Objective: Typhoons, characterized by sudden onset, extensive geographic impact, and considerable destructive power, pose recurring threats to the stability and safety of power grid systems, particularly in China's coastal regions. As extreme weather events become more frequent due to climate change, conventional emergency management approaches are inadequate. These methods often suffer from fragmented knowledge sources, inefficient information extraction, and limited support for intelligent decision-making. Hence, this article proposes an integrated technical framework that combines knowledge graphs with large language models (LLMs). This study aimed to improve risk perception, enhance decision-making accuracy, and bolster emergency response effectiveness in typhoon-triggered power grid incidents. Zhejiang Province, a coastal area frequently impacted by typhoons, was selected as the demonstration case for the framework. Methods: The framework of knowledge graph construction included the design of two graph types: one derived from accident reports and the other derived from emergency plans. The accident-based knowledge graph was structured according to the Triangular Framework for Public Security Science and Technology at the schema layer. It organized knowledge into three primary dimensions: emergency events, affected infrastructure, and corresponding emergency management strategies. Meanwhile, the emergency-plan-based knowledge graph was structured based on the electric power production life cycle, covering key stages such as power transmission, power transformation, power distribution, power utilization, and energy storage. Both graphs worked together to support emergency planning. The system employed a hybrid approach at the data processing layer that integrated BERT with a bidirectional long short-term memory network. This hybrid model performed named entity recognition and relationship extraction. The extracted entities and relationships were visualized to improve model interpretability, enabling domain experts to validate and understand the underlying information. An enhanced discriminative similarity algorithm was introduced in the knowledge fusion process. Initially, cosine similarity and Pearson correlation filtered out low-relevance entity pairs. High-similarity entities were then semantically validated using the LLM, ensuring accurate fusion and reducing erroneous entity alignments. Experimental results showed a 10.11% improvement in accuracy compared to conventional methods. The final knowledge graphs were stored in the Neo4j graph database, which supported interactive visualization and real-time query functionalities. The system enabled intelligent reasoning for handling real-world disaster scenarios in the application stage. Using the Cypher query language, this study conducted a fuzzy query based on disaster descriptions. Relevant information was retrieved from the knowledge graph as a structured knowledge base. The ECO-STAR prompting template was used to guide the model in generating targeted risk analyses and emergency recommendations. Results: A case study was conducted in Zhejiang Province to validate the proposed framework. The results showed that integrating knowledge graphs and LLMs improved semantic precision. The integration also enhanced the relevance of decision support outputs. In addition, it reduced hallucination phenomena that often occurred when general-purpose LLMs were applied in specialized domains. Conclusions: This study highlights the value of leveraging LLMs in constructing an emergency knowledge graph for power grids during typhoons. The proposed method offers a scalable, intelligent solution for managing power grid emergencies during typhoons and serves as a valuable reference for enhancing the disaster resilience of energy systems.