Neural Networks for Aircraft Trajectory Prediction: Answering Open Questions About Their Performance

Abstract

The increase in air traffic in the recent years has motivated the development of technologies to monitor air space and warn about possible collisions by predicting the trajectories that will be followed by aircraft. In this field, neural networks have become prominent thanks to their potential to learn to predict maneuvers without providing aspects that are difficult to model such as atmospheric conditions, or detailed aircraft parameters. A variety of models have been proposed; however, these are often tested in very limited setups, leaving many unanswered questions about how they perform in certain conditions, or whether or not their accuracy can be improved by training models for specific trajectories, using additional features, predicting more distant points directly, etc. This may be problematic for researchers or developers of these systems, who have no way of knowing what strategies will yield the best results. We have identified ten open research questions that have not been answered through in-depth testing. This motivated us to carry out a novel experimental study that performs aircraft trajectory prediction with several dozens configuration variants to answer the aforementioned questions by means of a much more complete evaluation. Some of the conclusions of our study stand in contrast with some popular practices in the state of the art, which casts some doubts on the simplicity of their application; for example, differential features are crucial for proper performance but are not mentioned by most studies, while complex, more elaborate models may lead to worse results than simple ones. Other important insights include the benefit from specialized models in more challenging scenarios, the influence of the known trajectory length in said scenarios, the step degradation of predictions when predicting further into the future, or the detrimental effect of adding additional features. These insights should help guide future research about the application of neural networks when it comes to aircraft trajectory prediction and their eventual inclusion in final systems.