TM-LLM: Token-Motion Embedding Fusion for Large Language Model-Based Vessel Trajectory Prediction

Automatic Identification System (AIS) is a key resource in monitoring maritime traffic, providing essential data such as vessel positions, speed, and course changes over time. Despite extensive research, predicting vessel trajectories using AIS time-series data remains challenging due to the intricate movement patterns and the diverse characteristics of vessels. Recently, Large Language Model (LLM) have gained recognition for their ability to transfer knowledge across time-series analysis. However, LLMs face significant difficulties in effectively representing the motion patterns in AIS time-series data. To tackle this problem, we propose a novel model named TM-LLM, designed for vessel trajectory prediction using historical AIS data. In TM-LLM, we develop a token and motion embedding layer that effectively captures both local and global patterns within sequential AIS data. These embeddings are subsequently combined uniformly using a convolutional fusion layer, making them compatible for input into the LLM. We then fine-tune a pre-trained LLM to help it learn the patterns in historical AIS data and effectively predict future vessel trajectories. Experimental results demonstrate that TM - LLM achieves prediction performance comparable to state-of-the-art deep learning models. Furthermore, TM - LLM has a significantly shorter prediction time, making it well-suited for real-time prediction scenarios.