Monitoring sensor prediction is a critical component of power transformer operation and maintenance. Traditional time-series-based methods for predicting variables such as oil temperature suffer from two major limitations: (a) they require domain-specific expertise, as engineers must understand the algorithm's underlying principles and parameter configurations; (b) they cannot leverage textual knowledge documents. To address the first issue, this study fine-tunes a large language model (LLM) for sensor prediction, leveraging their conversational interface, which allows engineers to obtain prediction results simply by posing queries. This fine-tuning is achieved through Low-Rank Adaptation (LoRA), a method that accelerates training and efficiently updates parameters. To address the second issue, we employ retrieval-augmented generation (RAG) techniques to enable the integration of local knowledge bases. This transforms the LLM into a domain-specific expert capable of reading and utilizing maintenance manuals, technical patents, and operations and maintenance (O&M) logs to provide informed expert operational suggestions. Experiments and visualizations validate the effectiveness of the proposed approach, highlighting the potential of LLMs in the field of power system maintenance.

Large Language Model (LLM),power transformer,Monitoring Sensor Prediction,oil temperature prediction