Flow Network Based Hybrid Models for Reservoir Applications

Understanding reservoir and predicting well performance in a timely manner is essential for closed loop reservoir management, improving operational efficiency, and maximizing value. Can we accurately predict water breakthrough, infer well connectivity, estimate resource volume, forecast well performance at field scale reliably through all the operational variations?

Traditional reservoir management methods are often interpretive and do not scale for manual surveillance of either large fields or those with large data volumes. Typically, the limitations may be due to (1.) unknown physics or lack of ways of formulating relevant physics; (2.) expensive or unavailable reservoir characterization data inputs at field scale and (3.) prohibitive speed for model building, calibration, and prediction to meet operational decision time cycles.

Data-driven modeling has been revitalizing the energy industry as a powerful approach for forecasting, well performance analysis and reservoir management. However, key challenges of data-driven applications in reservoir engineering include dealing with interpretability and data availability, where the amount of relevant data is limited or acquiring labeled data is unsustainable.

In this talk, we will discuss a new generation of reservoir modeling tools, referred to as reservoir graph network (RGNet). It combines physics and machine learning that can be built using routinely collected field measurements for practical reservoir model calibration, characterization, forecasting and optimization applications. These hybrid models combine the agility of the data-driven methods, while still providing adequate fidelity over long term or outside the training sample data support through underlying fundamental physics. The field case studies will illustrate the business value through several examples such as estimating reservoir resource size, well connectivity, uncertainty analysis, flood optimization and forecasting. All content contained within this webinar is copyrighted by Zhenyu Gho and its use and/or reproduction outside the portal requires express permission from Zhenyu Guo.

SPE Webinars are FREE to members courtesy of the