Design and development of data-driven screening tools for enhanced oil recovery processes

Computationally efficient screening and forecasting tools can offer faster decision-making and value-creation opportunities for enhanced oil recovery (EOR) operations without requiring a high-fidelity reservoir model. In this paper, we present a data-driven modeling approach utilizing numerical models and neural networks (ANN) to screen EOR methods in a rapid way. Numerical modeling is employed to generate the data for the training of the neural-network based data-driven model. It is aimed to develop comprehensive and globally applicable screening tools that can be used to identify reservoirs where the EOR method would be applicable through estimation of performance indicators, such as utilization efficiency. Important process variables related to the EOR method are identified and grouped, while also defining their limits and statistical distributions. A large number of scenarios are generated and run using the numerical model that represents the EOR method under consideration. A simple yet representative performance indicator is defined and calculated for each scenario, which takes into account both additional income due to incremental oil recovery and the cost of the injected agent. Considering annual volumes in the calculation and evaluating the time-dependency of the performance indicator allow to incorporate time-value of the money. Finally, the knowledge base is used to train a neural network that can capture the signatures within the dataset through an iterative training process. This methodology is explained by highlighting important components of the workflow as well as best practices of each step. Results of two example applications are summarized: 1) Cyclic steam injection, 2) Polymer flooding. The results indicated that presented methodology can be successfully followed for different EOR methods. In both cases, the data-driven screening model was able to predict the efficiency indicator within acceptable accuracy levels and identified the degree of success of each method under different reservoir and operational conditions. Comparison of sensitivities between numerical and data-driven models showed that the data-driven model captured the physics of both problems as reflected by the numerical model. Predicting the indicator at 2-year intervals allowed to estimate the feasibility as a function of time.