Hybrid deep learning models for ship trajectory prediction in complex scenarios based on AIS data

Ship trajectory prediction plays a vital role in situation awareness and maritime safety monitoring systems. Currently, the mainstream ship trajectory methods focus on single ships, and little work has been done to consider the interaction between ships. Therefore, aiming at improving the ship trajectory prediction accuracy and giving a comprehensive perspective of maritime surveillance, we proposed an integrated model with two sub-models. (1) the S-TGP model, combining Time Convolutional Network (TCN) and Gated Recurrent Unit (GRU) for single-ship trajectory with high accuracy and high generalization. The S-TGP model takes advantage of the parallel computing ability of TCN and the ability to estimate long-term correlation in the historical data. (2) the MVS-TGP model, integrating variational autoencoder (VAE) with S-TGP, for multi-ship trajectory prediction in complex scenarios. Our contributions include: (1) enhancing the accuracy of single-ship trajectory prediction with the S-TGP model; (2) improving collaborative prediction capabilities for multiple ships with the MVS-TGP model; and (3) providing real-time prediction and monitoring capabilities for maritime surveillance. Validated on AIS data from three regions, our models demonstrate superior performance and robustness compared to existing methods. The results show that the proposed models are effective in different environments and outperform the other models quantitively and qualitatively.