Integrated Airline Fleet and Crew Recovery through Local Search

TL;DR

The paper tackles integrated airline fleet and crew disruption recovery under real-time constraints. It introduces a fast, directed local-search algorithm based on simulated annealing that relaxes hard constraints with penalties to quickly explore feasible recovery options, including multi-type aircraft swaps and crew changes. Through extensive experiments on real KLM data, the integrated approach significantly outperforms naive and sequential methods, achieving about a 40% reduction in non-performance costs and resolving disruptions within around 30 seconds. The results support the practicality of integrated recovery for reducing delays and passenger impact, while offering flexible cost models and strategies for different airline priorities.

Abstract

Airline operations are prone to delays and disruptions, since the schedules are generally tight and depend on a lot of resources. When disruptions occur, the flight schedule needs to be adjusted such that the operation can continue. Since this happens during the day of operations, this needs to be done as close to real time as possible, posing a challenge with respect to computation time. Moreover, to limit the impact of disruptions, we want a solution with minimal cost and passenger impact. Since airline operations include many interlinked decisions, an integrated approach leads to better overall solutions. We specifically look at resolving these disruptions in both the aircraft and crew schedules. Resolving these disruptions is complex, especially when it is done in an integrated way, i.e. including multiple different resources. To solve this problem in an integrated manner, we developed a fast simulated annealing approach. To the best of our knowledge, we are the first to develop a local search approach to resolve airline disruptions in an integrated way. This approach is compared with traditional approaches, and an experimental study is done to evaluate different neighbour generation methods, and to investigate different recovery scenarios and strategies. The comparison is done using real world data from KLM Royal Dutch Airlines. Here, we show that our approach resolves disruptions quickly and in a cost-efficient manner, and that it outperforms traditional approaches. Compared to naive delay propagation, our method saves 40% in non-performance costs. Moreover, while most airlines use tools that consider resources separately, our approach shows that integrated disruption management is possible within 30 seconds.

Figures (11)

• Figure 1: Example schedule consisting of two aircraft and three crew members. Here, the horizontal axis represents time, and the flights are drawn as blocks. The serrated edge of a block means that the flight departs or lands at an outstation.
• Figure 2: Illustration of the swap neighbour. Here, there is an overlap between the flights KL982 and KL923, which we try to fix. The rotation considered for moving ($R$) is highlighted orange (KL923 and KL924), and we try to swap this with the green highlighted flights (KL1531 and KL1532).
• Figure 3: Illustration of the move neighbour. Here, there is an overlap between the flights KL982 and KL923, which we try to fix. The rotation considered for moving ($R$) is highlighted orange (KL923 and KL924), and we try to move this in between the green highlighted flights (KL1822 and KL1711).
• Figure 4: Comparison of KPI values when recovering just the tail schedule in case of the FD scenario.
• Figure 5: Comparison of KPI values when recovering just the tail schedule in case of the UR scenario.