In the electric power industry, entity relationship extraction is essential to achieving deep semantic understanding and supporting organized knowledge discovery. However, when applied to electric power texts, this task faces several difficulties, including domain-specific terminology, complex relational structures, and contexts that span multiple sentences. This paper presents a relation extraction framework based on large-scale model multi-task fine-tuning, which consists of three main parts: an output selector based on prompt engineering techniques, progressive adapter activation with dynamic weight adjustment strategies, and a Multi-task Embedded Low-Rank Adaptation (LoRA) training mechanism. The study uses phased training to reduce the conflicts that arise from multi-task learning and applies output format constraints that are specific to the business needs of the electric power industry for methodological validation. Experimental results show that the multi-task LoRA fine-tuning approach and the proposed dynamic weight adjustment strategy significantly enhance the model’s ability to identify intricate entity relationships in the context of electric power, particularly by allowing for structured control of large language model outputs. Furthermore, especially in situations with constrained data resources, the framework provides a robust and dependable technological technique and system solution for entity relation extraction from electric power texts.