LLM-Based Maritime Risk Scenario Generation for Intelligent Vessel Traffic Services

Abstract

Maritime accident prevention requires timely, interpretable risk assessment that can support operational decision-making in Vessel Traffic Service (VTS) environments. We propose an integrated framework that combines Large Language Models (LLMs) with machine-learning-based risk inference to unify accident risk probability estimation and dynamic scenario generation. Heterogeneous data sources–Automatic Identification System (AIS) trajectories, meteorological observations, historical accident statistics, and tribunal adjudication reports–are standardized through a two-path LLM pipeline and aligned on a uniform spatial grid. A gradientboosting classifier with SHAP (SHapley Additive exPlanations)-based explainability identifies key risk drivers, and isotonic regression calibration produces well-calibrated, continuous risk scores. For scenario generation, a hypergraph-based retrievalaugmented generation (RAG) knowledge base encodes multiway relationships among accident cases, causal factors, and maritime regulations; retrieved evidence is combined with calibrated risk scores in few-shot, Chain-of-Thought (CoT) prompts to generate interpretable accident scenarios with a cause-progression-outcome structure. A web-based dashboard with on-device text-to-speech delivers grid-level risk visualization, narrative risk scenarios, and actionable navigational advisories within a unified prediction-explanation-response interface for maritime safety management. Additional results are available at the project page: https://thrillcrazyer.github.io/MaritimeRiskScenarioGen