Prediction of aircraft trajectories for air traffic control using machine learning approaches

Marko Hrastovec (2018) Prediction of aircraft trajectories for air traffic control using machine learning approaches. PhD thesis.

Preview

PDF Download (4Mb)

Abstract

Air traffic is facing great challenges for the future. The economic crisis has brought a burden of cost savings, while the increase of traffic requires investments in research and development to find new paradigms for safe operations. One of the most important aspects in all future plans is better trajectory calculation, or better knowledge where the aircraft is going to be at a certain time. When positions are known, the planning can optimize flying paths to be cost efficient and safe, which is very important as the traffic becomes denser every day. Aircraft operators are planning flight paths with minimum costs, but they are not optimizing them for conflicts with other aircraft, and for airspace optimizations. Air traffic control and airspace restrictions are taking care of that. Soon, this present model will not provide enough throughput for all aircraft that want to fly. Our research is putting a stone in the mosaic of trajectory prediction and airspace optimization. In the future, aircraft will share data about their planned paths with air traffic control and aircraft in vicinity. Since air traffic is a highly regulated and expensive business, it takes a very long time before changes are implemented. Until then, we have to find alternative ways for better trajectory predictions, which will allow us to plan and optimize traffic, and to increase throughput. The ground control records the data about actual flight paths acquired by radars. Some weather data can be also acquired with a new generation of Mode-S radars. Pure aircraft performance data are enriched with weather and flight plan data into a joint knowledge database. For every new flight, we search in the database for flights similar to the incoming one. If we know how similar flights behaved in the past, we can predict the performances of a new flight, and can calculate the planned flight trajectory more accurately. Our goal is to predict trajectories better than using static models of aircraft performances. With existing prediction methods we predict for the same type of aircraft on a specific path the same trajectory every time. In that way, we have a prediction that deviates the least from the majority of flights. On the other hand, we predict a trajectory that does not fit any flight. With our approach, we want to take into account other factors such as aircraft operator, final destination, time of flight, etc., and every time predict a different trajectory suited to fit exactly to the considered flight. Operator and similar attributes are all factors that do not influence the flight directly. The destination, for instance, determines the distance of flight and therefore determines, how much fuel is on-board. More fuel means more weight and different flight characteristics. Similarly, we can assume that each operator operates airplanes differently than others, or carries different type of passengers that have usually more or less luggage than others. All these factors are not very well measurable, but they do affect flight performances. We use machine learning to find the flights in the database that are the closest to the one being predicted. With the assumption that flights with similar features flight similarly, we expect to predict more accurate trajectories than with static models and default parameters. We tested many machine learning methods and found the ones that perform the best on our data. We also adapted standard machine learning algorithms for our needs and large amounts of data. We have used machine learning predictions instead of static nominal values in widely used trajectory calculation model. The methods using only aircraft type are widely used in aviation, but they lack the capability to adapt to each flight individually. In our opinion, such rigid and static usage of aircraft type is an important cause for poor predictions. The results show that our predictions methods using individually customized predictions are more accurate than predictions based on aircraft type. We have shown that our methods are comparable with standard machine learning methods. The solution that we propose, is deployed as a web service, to which users can send flight details and get back parameters suited for a particular flight. Because the parameters are in the same form as in the widely used Base of Aircaft Data (BADA) model, legacy air control applications could use this service instead of static BADA database, and improve their trajectory calculations. In that way, a minimal change of the air control applications is needed. Trajectory calculations can remain unchanged, but with better input parameters, they can predict more accurately.

Actions (login required)