Why are we not seeing more data-driven VFMs in operation?

In time with improved instrumentation of petroleum assets and increasing amounts of data, the buzz around AI and Machine Learning reached the petroleum industry. Is there a reason why such solutions shouldn't kick-ass in production flow modeling?

Several mind-blowing achievements of data-driven solutions have been reported from other fields of research, such as AlphaGo, the computer program that beat the world champion in the game of Go. This has led to a prevalent “AI mindset” also in the oil and gas industry, with high but yet unfulfilled expectations of impact.

In recent years, there has been an increasing amount of literature on data-driven solutions to virtual flow meters (VFMs) - a soft-sensor technology able to predict the petroleum flow rates in an asset using mathematical models and already existing measurements. Some of the case-studies on real data report an error as low as 4% on predicting the oil flow rate from a petroleum well (AL-Qutami et. al. 2017). Yet, as far as we are aware, no commercial data-driven VFMs exist. Why is that?

Less than 10% error 90% of the time

Solution Seeker has several years of experience in working with data from petroleum assets and in modeling data-driven VFMs. We believe that the traditional approaches to data-driven VFM modeling may not be robust enough for commercial use. We define a method to be sufficiently robust if it achieves less than 10% error for at least 90% of the wells modeled. To illustrate this, consider the two figures on the right hand side. The figures illustrate the oil flow rate predictions from two VFMs developed for two individual petroleum production wells using an identical modeling approach. In Figure 1, the VFM achieves an error of 4%, whereas in Figure 2, the error is 45%.




Four challenges

We believe that there are four prevalent challenges that degrade the performance of data-driven VFMs:

  1. Low data volume
  2. Low data variety
  3. The available data have poor quality, they are noisy and biased
  4. The available data originate from a non-stationary process

In an upcoming series of articles, we will share and discuss recent findings of Solution Seeker’s research lab. You can find our article in the reference list below. The series begins by questioning the robustness and trustworthiness of the traditional approaches to data-driven VFMs in light of the four challenges listed above. The findings are based on experiments with a large and diverse set of data from 60 petroleum wells located on five petroleum assets. Stay tuned!

References

  • Tareq Aziz AL-Qutami, Rosdiazli Ibrahim, Idris Ismail, and Mohd Azmin Ishak. Development of soft sensor to estimate multiphase flow rates using neural networks and early stopping. In International Journal on Smart Sensing and Intelligent Systems, volume 10, pages 199–222, 2017
  • Grimstad, B., Hotvedt, M., Sandnes, A.T., Kolbjørnsen, O., Imsland, L.S.,2021. Bayesian neural networks for virtual flow metering: An empirical study.arXiv:2102.01391