Online Prediction of Mid-Flight Aircraft Trajectories with Multi-Timestep Markov Models

Abstract

Online trajectory prediction is central to the function of air traffic control of improving the flow of air traffic and preventing collisions, particularly considering the ever-increasing number of air travellers. In this thesis, we propose an approach to predict the mid-flight trajectory of an aircraft using models learned from historical trajectories. The main idea is based on Markov Models with transition probabilities for multiple timesteps, representing the location of aircraft as states and movement of aircraft over a certain number of minutes (i.e. timestep) as transition probabilities between states. Using our approach, one is able to make predictions of future positions of mid-flight aircraft for each minute up to twenty minutes into the future, and concatenate them to form the predicted trajectory of the aircraft for the next twenty minutes. We evaluated the effectiveness of the proposed approach using a dataset of historical trajectories over the USA. Using prediction accuracy metrics from the aviation domain, we demonstrated that our approach was able to make accurate predictions of future trajectories of mid-flight aircraft, achieving an improvement of 24.6% in horizontal error and 34.2% in vertical error over baseline models from conventional approaches, with each prediction requiring mere milliseconds to compute.