Fusion of Large Language Model and Knowledge Graph for Intelligent Power-System Q&A

Abstract

Intelligent question-answering (Q&A) models are needed for modern power systems that can provide accurate, reliable, and explainable answers to engineers and operators. Conventional Q&A systems cannot always understand the question semantically or combine domain knowledge, and stand-alone Large Language Models (LLMs) can generate hallucinated or non-grounded answers. To tackle these challenges, this research introduces a fusion model that combines the power of LLMs with a Power Knowledge Graph (PKG) for intelligent power-system question answering. The proposed approach integrates LLMs' ability to understand language with that of structured knowledge representation by PKGs. First, z-score normalization is used to normalize the operational parameters, such as voltage, current, temperature, and load. After that, Bidirectional Encoder Representations from Transformers (BERT) is used to obtain semantic relations and domain-specific elements from textual data related to the power system. The PKG stores fault, operational, and maintenance knowledge, and a knowledge-enhanced retrieval mechanism extracts relevant information and adds it to the LLM prompt to boost the accuracy of reasoning and diminish hallucination. The model is assessed with the following tasks: Fault Type Identification, Environmental and Operational Causation, and Maintenance and Recovery Decision. The proposed Python-based implementation achieves BERT Score F1 values of 71.5%, 69.2%, and 67.8%, Faithfulness Scores of 64.8%, 61.5%, and 59.7%, and Hallucination Rates of 18.6%, 21.4%, and 24.1%, respectively. The results validate the proposed model's efficiency in intelligent power-system decision support through semantic consistency, factual accuracy, and context relevance.