In modern power systems, ring main units (RMUs) are crucial components of distribution networks, responsible for the key tasks of power distribution and grid fault protection. However, due to insulation aging and harsh environmental conditions inside RMUs, partial discharge (PD) issues frequently arise, threatening the normal operation of the equipment. Traditional PD identification methods often rely on ultra-high frequency (UHF) signal detection and conventional machine learning algorithms, which, despite offering certain detection speed, have limited generalization capabilities and struggle to handle diverse PD patterns. In recent years, deep learning technology has gradually been applied to PD identification with significant success. However, traditional deep learning methods still fall short in terms of intelligent operation and command processing, limiting their application scenarios. This paper proposes a PD identification method based on a multimodal large model, which achieves more efficient PD identification by integrating UHF signal time-frequency features with other operational data. The proposed method designs a multimodal large model architecture capable of synergistically processing multimodal data and introduces an efficient pre-training and supervised fine-tuning algorithm. Experimental results demonstrate that the proposed model surpasses traditional deep learning algorithms in terms of accuracy and operational capability, showcasing its potential for application in complex power scenarios.