Integrated real-time airline control system using machine learning

点击翻译按钮获取中文摘要

Aircraft turnaround (TA) covers all processes that are required to prepare the arriving aircraft at the airport for its next flight segment. The interaction and sometimes deficient coordination of staff, passengers and systems involved in the process as well as the limited availability of resources can lead to disruptions which, together with reactionary network effects, are the main causes of flight delays in densely timed flight schedules. On the light side, TA offers the potential to control ground operations, thus supporting on-time performance and avoiding reactionary delays and missed passenger, crew, and aircraft connections. The appropriate TA control measures are based on alternative process executions, changed task priorities, and a carefully selected allocation of resources. However, these actions are decided by dispatchers, who currently act based on incomplete information and poor or even missing predictions. Moreover, the choice of these recovery measures is not holistically made by passenger, aircraft, and crew recovery, which are labeled as "integrated" optimization in the literature already when only two of these planning steps are handled simultaneously. The combination of these models plus the TA under constrained conditions (limited resources, capacity) at the airport and in the airspace is entirely new. One reason for this is the excessive runtime of conventional solvers for this setup. We plan to overcome this barrier by supervised machine learning (ML) methods based on pre-computed and similar classified schedule disruptions. Initial research results show that ML-based approaches offer runtime improvement for trivial heuristics. Based on the previously published research results on decision models in TA and at the level of individual fleet rotations, these are integrated into an MI(N)LP optimization model and extended by network-wide measures such as aircraft swap or delay management. This enables the prediction of reactionary delays across the remaining time horizon of the schedule and transforms transfer relationships concerning passengers, aircraft, and crew into local evaluation functions. The selected ML method will be implemented to allow comparison of disruptions during flight operations with ML predicted disruptions being solved with relevant lead time. This fixes problem-specific expected optimal measures as additional constraints to restrict the solution space of classical optimization and supports common solvers in convergence. We aim to identify a suitable supervised training based ML solution procedure for the automated solution of decision processes within dynamic situations at airports for different airline networks and in a time-rolling manner.