Day-ahead aircraft routing with data-driven primary delay predictions

Abstract

Flight delays are major sources of disruptions in airline operations. To mitigate them, day-ahead aircraft routing aims to create flight sequences that can absorb delays and minimize their propagation. However, flight delays are unknown ahead of operations; moreover, predicting delays is complicated by the fact that historical data encompass both primary delays (arising from exogenous sources) and propagated delays (arising from cascading effects in an airline’s network). This paper thus develops predictive and prescriptive analytics models to forecast primary delays and to optimize day-ahead aircraft routing toward delay mitigation. We develop a quantile regression model to reconstruct primary delays from historical data, and an ensemble machine learning model to predict them based on flight-level features, environmental features, and traffic features—estimated via a queuing model of airport operations. Then, we formulate deterministic and stochastic optimization models to support day-ahead aircraft routing. Using real-world data from Vueling Airlines, we evaluate the models out of sample against real-world counterfactuals. Results show that our predictive model achieves a mean absolute error of 7–8 minutes and that our prescriptive models can reduce delay costs by 3–5%. This paper shows the benefits of predictive and prescriptive analytics to enhance the robustness of airline operations by (i) creating shorter aircraft rotations, and (ii) strategically allocating schedule slack to avoid the propagation of long delays in later phases of the day. This research led to the deployment of the models in collaboration with the Vueling data science unit.