Exploration of the Application of General Large Model Fine-tuning Technology of Natural Language in the Internal Training of Power Grid Dispatchers

Abstract Power grid dispatchers play a critical role in ensuring the stable, efficient, and safe operation of electrical networks. As power systems continue to expand in scale and complexity, traditional training methods increasingly struggle to equip dispatchers with the necessary skills for multitasking, rapid decision-making, and effective human-machine collaboration. This research investigates the application of fine-tuning general large language models (LLMs) using domain-specific data to enhance internal training processes for power grid dispatchers. The Archerfish Hunting Fine-tuned Span Bidirectional Encoder Representations from Transformers (AH-SpanBERT) model is designed to support a wide range of power system operational tasks and decision-making scenarios, including general, dispatch, operation monitoring, and black start procedures. A comprehensive dataset was compiled with simulated operational records that cover real-world scenarios such as equipment failures, grid fluctuations, emergency actions, and routine monitoring activities. The data was preprocessed using tokenization and domain-specific term normalization to ensure consistency and contextual relevance. The research fine-tuned the general LLM to acquire specialized knowledge and domain-specific contextual understanding for training. Prompt strategies were developed to simulate realistic dispatch scenarios, fostering interactive, scenario-based learning for trainees. The model’s power dispatch performance was evaluated with scenarios to assess LLMs on key parameters such as factuality, logicality, stability, and security for effective power system management in a black start. Experimental results involving dispatchers of varying experience levels revealed significant improvements in factuality (8.48 in operation monitoring), logicality, stability, and security. This research demonstrates that the proposed General Large Model significantly enhances decision-making capabilities, operational efficiency, and human-machine interaction within power dispatch operations.