Abstract:

Underground station facilities are tightly coupled through physical connections, functional dependencies, and information interactions. During flood events, such coupling relationships can result in cascading failures, where damage to a critical facility may trigger systemic risks. Existing flood vulnerability assessment methods often regard facilities as isolated units, ignoring the coupling effects and risk transmission mechanisms, making it difficult to accurately characterize the damage propagation paths. This paper proposes a flood vulnerability cascading analysis method for underground station facilities by integrating knowledge graphs (KGs) and large language models (LLMs). First, a three-domain knowledge graph consisting of “object-behavior-state” is constructed to represent facility relationships. Second, a cellular automaton model is developed to simulate flood evolution coupled with facility component interactions. Third, flood vulnerability assessment and cascading effect inference are performed by constraining the LLMs with the knowledge graph. Finally, a large-scale underground station in Daxing District, Beijing, is selected as a case study, along with the DeepSeek-R1 series model, for experimental analysis. The results show that the proposed method can effectively identify spatial and functional changes of facilities under flood scenarios and reveal risk propagation paths. The reasoning process exhibits strong robustness and high interpretability. Compared with expert-defined benchmark cascade paths, the method achieves higher accuracy and logical consistency in terms of node matching rates and sequence matching accuracy. The findings provide theoretical support and technical reference for the formulation of emergency strategies and the enhancement of system resilience in underground stations.

Key words: underground station, flood vulnerability, cascading effect, knowledge graph, large language models