胜利油田油气认知大模型建设与应用  

Large Language Model of Oil and Gas Cognition Constructed and Applied in Shengli Oilfield

作　　者：段鸿杰[1] 马承杰[1] 王振 公绪超 景瑞林[1] 刘鹤 Duan Hongjie;Ma Chengjie;Wang Zhen;Gong Xuchao;Jing Ruilin;Liu He(Digital Intelligence Management Service Center,Sinopec Shengli Oilfield Company,Dongying 257015,China)

机构地区：[1]中国石化胜利油田分公司数智化管理服务中心

出　　处：《石油科技论坛》2024年第6期46-55,共10页PETROLEUM SCIENCE AND TECHNOLOGY FORUM

基　　金：中国石油化工股份有限公司科技项目“基于大数据的油藏井筒一体化智能诊断方法研究”(编号:P20058-1)。

摘　　要：为推动企业数字化转型智能化发展,胜利油田大力推进大模型技术在勘探开发领域的研发和应用,2023年成功上线推广了“胜小利”智能助理,数万人安全使用。文章介绍了人工智能与大模型概念,油气行业大模型发展现状。针对油气认知大模型建设过程中存在的海量多源异构数据整合处理、计算资源和训练效率要求高等难题,“胜小利”大模型在专业知识构建、体系技术架构、训练推理速度优化中采用先进适用技术,学习了数百万油气领域专业语料知识,打造基于千亿级自然语言大模型的油田信息资产智能服务平台,实现数据应用的高效交互、专业知识的共享传承。通过定制化知识运营、构建专长大模型助手,支撑了十几个大模型虚拟专家应用,为用户节省了大量写作调研、查数据查系统的时间,工作效率显著提升,为实现“一企一模型、一人一助理”的目标进行了有益探索。To promote the transition of corporate digitalization to intelligence,Shengli Oilfield was committed to accelerate R&D and application of large language model technology in the area of exploration and development.In 2023,it successfully launched the“Shengxiaoli”intelligent assistant online to serve thousands of thousands of users safely.This article briefs about the AI and large language model conception as well as the current development conditions of large language model in the oil and gas sector.Focusing on some bottlenecks in the construction process of the large language model,such as integration and processing of a huge magnitude of multi-source heterogeneous data and a high demand for calculation resources and training efficiency,the“Shengxiaoli”large language model adopts the latest adaptation technology in construction of specialized knowledge,configuration of system technology and optimization of training and reasoning speed.It also learns millions of millions of specialized language materials and knowledge in the oil and gas area,creates the oilfield information assets intelligent service platform based on a large model with billions of natural language resources,and brings about application of high-efficiency exchanged data and sharing of specialized knowledge.Based on operation of the tailored knowledge and construction of the specialized large language model assistant,a dozen of large language models for virtual expert application are established to save users a large amount of time for writing,investigation,consultation of the data system,obviously improving work efficiency.This is a meaningful effort for realization of the“one model for one enterprise and one assistant for one person”goals.

关 键 词：油气大模型 油气生产 勘探开发 数字化 智能化 信息交互 

分 类 号：TP39[自动化与计算机技术—计算机应用技术] TE1[自动化与计算机技术—计算机科学与技术]