A proposed method for predicting civil aircraft trajectories using the ADS-B system based on deep tabular learning.

Abstract

The continuous advancement of Air Traffic Control (ATC) has resulted in the increasing importance of efficient and reliable airspace management for the safety and efficiency of modern air transportation. The rapid growth of global air traffic has created a great demand for intelligent systems to accurately monitor and control aircraft operations. Recent advances in artificial intelligence have opened up the possibility of data-driven approaches that can significantly improve the performance of air navigation and trajectory prediction systems.In this study, two deep tabular learning models, TabNet and TabNSA, were employed for aircraft trajectory prediction with Automatic Dependent Surveillance-Broadcast (ADS-B) data. The experimental results showed that the proposed models performed better than a number of popular deep learning approaches such as CNN-LSTM and CNN-BiLSTM. This performance gain was particularly clear in altitude prediction, which remains one of the most difficult tasks in aircraft trajectory forecasting. In particular, TabNSA outperformed TabNet with around 7% lower RMSE and CNN-BiLSTM with 39% lower RMSE. TabNSA further improved the prediction accuracy in terms of both RMSE and MAE by around 50% over TabNet for latitude and longitude prediction. Additionally, qualitative comparisons of 2D and 3D flight trajectories showed that the predicted trajectories by TabNSA were highly similar to the actual flight trajectories, consistently better than the benchmark models and also exhibiting a strong correlation with the TabNet predictions.