Abstract | Machine learning techniques to support decision making processes are in trend. These are particularly relevant in the context of flight management where large datasets of planned and realised operations are available. Current operations experience discrepancies between planned and executed flight plan, these might be due to external factors (e.g. weather, congestion) and might lead to sub-optimal decisions (e.g. recovering delay (burning extra fuel) when no holding is expected at arrival and therefore it was no needed). Dispatcher3 produces a set of machine learning models to support flight crew pre-departure, with estimations on expected holding at arrival, runway in use and fuel usage, and the airline’s duty manager on pre-tactical actions, with models trained with a larger look ahead time for ATFM and reactionary delay estimations. This paper describes the prototype architecture and approach of Dispatcher3 with particular focus on the challenges faced by this type of data-driven machine learning models in the field of air transport ranging: from technical aspects such as data leakage to operational requirements such as the consideration and estimation of uncertainty. These considerations should be relevant for projects which try to use machine learning in the field of aviation in general. |
---|