On a hybrid approach to model learning applied to virtual flow metering

Abstract

Process modeling using first-principle equations has existed for centuries as a methodology to represent and analyze real-world processes. In time with increasing computing power and sensor data availability, data-driven modeling for processes has gained attention. Even though data-driven modeling, or machine learning, has shown remarkable results in fields such as image classification and speech recognition, it has yet to be adopted as the preferred approach for process modeling. Arguably, this is due to the long history of first-principles modeling, along with the inherent black-box nature of datadriven models. The latter causes a lack of model explainability, which, in turn, can result in distrusting the predictions originating from data-driven models. Furthermore, disregarding physical laws that have been acknowledged for centuries to model processes can seem irrational. Hybrid, or gray-box, modeling is a methodology with a vision to utilize all available knowledge, both physics and data, to model processes. It combines firstprinciple equations with data-driven techniques and is especially intriguing for inherent complex processes where the physical behavior is partly unknown or challenging to model with first principles. One such process is the petroleum production system. The multiphase flow rate through the production system is challenging to model with required precision using first principles due to uncertain subsurface properties and complex dynamic behavior. Furthermore, available sensor data is often limited or of low quality. Therefore, a hybrid modeling approach seems of significant importance to models predicting the multiphase flow rates as it attempts to exploit all available information to its full extent. The work leading to this thesis has explored hybrid solutions for virtual flow metering. A virtual flow meter (VFM) is a soft-sensor that utilizes process models and already existing sensor measurements, such as pressures and temperatures, to compute the multiphase flow rate at strategic locations in a petroleum asset. The main part of this thesis is a collection of six peer-reviewed papers, three journal publications, and three conference publications. In addition to the paper collection, this thesis introduces the topic of hybrid modeling for virtual flow metering to provide context to the publications. The main contributions of the six publications can be summarized as follows: firstly, a framework for simultaneous estimation of all parameters in a model with varying degrees of hybridity has been proposed. Secondly, six hybrid VFM model types were developed from real and historical production data from a petroleum asset. Thirdly, several hybrid model properties such as explainability, scientific consistency, flexibility, and accuracy have been examined. Lastly, two methods, one to include uncertainty in the modeling, and one to address the inherent nonstationarity of the underlying process to sustain the long-term VFM performance, have been proposed. The key takeaway of the work leading to this thesis is that hybrid modeling is challenging, yet, also essential for obtaining high accuracy VFMs in certain scenarios. The contributions have shown that the task of balancing learning from physics and learning from data is nontrivial, and if incautious, the hybrid model can exploit the disadvantages of both the mechanistic and data-driven modeling domain instead of the advantages. On the other hand, the results also showed that for processes with unknown or unmodeled physics, a hybrid model can offer improved performance over a mechanistic model, and with little available process data, a hybrid model can obtain a higher performance than a data-driven model. Moreover, in the presence of nonstationarity and little data, frequent updating of a hybrid VFM has shown essential to sustain the prediction accuracy over time. From the results, it is believed that hybrid modeling can be generalized to other applications and can offer improved performance over a mechanistic and datadriven approach. Furthermore, the solution for hybrid modeling presented in this thesis can be conveniently integrated with existing mechanistic process models in the industry. Naturally, the domain of hybrid modeling for virtual flow metering has not been fully explored. The most promising future research direction is combining hybrid modeling with methods that enable learning from more than one petroleum well at a time.