Enhancing domain-specific text generation for power grid maintenance with P2FT.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Power Systems Engineering

Background:

• Digitizing power grid operations requires extensive data for smart decision-making.
• Intelligent decision generation in specialized fields like power grids demands domain-specific knowledge and robust language processing.
• Pre-trained language models (PLMs) offer potential for specialized tasks, even with authorization constraints.

Purpose of the Study:

• To explore pre-trained model specialization for generating maintenance strategies in the complex power grid domain.
• To address the challenge of specialized terminology and knowledge within power grid textual data.
• To evaluate a novel two-stage fine-tuning approach (P2FT) for this specific task.

Main Methods:

• Utilized a large-scale pre-training model for natural language processing.
• Implemented a two-stage fine-tuning approach (P2FT) tailored for power grid domain specialization.
• Evaluated the method's efficacy and practical value against advanced low-parameter and parameter-free fine-tuning techniques.

Main Results:

• The P2FT approach demonstrated significant feasibility and practical application value for pre-trained model specialization.
• Experimental outcomes validated the effectiveness of the proposed method in generating power grid maintenance strategies.
• The study's findings were corroborated through meticulous analysis and validation against other advanced methods.

Conclusions:

• Pre-trained model specialization using the P2FT method is a viable and valuable approach for the power grid domain.
• This research provides practical guidance for text generation tasks in specialized Chinese language and power grid contexts.
• The findings contribute to advancing intelligent decision-making in power grid operations and maintenance.