Multi-fidelity Simulation Optimisation for Airline Disruption Management

Disruption to schedules is one of the major issues faced by the airline industry. Airline scheduling itself has seen much focus at the planning stage, allowing for ever more robust and efficient use of the resources available to airlines. A great effort in time and money is spent on producing an optimal schedule. However, due to the complex and stochastic nature of the industry, it is rare that a flight programme is operated as intended. Disruptive events arise due to various circumstances, from weather conditions to aircraft technical failures requiring unplanned maintenance. The impacts of such events can propagate through the system causing further delays and cancellations, particularly when the airline has high aircraft utilisation. In such an eventuality, the Airline Operations Control Centre (AOCC) must alter its schedule to reduce the repercussions for its finances, its reputation and passengers. Such alterations could include delaying or cancelling flights or exchanging aircraft. This paper will focus on the Aircraft Recovery Problem (ARP).

The complexity of the industry can make it difficult to determine what the consequences of any schedule alterations may be. Kohl et al. (2007) suggest that automation could play a large part in the identification and evaluation of potential recovery actions. Some of the complexity has been considered using deterministic models, a number of which have been proposed. However, these models have a limited capability of accounting for the various stochastic elements of the environment. Turn times, airport queueing times, maintenance times and flight durations are all stochastic, which may lead to further disruptions that the AOCC would wish to consider. Probabilistic models that achieve the levels of detail required are unlikely to be analytically tractable. Thus simulation seems to be a natural way to model the airline’s operations. However, a high-fidelity simulation model would have a non-negligible computation time. This creates issues for searching through the large solution space. To make the most of the simulation, it must be used selectively on solutions believed to be relevant.

Related posts