Xinjiang Oilfield Puts Into Use an Industrial LLM for Pumping Well Production Optimization

A pumping well production optimization industrial large model, described as the first in the domestic oil and gas industry, has been applied at scale on more than 3,800 pumping wells at Xinjiang Oilfield since going online. Its average accuracy in diagnosing abnormal operating conditions exceeds 90%, the time to discover an abnormal well has dropped from the day level to the minute level, and wells managed per person rose from 3.5 to 11.5. The CNPC Exploration and Development Research Institute and Xinjiang Oilfield developed it independently. They combined production data from 78,000 pumping wells at the Daqing, Liaohe, Changqing and Xinjiang oilfields, raising data availability from 40% to over 95%, and used a Transformer architecture and contrastive learning to train a one-billion-parameter model, which was fine-tuned into scenario models for each oil production plant. Operating-condition diagnosis, the first scenario, covers all thin-oil pumping wells at the No. 1 and No. 2 plants and Baikouquan. Wells are color-coded by health, and clicking an abnormal well shows a report with the diagnosis, its basis and disposal suggestions. The team plans further iteration and embodied AI agent research.