| Resumo : |
Accurately predicting the actual performance of a flight during its en-route phase, particularly deviations from the planned flight path, is crucial for optimizing fuel planning and airspace utilization. This study addresses the predictive modeling of en-route operational performance within the Brazilian airspace using machine learning techniques. We develop a multi-quantile regression model to estimate the deviation between the actual flown distance and the planned distance during the en-route flight phase. The model, learned with the CatBoost algorithm based on gradient-boosted decision trees, provides probabilistic forecasts and quantifies predictive un- certainty. Local interpretability is achieved through Shapley Additive Explanations (SHAP), providing insights into the relative influence of explanatory features. Using one year of operational data comprising aircraft surveil- lance and flight plan information, the proposed method outperforms baseline statistical approaches, reducing the multi-quantile error by 77%. By integrating machine learning techniques that combine predictive accuracy with interpretability, the proposed approach aims to deliver valuable decision support for airlines and air traffic management, particularly in areas such as fuel planning and traffic flow management. |