Scene-centric multi-vessel trajectory prediction via GRU-SAT

The continued growth of vessel traffic in strategic waters demands analytical tools that strengthen the situational awareness provided by conventional Vessel Traffic Services (VTS). Most AIS-based trajectory predictors model each vessel independently, limiting their ability to capture inter-vessel dependencies in dense traffic. This study presents GRU-SAT, a scene-centric multi-vessel trajectory prediction framework that pairs a Gated Recurrent Unit (GRU) temporal encoder with an inter-vessel Self-Attention Transformer (SAT) module, jointly predicting the future trajectories of all vessels in a scene in a single forward pass. A pairwise-distance regression head, supervised by the minimum inter-vessel distance over the prediction horizon, is attached to the shared scene representation as a multi-task regularizer and provides a continuous proximity score at inference. Evaluated on real-world AIS scenes from Busan Port, GRU-SAT improves trajectory accuracy over six scene-centric baselines spanning sequence-only and interaction-aware predictors, achieving an Average Displacement Error (ADE) of 1.036 km and a Final Displacement Error (FDE) of 1.584 km on the held-out test partition. The auxiliary objective further reduces close-pair pairwise-distance error compared with a trajectory-derived baseline, providing VTS operators with a continuous situational-awareness aid rather than a safety-critical collision-alerting system.