Power Utility Ticket Processing via Domain Knowledge-Infused Large Language Models

Abstract

Power grids face growing challenges from extreme weather, aging infrastructure, and increasing penetration of electric vehicles and distributed energy resources, making incident response more complex. Maintenance tickets provide valuable records, but their unstructured text requires manual interpretation, which is labor-intensive and error-prone. To address this, this paper proposes a domain knowledge-infused pre-trained transformer for automated multi-label classification of utility tickets to predict outage causes. An in-house DistilBERT-based model is developed as an efficient encoder-only transformer, offering a lightweight and faster alternative while maintaining strong performance. The model is pre-trained on domain-specific power systems text to capture technical vocabulary and operational semantics, then fine-tuned on labeled tickets. It is first evaluated on the IEEE PQ dataset, where the domain-infused version outperforms the base model on short, terminology-dense descriptions. The approach is further validated on a large real-world dataset from a major U.S. utility, achieving performance comparable to DI-BERT (within 0.5% exact match) while running nearly $2\times $ faster, outperforming traditional machine learning methods and large LLM baselines, and enabling fast, privacy-preserving on-premises deployment.