A qualitative study of the impact of random shale barriers on SAGD performance using data analytics and machine learning

Shale barriers are common in all types of oil and gas reservoirs. Previous works on characterizing the impact of random shales are limited. Although for a gravity-based recovery method such as steam-assisted gravity drainage (SAGD), these shale barriers prove to be a major hurdle in production rate and ultimate oil recovery. Apart from obstructing the oil drainage path, shale barriers may also contribute to the additional heat losses. Hence characterizing these shale barriers is of importance to understand the magnitude of their impact on oil recovery. In this paper, first, we describe a procedure to construct the base 2D reservoir model and then populate it lognormally with various scenarios of random shales based on the mean and standard deviation. Unique features are extracted to parameterize these random shales. Next, the impact of reservoir heterogeneity on the recovery factor is described qualitatively using these shale features. Second, we prepare a non-linear model using a machine learning algorithm for predicting oil recovery factor using a synthetic dataset. We demonstrate that the machine learning approach can predict the oil recovery factor with reasonable accuracy compared with thermal numerical reservoir simulation results. The model resulted in an R2 value of 0.95 and RMSE of 5.06. Considering the complexity of characterization of randomly distributed shales, this is a considerable accuracy from a model trained on 226 simulation data sets. The approach described in this paper can be used as a technique to characterize the impact of randomly distributed shales. The predictive model may also be utilized to run a large set of sensitivity, which can prove effective for making a real-time decision. These techniques can be further advanced with a larger set of training data, including a variety of dynamic and static reservoir properties.

Automated summary

Random shale barriers in oil and gas reservoirs can significantly impact production rates and ultimate recovery. This paper presents a procedure to construct a 2D reservoir model with various scenarios of random shales and parameterizes their unique features. A machine learning model is used to predict oil recovery factor with high accuracy compared to traditional thermal numerical reservoir simulation results, suggesting its potential for real-time decision-making in oil recovery operations.