Machine learning assisted dynamic copy generation approach for integrated aircraft and passenger recovery problem

Frequently confronted with unexpected and cascading disturbances, airlines are obliged to develop cost-effective recovery schedules for their operational assets and passengers within constrained computational time frames. Despite the growing research interest in integrating aircraft and passenger recovery decisions, the critical arrival delay cost of passengers are either modelled from an individual passenger perspective with high complexity or approximated via discrete delay levels and aggregated flight delays. In this study, a mixed integer programming model, which integrates aircraft and passenger recovery subproblems, is developed based on dynamic flight copies with unfixed delaying values. To tackle the corresponding intractability regarding column-wise and row-wise increase of model size induced by the high cardinality of potential aircraft routes, passenger paths and underlying flight copies, a simultaneous column-and-row generation algorithm is developed to generate pricing variables and linking constraints on-the-fly. The algorithm is further accelerated by filtering low quality flight connection opportunities predicted by a deep graph neural network model. To alleviate the computational expense of preparing large-scale training problem instances with the column-and-row generation algorithm, we prepare small-scale datasets to train the supervised learning model. Subsequently, a graph sampling technique is applied to extend and generalize the pre-trained model to solve large-scale problem instances. Through extensive computational experiments using real-world operation data, our methodology efficiently derives high-quality recovery solutions within a 30 min time limit, achieving no more than a 1.13% optimality gap. Additionally, our experiments demonstrate that a significant reduction in passenger arrival delays can be expected using the presented model, yielding substantial benefits to passenger-centric airline operations.