Physics-constrained transformer learning for long-term ship trajectory prediction (BATFormer)

Accurate vessel trajectory prediction boosts navigation safety. Though Transformer excels at time-series modeling, mainstream methods simplify ships as inertial mass points without explicitly modeling steering intentions, causing large straight-line inertial bias and poor capture of abrupt course shifts in long-term forecasting. We present BATFormer, a behavior-aware multi-task framework with dual decoders for joint trajectory prediction and maneuver intention classification. A turn-rate-calibrated labeling strategy provides structured prior to heighten steering perception. A physical consistency inference module prunes kinematically invalid predictions via Top-k pruning and dynamic r-vicinity constraints. Tests on large datasets show BATFormer cuts 6-hour Average Displacement Error by 19.3% over SOTA models, with stronger robustness, physical compliance and cross-region generalization.