Abstract: Knowledge graphs can efficiently integrate the knowledge of a hydraulic project and advance digitalization significantly. However, traditional methods face challenges in cross-domain ontology construction, high annotation cost, and limited transferability. This study constructs an automated knowledge graph framework that leverages large language models (LLMs) for cross-domain intelligent hydraulic construction. The method has two parts: (1) constructing a shared ontology through terminology discovery, co-occurrence networks, and LLM reasoning to resolve cross-domain semantic inconsistencies; (2) extracting enhanced knowledge, combining prior knowledge, hybrid retrieval, dynamic prompting, and chain-of-thought reasoning to reduce LLM hallucinations. Numerical experiments show the shared ontology achieves structural consistency, with cross-domain knowledge extraction reaching an average F1 score of 84.5, outperforming conventional models. This validates the method's effectiveness in multi-subdomain knowledge integration with reduced annotation requirements.

Key words: hydraulic construction, knowledge graph, large language model, prompt engineering, text retrieval