arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2305.11202cs.HCcs.SEcs.SYeess.SY
收藏

从常规到新颖任务的电力工程LLM框架

LLM-based Frameworks for Power Engineering from Routine to Novel Tasks

Ran Li, Chuanqing Pu, Junyi Tao, Canbing Li, Feilong Fan, Yue Xiang, Sijie Chen

AI总结:

本文提出基于LLM的端到端和人在回路框架,分别用于电力系统的常规编程任务和复杂任务,发现人在回路特征对解决LLM领域知识不足的问题至关重要。

摘要:

能源领域的数字化扩大了电力工程师和研究人员的编码职责。本文探讨了利用大型语言模型(LLMs)减轻这一负担的潜力。我们提出了针对电力系统不同编程任务的基于LLM的框架。对于定义明确和常规的任务(如经典的机组组合(UC)问题),我们部署了端到端框架,以系统地评估四个领先的LLM——ChatGPT 3.5、ChatGPT 4.0、Claude和Google Bard——在成功率、一致性和鲁棒性方面的表现。对于先验知识有限的复杂任务,我们提出了人在回路框架,使工程师和LLM能够通过方法推荐、问题分解、子任务编程和综合的交互式学习来协作解决问题。通过两个框架的比较研究,我们发现网络访问、结合领域知识的问题分解和人工辅助代码综合等人在回路特征至关重要,因为LLM目前在获取前沿和特定领域知识以完成整体问题解决项目方面仍存在不足。

英文摘要:

The digitalization of energy sectors has expanded the coding responsibilities for power engineers and researchers. This research article explores the potential of leveraging Large Language Models (LLMs) to alleviate this burden. Here, we propose LLM-based frameworks for different programming tasks in power systems. For well-defined and routine tasks like the classic unit commitment (UC) problem, we deploy an end-to-end framework to systematically assesses four leading LLMs-ChatGPT 3.5, ChatGPT 4.0, Claude and Google Bard in terms of success rate, consistency, and robustness. For complex tasks with limited prior knowledge, we propose a human-in-the-loop framework to enable engineers and LLMs to collaboratively solve the problem through interactive-learning of method recommendation, problem de-composition, subtask programming and synthesis. Through a comparative study between two frameworks, we find that human-in-the-loop features like web access, problem decomposition with field knowledge and human-assisted code synthesis are essential as LLMs currently still fall short in acquiring cutting-edge and domain-specific knowledge to complete a holistic problem-solving project.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑