Safety-screened physics-informed hierarchical diffusion for near-real-time AIS-based vessel trajectory prediction

IntroductionAutomatic Identification System (AIS)-based vessel trajectory prediction is an important component of maritime traffic safety and decision support in busy waterways, port approaches, and restricted waters. In such areas, prediction accuracy alone is insufficient: predicted motion should remain physically plausible, evidently unsafe outputs should be filtered or corrected, and inference should be fast enough for operational use. Physics-based models are interpretable but require vessel or environmental information that is not fully available from AIS records, whereas data-driven models can learn complex motion patterns yet may produce physically inconsistent or unsafe trajectories in dense encounters.MethodsThis paper presents Risk-Prior Hierarchical Joint Diffusion (RP-HJD), a hierarchical diffusion method with physics-informed conditioning and safety-aware candidate screening for near-real-time vessel trajectory prediction. RP-HJD extracts behavior-related physical priors from historical AIS states and local scene information, generates multimodal future trajectories through hierarchical diffusion, and ranks the retained candidates after safety screening and local correction.ResultsOn Yangtze River AIS data with a 10-min prediction horizon, RP-HJD achieves a mean average displacement error (ADE) of 68.9 m (23% lower than AgentFormer), a safety violation rate (SVR) of 0.27%, and a mean inference latency of 76 ms per sample on an NVIDIA V100 GPU.DiscussionAblation, scenario-difficulty, high-risk-encounter, scalability, and Danish Maritime Authority public-data analyses provide complementary evidence on the contributions of the staged pipeline, safety-related output behavior, and performance under the evaluated settings.