Aircraft Disruption Management: Increasing Performance with Machine Learning Predictions

Airlines experience schedule disruptions on a daily basis. Poor weather conditions, unscheduled aircraft maintenance and congested air spaces are just a few of the causes that prevent airlines from operating their flight schedules as planned. In the third quarter of 2017 over 20% of all scheduled flights in Europe suffered from delays. Operation Research based decision support systems (DSS) help airlines with their disruption management processes and provide suggestions for recovery options.For large hub-and-spoke carriers, with an extensive network and a large number of aircraft and flights, the computation time required to find the optimal recovery solution after a disruption increases rapidly. As airlines require fast recovery solutions when disruptions occur, there is an ongoing trade-off between computation time and system sophistication. In the majority of disruption cases, no or a limited number of undisrupted aircraft are required to find the optimal recovery solution. By selecting a limited number of aircraft, flights and airports used to find a recovery solution, the computation time is reduced exponentially. The challenge is determining which aircraft and flights should be selected.This research aims to develop a decision support system for the schedule and aircraft recovery process that is able to present a feasible solution to a disruption in less than 120 seconds. An aircraft recovery model will be developed based on the integer linear programming model that was created by Vink et al. (2019) and Vos et al. (2015). Crew and passenger recovery are not considered. To recover disruptions, the optimization model can delay and cancel flights as well as perform tails swaps, where the flights from two aircraft are switched. The novelty of the work is that machine learning is used to predict which undisrupted aircraft will help recover a disruption. Based on those predictions a sub-network selection algorithm will select the subset of aircraft to be included in the optimization instead of the entire aircraft fleet.The performance of the system is tested on a case study for the domestic hub-and-spoke network of Delta Airlines. The dataset for the study consists of 2200 daily flights, 147 airports and 827 aircraft in 8 aircraft families. The results of the system are compared with the optimal solution, where no aircraft selections were made. The case study shows that the system is able to make an aircraft selection where 50% of the fleet is discarded, while still finding the optimal solution in 98.9% of the 556 disruptions tested. Furthermore, the system reduces the computation time by 45%, resulting in an average time of 48 seconds. For the disruptions, the computation time varied between 9 and 180 seconds.Aerospace Engineerin