Big Data Machine Learning for Flight Planning

Abstract

Weather forecasts serve as a fundamentally important input to the flight planning process and can carry an inherent measure of uncertainty. Such uncertainties thus lead to a trajectory being planned that does not represent the most cost-optimal option. Weather forecast generation relies on numerical simulations of the earth’s atmosphere, which in turn rely on models imitating the physical processes involved. Technological advances have meanwhile lead to a surge in means to more efficiently process large amounts of data, commonly termed Big Data. Such processing includes the possibility of applying analysis and Machine Learning techniques. It is therefore of interest whether forecast uncertainties can be predicted using these means and whether these predictions in turn yield a benefit for the flight planning process. This thesis provides a feasibility evaluation of a data-centric approach to weather forecast uncertainty prediction and a subsequent validation of potential benefits to a flight planning engine’s measure of predictability. Eight Machine Learning algorithms are trained on this data using the discrepancy between forecast and re-analysis data. Doing so ensures that the algorithms learn an underlying pattern of forecast errors or uncertainties. A second algorithmic layer is further realized which leverages this information to determine the algorithm generating the most accurate prediction, per forecast instance. A validation data set spanning a year of data is utilized to serve as input data for the flight plan generation of three flights. These are then compared to the flight’s actual flown trajectories. Results indicate that algorithms’ predictions are able to decrease forecast uncertainty in a majority of cases. A heavy dependence on the world region the flight is performed in, is recorded. As such, no benefit to flight plan predictability is observed for a short haul flight in South East Asia, while a slight benefit is recorded for an intercontinental long haul flight. An operational realization is not recommended at the time of writing, as further validations covering more areas and a greater number of flights need to be performed to better gauge the boundaries in which the method is beneficial to the flight planning process. Further research is needed to understand the underlying patterns in algorithmic prediction performance and increase reliability.