Open Access

ARTICLE

Enhancing Power Enterprise Inspection and Supervision: A LoRA-Based Lightweight LLM Framework Integrating Retrieval-Augmented Generation and Prompt Engineering

Jianfeng Liu¹, Yongjiao Yang¹, Kangyi Yang¹, Changhua Hu¹, Zijia Xu¹, Qingguo Shi², Yi Su^2,*

1 Guangdong Power Grid Co., Ltd., Zhongshan, China
2 Faculty of Automation and Electronic Information, Xiangtan University, Xiangtan, China

* Corresponding Author: Yi Su. Email:

(This article belongs to the Special Issue: Generative Artificial Intelligence and Large Language Models: Methods, Architectures, and Applications)

Computers, Materials & Continua 2026, 88(2), 95 https://doi.org/10.32604/cmc.2026.082804

Received 23 March 2026; Accepted 08 May 2026; Issue published 15 June 2026

Abstract

Power enterprise inspection and supervision require greater intelligence, efficiency, and standardization; however, existing approaches are limited by inefficient knowledge retrieval, inaccurate issue identification, and insufficient support for standardized reporting and rectification tracking. This study proposes a lightweight, domain-adaptive large language model (LLM) framework based on Low-Rank Adaptation (LoRA), integrating Retrieval-Augmented Generation (RAG) and structured prompt engineering to enable evidence-grounded inspection tasks. The framework achieves parameter-efficient adaptation through low-rank decomposition and constructs a domain-specific multimodal knowledge base, enhancing output traceability, consistency, and task generalization. A key contribution is the introduction of a Sensitive Information Control Gate, which enforces role-based access control and automated redaction, ensuring secure and compliant generation in regulated environments while preserving traceability. Experimental results demonstrate that the proposed method achieves improved performance over the base model and demonstrates competitive effectiveness under the evaluated conditions, supported by statistical analysis (paired t-test, p < 0.01, bootstrap 95% confidence intervals), while maintaining high parameter efficiency with only 0.4%–0.5% trainable parameters.

---

Keywords

---