Domain-Specific Large Language Model for Maintenance Decision-Making on Wind Farms by Labeled-Data-Supervised Fine-Tuning
Dongming Fan, Meng Liu, Yi Shao, Linchao Yang, Yiliu Liu, Yue Zhang, Yi Ren, Zili Wang
Abstract
Wind farm operators always need a better maintenance strategy to increase resource utilization efficiency while controlling operation and maintenance costs. However, conventional maintenance decision-making approaches are time-consuming and have poor flexibility and adaptability to various scenarios. This study addressed these challenges by using a large language model (LLM) to understand, generate, and plan maintenance strategies for wind farms characterized by various failure modes and maintenance costs. A labelled-data-supervised fine-tuning LLM for maintenance, named LLM4M, is proposed. The proposed LLM4M model is trained on an extensive dataset of mathematical programs for maintenance to generate optimal strategies for wind farms. Compared with other large parameter LLMs, the fine-tuned LLM4M model demonstrates remarkable accuracy, with an error of approximately 2% from the optimal strategy. In addition, the generalization of the proposed LLM4M model has achieved remarkable results. If the LLM4M model correctly generates the maintenance strategy, the maintenance cost deviates from the optimal solution by only approximately 5%. Furthermore, phase transition behavior is observed, which provides considerable guidance for the development of domain-specific LLMs for the maintenance domain.