(193g) Physics-Aware Machine Learning Algorithms with Improved Accuracy and Explainability Applied to Multiphase Flowrate Estimation | AIChE

(193g) Physics-Aware Machine Learning Algorithms with Improved Accuracy and Explainability Applied to Multiphase Flowrate Estimation

Authors 

Bikmukhametov, T. - Presenter, Norwegian University of Science and Technology
Jäschke, J., Norwegian University of Science and Technology
Estimation of oil, gas and water flowrates from wells in petroleum production systems is one of the challenges to be solved in order to perform efficient production optimization, accurate update of a reservoir model and deliver safe operation in flow assurance aspects [1]. One solution to this problem is to use hardware metering devices which produce good flowrate estimates but are very expensive and exposed to degradation. Another approach is to use a mathematical model of the production system and solve an inverse problem to estimate the flowrates from readily available pressure and temperature measurements. This method is called Virtual Flow Metering (VFM). The mathematical problem can be formulated using first principles models which can include mass, momentum and energy balances of the multiphase flow mixture. The behavior of these models can be well explained, and it can operate under limited amount of historical data [2]. However, due to the complex multiphase flow nature, these models are hard to create, and computational cost can be high due to the need of solving discretized partial differential equations and a non-linear optimization problem. At the same time, machine learning algorithms can be used in order to estimate the flowrates using historical data, however, they typically do not consider the specifics of the process and estimate the flowrates from data directly. These algorithms can be easier to construct and use than the first principles models, however, because of its black-box behavior, it is hard to interpret the algorithm outcomes, and petroleum engineers commonly prefer to use conventional approaches.

In this work, we propose a method which combines first principles and machine learning VFM approaches. More specifically, we create machine learning algorithms which are aware of the multiphase flow physics through input features and represent a specific system part, for instance, a well tubing or a production choke. To do this, we use relatively simple first principle models such as Bernoulli choke model with multiphase mixture properties and No-Pressure-Wave momentum equation [3] as the well tubing model and use capabilities of machine learning algorithms to utilize these models in describing the multiphase flow in wells. This is different from the traditional machine learning VFM systems in which raw measurements are used directly [4]. We study different algorithms such as tree-based algorithms (gradient boosting regression trees [5] and random forest [6]) and neural networks (feed-forward and Long-Short Term Memory [7]). In addition to using the algorithms directly, we also combine them using a linear meta-model. All the approaches are tested on real field data from a subsea well located in the North Sea.

The results show that by using physics-aware machine learning algorithms, it is possible not only to improve the predictive accuracy, but also explainability of the machine learning algorithms. The first method to expand the algorithm explainability is to evaluate the importance of the physics-aware features. Another method is to use the coefficients from the linear meta-model which shows the importance of the used algorithms. Since each algorithm is responsible for a specific system part, for instance, a well tubing and a production choke, the meta-model explains how a particular algorithm contributes to the final solution. As such, by using the approach of combining physics-aware machine learning algorithms, two goals can be achieved: 1) improved accuracy of the model using physically meaningful features and simple meta-models; 2) improved explainability of the model which is important in real time operation.

References

[1] Falcone, G., Hewitt, G. F., & Alimonti, C. (2009). Multiphase Flow Metering: Principles and Applications. Amsterdam: Elsevier.

[2] Holmås, K., & Løvli, A. (2011). FlowManager Dynamic: A multiphase flow simulator for online surveillance, optimization and prediction of subsea oil and gas production. 15th International Conference on Multiphase Production. Cannes, France.

[3] Aarsnes, Ulf Jakob F., Adrian Ambrus, Florent Di Meglio, Ali Karimi Vajargah, Ole Morten Aamo, and Eric van Oort. "A simplified two-phase flow model using a quasi-equilibrium momentum balance." International Journal of Multiphase Flow 83 (2016): 77-85.

[4] Al-Qutami, T., Ibrahim, R., Ismail, I., & Ishak, M. (2017) (Al-Qutami et al. 2017a) Development of Soft Sensor To Estimate Multiphase Flow Rates Using Neural Networks And Early Stopping. International Jouranl on Smart Sensing and Intelligent Systems, 10(1), 199-222. doi:10.21307/ijssis-2017-209.

[5] Chen, Tianqi, and Carlos Guestrin. "Xgboost: A scalable tree boosting system." In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pp. 785-794. ACM, 2016.

[6] Breiman, Leo. "Random forests." Machine learning 45, no. 1 (2001): 5-32.

[7] Hochreiter, Sepp, and Jürgen Schmidhuber. "Long short-term memory." Neural computation 9, no. 8 (1997): 1735-1780.