Home > Research > Publications & Outputs > Multi-fidelity modelling approach for airline d...

Electronic data

  • 2020rhodes-leaderphd

    Final published version, 2.07 MB, PDF document

    Available under license: CC BY: Creative Commons Attribution 4.0 International License

Text available via DOI:

View graph of relations

Multi-fidelity modelling approach for airline disruption management using simulation

Research output: ThesisDoctoral Thesis

Publication date2020
Number of pages316
Awarding Institution
Thesis sponsors
  • Rolls Royce
Award date15/06/2020
  • Lancaster University
<mark>Original language</mark>English


Disruption to airline schedules is a key issue for the industry. There are various causes for disruption, ranging from weather events through to technical problems grounding aircraft. Delays can quickly propagate through a schedule, leading to high financial and reputational costs. Mitigating the impact of a disruption by adjusting the schedule is a high priority for the airlines. The problem involves rearranging aircraft, crew and passengers, often with large fleets and many uncertain elements. The multiple objectives, cost, delay and minimising schedule alterations, create a trade-off. In addition, the new schedule should be achievable without over-promising. This thesis considers the rescheduling of aircraft, the Aircraft Recovery Problem.

The Aircraft Recovery Problem is well studied, though the literature mostly focusses on deterministic approaches, capable of modelling the complexity of the industry but with limited ability to capture the inherent uncertainty. Simulation offers a natural modelling framework, handling both the complexity and variability. However, the combinatorial aircraft allocation constraints are difficult for many simulation optimisation approaches, suggesting that a more tailored approach is required. This thesis proposes a two-stage multi-fidelity modelling approach, combining a low-fidelity Integer Program and a simulation. The deterministic Integer Program allocates aircraft to flights and gives an initial estimate of the delay of each flight. By solving in a multi-objective manner, it can quickly produce a set of promising solutions representing different trade-offs between disruption costs, total delay and the number of schedule alterations. The simulation is used to evaluate the candidate solutions and look for further local improvement. The aircraft allocation is fixed whilst a local search is performed over the flight delays, a continuous valued problem, aiming reduce costs. This is done by developing an adapted version of STRONG, a stochastic trust-region approach. The extension incorporates experimental design principles and projected gradient steps into STRONG to enable it to handle bound constraints. This method is demonstrated and evaluated with computational experiments on a set of disruptions with different fleet sizes and different numbers of disrupted aircraft. The results suggest that this multi-fidelity combination can produce good solutions to the Aircraft Recovery Problem.

A more theoretical treatment of the extended trust-region simulation optimisation is also presented. The conditions under which a guarantee of the algorithm's asymptotic performance may be possible and a framework for proving these guarantees is presented. Some of the work towards this is discussed and we highlight where further work is required.

This multi-fidelity approach could be used to implement a simulation-based decision support system for real-time disruption handling. The use of simulation for operational decisions raises the issue of how to evaluate a simulation-based tool and its predictions. It is argued that this is not a straightforward question of the real-world result being good or bad, as natural system variability can mask the results. This problem is formalised and a method is proposed for detecting systematic errors that could lead to poor decision making. The method is based on the Probability Integral Transformation using the simulation Empirical Cumulative Distribution Function and goodness of fit hypothesis tests for uniformity. This method is tested by applying it to the airline disruption problem previously discussed. Another simulation acts as a proxy real world, which deviates from the simulation in the runway service times. The results suggest that the method has high power when the deviations have a high impact on the performance measure of interest (more than 20%), but low power when the impact is less than 5%.