Subnetwork prediction approach for aircraft schedule recovery

Abstract

The combinatorial optimization techniques for the aircraft schedule recovery problem are too expensive to compute. However, we require a reasonable solution within 1–2 min for real-time application. The computational time of traditional optimization algorithms will drastically reduce if we can accurately predict the subnetwork affected by the disruption. In this study, we advocate the use of machine learning (ML) as a promising tool to predict the subnetwork. The optimal recovery schedules for a large number of disruptions are used for offline training, and subsequently, the trained ML model is employed to reduce the search space for new disruption cases substantially. To achieve this, we employ a unique feature space specifically designed to capture the essential characteristics of the problem. A notable contribution is constructing a novel feature space based on revised aircraft schedules designed to capture the essential characteristics of the problem. The recursive feature elimination technique is employed for optimal feature selection. Five machine learning classifiers: Logistic Regression, Random Forest, Extreme gradient boosting, Light gradient boosting, and Categorical Boosting are compared. The performance is evaluated on real data obtained from an airline company. Our study demonstrates that, with the subnetwork of aircraft predicted by the classifier, the computational time of the column generation approach is remarkably reduced without deteriorating the solution quality in all the disruption instances tested, with more than 98% of the instances being solved within 60 s. The results highlight the remarkable advantage of integrating ML in the pre-optimization phase.