Data-Driven Reservoir Modeling: Application of Proxy Models in Reservoir Management

Abstract

This Ph.D. thesis consists of 8 papers that summarize the main contents of the research work done over the past 3 years. Due to the ability of machine learning (ML) in capturing high nonlinearity, the thesis mainly touches upon its use in data-driven modeling to provide aids in reservoir management. Data-driven models are referred to as “proxy models” as they act on behalf of the reservoir simulator. Proxy models are deemed practically useful if they can provide fast and desirably accurate solutions. In this thesis, a survey on the use of ML and metaheuristic algorithms in developing proxy models for reservoir simulation was presented to enlighten the readers. We also explained the methodology of proxy modeling with an associated case study, viz. the waterflooding process. The proxy modeling of a synthetic reservoir model was first formulated on which further works were done as improvements. These improvements, including the integration of sampling techniques and the use of more complex reservoir models, proposed the fundamentals of the proxy modeling methodology in more realistic application cases. Upon the completion of these steps, adaptive sampling and retraining were applied to address the geological uncertainties. Also, two classes of proxy modeling, namely local and global proxy modeling, were implemented to handle optimization problems with higher dimensions. Furthermore, additional works were illustrated to provide a scaffold for the maturity of the methodology. These works pertain to research on applying ML methods in predictive modeling and a decision analysis framework. One of them illustrated the establishment of ML-based predictive models with splendid predictability. The work also includes a discussion about the steps of predictive modeling for well production forecast based on real field data. The other one displayed coupling of ML with a mathematical algorithm to approximate the Value of Information that was used for optimization under uncertainties. These studies are not only related to those described earlier but also illustrate the robust application of machine learning. In summary, this research project portrayed the establishment of a methodology that could yield proxy models to facilitate the resolution of reservoir management issues with less computational efforts as compared with reservoir simulator without compromising the accuracy.