A Structured Question–Answer Dataset for Large Language Models in Power System Dynamics and Operations

This paper introduces Power Grid Question and Answer (PowerGridQA), a curated question–answer dataset for large language models in power system dynamics and grid operations. The dataset combines content from authoritative sources, including scholarly articles, Textbooks and official North American Electrical Reliability Corporation (NERC) reliability publications. The resulting corpus of over 28,000+ entries divide into Power System Theory, NERC reports and technical documents, and a handful of reasoning type entries. The dataset is intended to support research on knowledge grounding, reasoning fidelity, and source-aware question answering for power-system applications. Each are technically verified Q/A pairs, emphasizing definitions, equations, and cause–effect relationships. PowerGridQA bridges the gap between traditional power system literature and modern AI models, supporting the development of factual, reliability-compliant language systems for real-world grid intelligence. A comprehensive analysis is done using a combination of traditional lexical metrics like Exact Match (EM), F – Score (F1), advanced semantic metrics like Bidirectional Encoder Representations from Transformers (BERTScore), and graph-based structural tests, including syntactic divergence analysis and thematic community detection. Furthermore, the structural analysis visually confirms the thematic coherence of the datasets through distinct topic clusters and illustrates the large syntactic gap between questions and answers in the reasoning subset.