Long-horizon vessel trajectory prediction with step-wise traffic-pattern priors and sequential learning

Reliable long-horizon vessel trajectory forecasting is essential for Vessel Traffic Services (VTS) to support early traffic organization and safety monitoring. However, purely data-driven multi-step predictors often produce operationally implausible forecasts, which can undermine trust in strategic look-ahead warnings. This paper proposes a traffic-pattern-informed framework that continuously conditions long-horizon forecasting on local historical Speed-Over-Ground and Course-Over-Ground distributions extracted around the current vessel state, so as to encourage more plausible step-wise predictions. In addition, a sequential learning strategy with horizon-specific predictors is employed to improve long-horizon stability by conditioning each step on previously predicted state and updated local patterns. Experiments on two real-world AIS datasets (US Coast Guard and Danish Maritime Authority) show consistent improvements over representative baselines, reducing cumulative displacement error by up to 53.9%, with notable gains in constrained scenarios such as port approaches. The proposed framework is suited for strategic look-ahead in VTS, where stable long-horizon forecasts are required to support early conflict screening and traffic organization. • A traffic-pattern-informed framework is proposed for long-horizon trajectory prediction. • Grid-based SOG/COG distributions are used as local traffic-pattern priors. • Traffic-pattern priors are updated step-wise during recursive prediction. • Horizon-wise predictors are trained to reduce error growth in recursive forecasting. • Results reveal when priors help and where failures may arise.