Noise robust aircraft trajectory prediction via autoregressive transformers with hybrid positional encoding

Aircraft trajectory prediction is vital for ensuring safe and efficient air travel while addressing challenges in complex and dynamic environments. Current trajectory prediction models often struggle in noisy scenarios due to their lack of robustness. This study introduces the Noise-Robust Autoregressive Transformer, a novel model that enhances prediction reliability by integrating noise-regularized embeddings within a multi-head attention equipped with hybrid positional encoding. This model effectively captures essential temporal-spatial relationships and manages positional information more precisely across varied trajectories. Moreover, we formulate the robust trajectory prediction problem as an autoregressive approach that models the encoding of historical data and the decoding of future positions as a sequence-to-sequence learning problem. Our approach effectively captures positional encodings for the complex spatial-temporal variations in aircraft trajectory prediction, improving long-term prediction accuracy while achieving real-time responsiveness. Extensive experiments on multiple datasets demonstrate our improvement over existing aircraft trajectory prediction methods.