation between take-off and landing weight, substantially improves predictive accuracy. Models trained on fuel flow consistently perform better than those trained directly on fuel consumption, because the fuel flow target exhibits lower skewness and better reflects the underlying physics. The final configuration achieves an RMSE below 200 on fuel flow prediction, demonstrating the value of combining physics-informed features, statistical embeddings, and robust mass estimation. These findings highlight the importance of high-quality preprocessing and feature engineering for large-scale fuel-burn inference from sparse operational data.