DYNAMIC DATA-DRIVEN SMART PROXY MODELING FOR NUMERICAL RESERVOIR SIMULATION

Abstract

A successful Geologic Carbon Dioxide (CO2) Storage (GCS) operation requires the ability to make quick and reliable subsurface modeling decisions; such decisions must be made based on an accurate and realistic modeling of the reservoir. Numerical reservoir simulation is the most common tool used for predicting fluid flow behavior and analyzing uncertainties in the subsurface reservoirs. In general, a numerical reservoir simulation model has tens of millions of grid blocks and requires intensive computations to be performed at each time-step of the simulation, therefore, they are computationally expensive and time-consuming. As a result, studies (such as uncertainty analysis of GCS) which may require hundreds to thousands of simulation runs become impractical due to the very large amount of time needed to make the runs. This research study employs Artificial Neural Networks (ANN) and data-driven techniques to lower the high computational footprint necessary to produce reservoir simulation modeling results by developing Dynamic Smart Proxy models. The Dynamic Smart Proxy is a data-driven machine learning model that successfully replicates the output of a sophisticated numerical reservoir simulation model for each time step in a short amount of time (fraction of minutes). To develop the Smart Proxy Model, algorithms in ANN must be trained on large volumes of sub-surface data to learn the complex patterns of the fluid in the Numerical reservoir simulation. Therefore, a few reservoir simulation realizations were developed with varying geological properties such as porosity and permeability. Each reservoir simulation included a certain number of Injectors. The simulation results, including the geological reservoir properties were utilized to develop the Smart Proxy model. The developed Smart Proxy Model reproduced the output (such as pressure distribution and CO2 saturation) of an entirely new reservoir simulation run at each grid block of the model during every time step of the simulation.