中国石油网消息：“以前是人追着井跑，凭经验猜；现在是数据追着人跑，AI帮你把关。大模型让异常处置更精准了。”新疆油田采油二厂异常井技术人员感慨道。4月12日记者获悉，国内油气行业首个抽油机井生产优化工业大模型自上线以来，已在新疆油田超3800口抽油机井规模化应用，平均工况异常诊断准确率超90%，异常井发现周期从天级缩短至分钟级。

China Oil News: "In the past, people chased after the wells, relying on experience and guesswork. Now, data chase after people, and AI helps to verify. The large model makes abnormal handling more accurate." An abnormal well technician from the Second Oil Production Plant of Xinjiang Oilfield expressed his thoughts. On April 12th, reporters learned that the first industrial large model for optimizing the production of pumping wells in the domestic oil and gas industry has been in use in Xinjiang Oilfield for over 3,800 pumping wells. The average accuracy rate of abnormal condition diagnosis has exceeded 90%, and the discovery cycle of abnormal wells has been shortened from daily level to minute level.

该模型由中国石油勘探开发研究院联合新疆油田自主创新研发，聚焦油气生产痛点，深度整合来自大庆、辽河、长庆、新疆四大油田7.8万口抽油机井的生产数据，构建高质量数据集，将数据可用率从40%提升至95%以上。基于此，研发团队采用Transformer架构及对比学习算法，训练了10亿参数抽油机井生产优化大模型，并结合采油厂需求微调形成场景模型。

This model was jointly developed by the China National Petroleum Exploration and Development Research Institute and Xinjiang Oilfield through independent innovation. It focuses on the pain points in oil and gas production and deeply integrates the production data from 78,000 pumping wells in four major oilfields - Daqing, Liaohe, Changqing and Xinjiang. It constructs a high-quality data set, increasing the data availability rate from 40% to over 95%. Based on this, the research team adopted the Transformer architecture and contrastive learning algorithm to train a large model for optimizing pumping well production with 1 billion parameters. They then fine-tuned it according to the needs of the oil production plants to form a scenario model.

工况智能诊断作为首个落地场景，已在新疆油田采油一厂、二厂、百口泉等采油厂实现稀油抽油井的全覆盖运行。系统以不同颜色标注油井健康状况，告警、预警、问题井一目了然。技术人员点击异常井编号，即可查看抽油机井运行状态的“全身体检报告”及对应的诊断结论、判断依据和处置建议，彻底改变了凭经验判断的传统模式，人均管井数从3.5口提升至11.5口，有效提升了生产效率，降低了生产成本。

As the first implemented scenario, the intelligent diagnosis of working conditions has achieved full coverage operation of sparse oil pumping wells in the first and second oil production plants and Baikouquan oil production plant of Xinjiang Oilfield. The system marks the health status of oil wells with different colors, making alarms, warnings, and problem wells clearly visible. By clicking on the abnormal well number, technicians can view the "comprehensive physical examination report" of the pumping well's operation status, as well as the corresponding diagnosis conclusion, judgment basis, and disposal suggestions. This has completely changed the traditional mode of judgment based on experience, increasing the average number of wells managed per person from 3.5 to 11.5, effectively improving production efficiency and reducing production costs.

据悉，智能产液计量和生产参数智能优化作为新拓展场景，可实现单井产液量在线连续计算，同时自动筛选潜力井、推送参数调整方案，推动油田生产从“人工调参”向“智能寻优”跨越。下一步，研发团队将继续以大模型落地见效为目标，持续迭代优化，提升模型能力，并开展具身智能体研究。

It is known that the new expansion scenarios of intelligent liquid production measurement and intelligent optimization of production parameters can achieve online continuous calculation of single well liquid production volume, automatically screen potential wells and push parameter adjustment plans, promoting the transition of oilfield production from "manual parameter adjustment" to "intelligent optimization". In the next step, the R&D team will continue to aim at the goal of achieving effective implementation of large-scale models, continuously iterate and optimize, enhance the model capabilities, and conduct research on embodied intelligent agents.