AH-SpanBERT: Fine-Tuning SpanBERT with Archerfish Optimization for Power Grid Dispatcher Training

Abstract

Power grid dispatchers play a critical role in maintaining the stability and efficiency of electrical networks. As power systems grow in complexity, traditional training methods struggle to equip dispatchers with the necessary skills for rapid decision-making and human-machine collaboration. This research explores the application of fine-tuning general large language models (LLMs) to enhance internal training processes for power grid dispatchers. The research proposed Archerfish Hunting Fine-tuned Span Bidirectional Encoder Representations from Transformers (AHSpanBERT), a model that integrates the SpanBERT architecture with Archerfish Hunting (AH) optimization to improve decision-making and operational efficiency in power system management. To fine-tune the model, a comprehensive dataset of 1,000 simulated power grid operational records was created, covering scenarios such as equipment failures, grid fluctuations, and emergency responses. The data was preprocessed using domain-specific tokenization and term normalization to ensure consistency and contextual relevance. The AH-SpanBERT model was trained using this dataset, with specific prompt strategies designed to simulate real-world dispatch scenarios and foster interactive, scenario-based learning. The model’s performance was evaluated across multiple key metrics, including factuality, logicality, stability, and security. Results show significant improvements in factuality (8.48 in operation monitoring), logicality (9.74 in general scenarios), stability (9.15 in black start procedures), and security (9.62 in black start procedures). The AH-SpanBERT model outperforms existing LLMs such as GPT-4 and GAIA-70B in these areas, demonstrating its potential to enhance dispatcher decision-making and human-machine collaboration in critical power grid operations. This research highlights the effectiveness of fine-tuning general LLMs with domainspecific data to improve dispatcher training and operational performance in power grid management.