ABSTRACT Accurate prediction of ship trajectories using automatic identification system (AIS) data is critical for intelligent marine traffic management. Although Transformer‐based architectures have improved long‐term prediction performance, their short‐term accuracy remains suboptimal. This study introduces TG‐generative pre‐trained Transformer (GPT), a lightweight trajectory prediction model that embeds a gated recurrent unit (GRU) within a tiny GPT to enhance prediction precision while reducing computational complexity. Using the Danish Maritime Authority AIS dataset containing over 1.2 million trajectory records across multiple vessel types, TG‐GPT was evaluated against state‐of‐the‐art models. For specified routes, TG‐GPT achieves 72%, 85% and 87% reductions in mean absolute error (MAE) at 1, 2, and 3 h prediction horizons, respectively, compared with CLSA. For random trajectory predictions, MAE is reduced by 6%, 5% and 9% relative to TrAISformer. The model further achieves 20% fewer parameters, 10% faster training and 40% faster trajectory generation than Transformer‐based baselines. Unlike previous Transformer‐based models, TG‐GPT introduces a lightweight GRU‐enhanced architecture that significantly improves short‐term prediction accuracy with fewer parameters. These results highlight TG‐GPT's potential as a practical solution for real‐time maritime navigation and autonomous vessel control systems.
Chen et al. (Thu,) studied this question.