Predicting maritime accident risks in advance and quantitatively assessing the likelihood of accidents can serve as a key solution for enhancing maritime safety and preventing casualties. However, existing studies have predominantly relied on limited ...
Predicting maritime accident risks in advance and quantitatively assessing the likelihood of accidents can serve as a key solution for enhancing maritime safety and preventing casualties. However, existing studies have predominantly relied on limited structured data such as weather conditions or marine traffic information, and often remained at the level of binary accident classification. Moreover, most prior approaches lacked legal justification and interpretability in explaining prediction outcomes.
This study proposes an integrated maritime accident prediction framework that combines a quantitative prediction model developed using weather and marine traffic datasets with an LLM-based accident scenario generation module grounded in maritime legislation and tribunal rulings. Derived variables reflecting regional maritime operational characteristics were constructed, and probabilistic calibration was applied to improve the reliability of predicted risk scores. The prediction outputs were then integrated with legal and tribunal case data through a Retrieval-Augmented Generation (RAG) architecture, enabling automatic generation of accident causes, legal grounds, and recommended countermeasures.
The proposed model goes beyond conventional accident probability prediction by connecting predicted risk levels with legal reasoning and simultaneously generating actionable, explainable decision-support information. By linking prediction outcomes with an LLM-driven RAG system, the framework enables autonomous reasoning capable of answering why a situation is risky and which legal provisions justify specific mitigation actions. To evaluate the quality and reliability of the generated scenarios, an LLM-as-Judge quantitative assessment was conducted. The proposed model demonstrated superior performance across all evaluation metrics compared with both the baseline model and the single-input model, confirming its effectiveness in producing coherent, legally grounded, and context-aware accident reasoning outputs.
The proposed approach demonstrates the potential to serve as a core technological foundation for intelligent maritime early-warning and decision-support infrastructure, ultimately improving both safety and operational efficiency within maritime traffic systems.