Predicting 4D trajectories of aircraft using neural networks and gradient boosting machines
A data-driven aircraft trajectory prediction study
More Info
expand_more
Abstract
Data-driven trajectory prediction is one of the key pillars of the future ATM system. Recent research focuses on using novel data sources and machine learning algorithms to improve the performance of 4D trajectory prediction, enabling safer and more efficient routing of aircraft. In this paper a framework for data sourcing and preparation for such predictors is presented, as well as a comparison of three of the best performing prediction algorithms from literature to a baseline method. Currently such comparisons are lacking, making it hard to determine which techniques provide the best results. Using an ADS-B antenna and various online data sources a trajectory set of 40,000 trajectories is built. Two clustering methods are tested and it is found that clustering trajectories using Density Based Clus- tering for Applications with Noise (DBSCAN) performs poorly on our data set of arriving flights. Too many trajectories are classified as outliers while DBSCAN is not capable of separating the trajectories in distinct clusters. A clustering method based on the STARs of the airport is proposed, which performs better in terms of accuracy and efficiency. Finally, a baseline simulation using Aircraft Performance Models is compared to a deep neural network, a Long Short-Term Memory (LSTM) network and to Gradient Boosting Machines (GBM) for trajectory prediction. It is found that the latter outperforms the other methods overall, while it was expected that predictors based on LSTMs would provide more accurate results. It is concluded that long-term dependencies in trajectory data, on which LSTMs perform well, are less important than categorical indicators, on which GBMs perform better, in trajectory prediction.