Improving Aircraft Fuel Consumption Prediction through Ensemble Learning

Abstract: Performance models provided by aircraft manufacturers are used by aircraft operators to perform flight path simulations aiming to reduce aircraft fuel consumption. However, performance models are generic and does not account for the performance deviations of each aircraft individual. The performance deviations, particularly in terms of fuel consumption, will affect the dynamic programming of flight path simulations. This may result in a less optimal flight path and ultimately lead to higher fuel consumption than expected. In hope of reducing this risk, a collection of local performance factors were derived. These factors describe the percentual deviation between the real fuel flow and the levels predicted by the performance model, and are allocated with respect to a range of flight parameters in a data library known as the performance library. A test environment is then constructed to simulate a continuous flow of flight data, where a new performance library is derived from the flight data of every month. The local performance factors of the previous month are then updated with the current; a learning process based on the weighted average ensemble approach. Further, the local performance factors are used in conjunction with the performance model to estimate the aircraft fuel consumption during cruise. The observed average prediction error is noticeably smaller than that of an equivalent global, scalar performance factor used by airlines today. The result also reveals that the prediction accuracy and versatility of the performance library is mainly determined by its resolution - higher resolution generally offers better accuracy at a cost of requiring more flight data, whereas lower resolutions are more versatile but of lower accuracy. Finally, the performance libraries of two identical aircraft are used to trace the performance deviation between them. The weighted average of all local performance factors in the performance library of respective aircraft reveal that the average fuel consumption is roughly -1.9 % and -2.5 % lower than the estimates by the performance model, ultimately proving that it is feasible to detect overall fuel efficiency deviation between two identical aircraft.