Enhancing Large Language Models with Graph-Based Node Sampling for Fault Attribution in Power Distribution Networks

Abstract

The increasing complexity of power distribution networks necessitates advanced methods for fault attribution analysis to uphold system reliability and stability. This paper introduces a novel approach that integrates large language models (LLMs) with domain-specific knowledge graphs to tackle the challenges posed by high-dimensional and intricate fault data in power distribution networks. A multi-dimensional fault ontology [13] is proposed to efficiently structure the fault data, facilitating the creation of a comprehensive fault knowledge graph. To enhance the diagnostic predictions of the LLM, a reinforcement learning-based node selection algorithm is implemented, strategically choosing pertinent nodes from the fault graph to enhance the model’s reasoning abilities. Experimental findings illustrate that this approach surpasses traditional statistical methods and direct LLM reasoning, achieving superior accuracy and efficiency in fault diagnosis. By incorporating selective knowledge graph node sampling, irrelevant noise is filtered out from the fault data, sharpening the LLM’s focus and eradicating “AI hallucinations,” thereby improving analytical precision. Validation on a real-world dataset from a power company confirms the method’s efficacy, promising swift and accurate fault analysis while reducing the time required for power grid fault diagnosis.