Trustworthy Knowledge Graphs for Grid Digitalization: Standards-Native CIM/CGMES vs. LLM-Extracted Graphs in Power-System GraphRAG

As utilities adopt LLM-based assistants over CIM/CGMES models, a basic reliability question arises: should the underlying knowledge graph come from parsing the engineering standard, or from an LLM extraction over the same content? We answer this inside an otherwise frozen GraphRAG pipeline on two ENTSO-E CGMES conformity networks (MicroGrid, 233 objects; SmallGrid, 1,398 objects), using the CGMES export itself as a schema-grounded oracle for fidelity. LLM extraction is precise but incomplete (near-zero schema contradiction, 34–53% of the standard’s structure recovered); raw CIM is exact but retrieval-verbose; on the larger network the LLM-extracted graph wins under a fixed 8,000-character context budget (98% for the LLM-extracted graph vs. 48% for the standards-native graph on SmallGrid). The advantage is representational selectivity, not higher graph correctness. A deterministic semantic compression of CGMES recovers part of that advantage at zero LLM cost. Code, prompts, the schema-grounded scorer, and audit trails are released as a reusable benchmark for trustworthy grid-AI evaluation.