LLM4STP: A large language model-driven multi-feature fusion method for ship trajectory prediction

• Propose LLM4STP, a novel multi-feature fusion method integrating LLMs and maritime knowledge for improved ship trajectory prediction. • Introduce an adaptive graph-masked Transformer for dynamic ship interaction modelling, capturing complex spatial relationships. • Combine local convolution and global LLM inference for effective multi-scale temporal modelling. • Establish a system for trajectory uncertainty quantification using 2D Gaussian heatmaps and Geohash-based encoding. • Outperform state-of-the-art methods on three AIS datasets, achieving lower ADE, FD, and FDE. Ship trajectory prediction (STP) is a critical research focus for enhancing maritime traffic situational awareness and supporting navigational decision-making in intelligent transportation systems. The accuracy and robustness of prediction models significantly affect maritime safety and shipping efficiency. Despite advances driven by Automatic Identification System (AIS) data and deep learning techniques, key challenges remain unresolved, including dynamic multi-ship interaction modelling in complex marine environments, multi-scale temporal dependency reasoning, trajectory uncertainty quantification, and effective integration of maritime domain knowledge. Existing methods based on Large Language Models (LLMs) improve generalisation through pre-trained knowledge but fall short in real-time interaction topology modelling, geospatial semantic representation, and uncertainty estimation. To address these limitations, this paper proposes LLM4STP, a novel LLM-driven multi-feature fusion method for STP. LLM4STP establishes a new paradigm by deeply integrating LLMs with maritime domain knowledge to collaboratively predict ship trajectories. The model features an adaptive graph-masked Transformer to dynamically capture ship interaction topologies, hierarchical temporal reasoning to jointly model local manoeuvring behaviours and macroscopic navigational intent, and an innovative fusion of Gaussian probability distribution heatmaps with GeoHash-based geospatial encoding to quantify trajectory uncertainty while preserving semantic continuity. Experiments on three representative water areas demonstrate that LLM4STP consistently outperforms nine state-of-the-art (SOTA) methods, as validated by key metrics including Average Displacement Error (ADE), Fréchet Distance (FD), and Final Displacement Error (FDE). Moreover, the few-shot learning experiments demonstrate that LLM4STP can match the performance of models trained on the full dataset using only 20 % of the training data, highlighting its efficiency and strong adaptability in data-scarce environments. The ablation studies empirically validate the significance and distinct contribution of each component within the proposed model architecture. This study integrates LLM into maritime traffic scenarios, making significant contributions to enhancing the robustness, accuracy, and interpretability of STP in high-interference environments. The source code is openly accessible at https://github.com/Joker-hang/LLM4STP .