VTLLM: A Vessel Trajectory Prediction Approach Based on Large Language Models

In light of the rapid expansion of maritime trade, the maritime transportation industry has experienced burgeoning growth and complexity. The deployment of trajectory prediction technology is paramount in safeguarding navigational safety. Due to limitations in design complexity and the high costs of data fusion, current deep learning methods struggle to effectively integrate high-level semantic cues, such as vessel type, geographical identifiers, and navigational states, within predictive frameworks. Yet, these data contain abundant information regarding vessel categories or operational scenarios. Inspired by the robust semantic comprehension exhibited by large language models (LLMs) in natural language processing, this study introduces a trajectory prediction method leveraging LLMs. Initially, Automatic Identification System (AIS) data undergoes processing to eliminate incomplete entries, thereby selecting trajectories of high quality. Distinct from prior research that concentrated solely on vessel position and velocity, this study integrates ship identity, spatiotemporal trajectory, and navigational information through prompt engineering, empowering the LLM to extract multidimensional semantic features of trajectories from comprehensive natural language narratives. Thus, the LLM can amalgamate multi-source semantics with zero marginal cost, significantly enhancing its understanding of complex maritime environments. Subsequently, a supervised fine-tuning approach rooted in Low-Rank Adaptation (LoRA) is applied to train the chosen LLMs. This enables rapid adaptation of the LLM to specific maritime areas or vessel classifications by modifying only a limited subset of parameters, thereby appreciably diminishing both data requirements and computational costs. Finally, representative metrics are utilized to evaluate the efficacy of the model training and to benchmark its performance against prevailing advanced models for ship trajectory prediction. The results indicate that the model demonstrates notable performance in short-term predictions fFor instance, with a prediction step of 1 h, the average distance errors for VTLLM and TrAISformer are 5.26 nmi and 6.12 nmi, respectively, resulting in a performance improvement of approximately 14.05%), having identified certain patterns and features, such as linear movements and turns, from the training data.