Construction and Application Technology of Controllable and Credible Power System Large Language Model

Addressing challenges such as incomplete knowledge in the field of electricity in areas like distribution network maintenance auxiliary decision-making, marketing customer service, and digital safety supervision and training, which leads to unreliable auxiliary decision-making conclusions generated by large models, potential inclusion of sensitive or non-compliant information in the content generated by large models, difficulty in tracing safety Q&A based on large models back to specific safety regulations, and lack of interpretability, we conduct research on the construction and application technology of a large-scale, controllable, and credible text model for power generation. Specifically, we improve the credibility and accuracy of content generation in the process of distribution network maintenance auxiliary decision-making by conducting research on pre-training technology for controllable and credible large language models based on mixture-of-experts and adaptive fine-tuning. We enhance the safety and compliance of intelligent Q&A content generation in marketing customer service through research on fine-tuning technology for enhancing the safety robustness and ethical compliance of large language models with knowledge-guided instructions. Additionally, we improve the credibility, traceability, and consistency of digital safety Q&A content through research on interpretability and consistency detection and verification technology for content generation by large language models.