A vessel trajectory prediction method integrating Conditional Variational Autoencoder with TrAISformer

Predicting vessel trajectories in complex maritime routes is challenging because several future paths may be plausible and autoregressive errors accumulate over long horizons. We propose CVAE-TrAISformer, an extension of TrAISformer inspired by conditional variational autoencoders (CVAEs) that augments its stochastic categorical predictor with context-conditioned latent features. A latent branch aggregates the available autoregressive context, samples latent variables through reparameterization, and fuses them with Automatic Identification System (AIS) embeddings before the Transformer; training uses categorical prediction losses and annealed Kullback-Leibler regularization. On the Danish Maritime Authority AIS benchmark, oracle best-of -16 mean prediction errors at 1, 2, and 3 h decrease from 0.48, 0.94, and 1.64 nautical miles for TrAISformer to 0.39, 0.77, and 1.30 nautical miles, respectively, with 0.96% more parameters. A separate paired ten-run fixed-latent evaluation identifies a significant reduction in 3-h endpoint error for complex trajectories (Holm-adjusted p = 0.023). Candidate-set analyses show lower displacement errors across evaluated set sizes but no increased spatial dispersion or consistent branch-coverage advantage. Kinematic diagnostics show more frequent large course-over-ground changes. These findings support improved oracle-evaluated long-horizon accuracy while identifying limits of latent conditioning.