Deep Neural Network For Steamflood Recovery Factor Prediction

Abstract

Proxy model to predict recovery factor is essential as a preliminary screening of steamflooding Enhanced Oil Recovery (EOR) design to find potential scenarios before it is simulated in reservoir simulator. This study uses Deep Neural Network to develop the steamflood proxy model. It consists of 2 hidden layers with 50 neurons each. The model development has considered bias-variance trade-off using dataset splitting (train-val-test split), and several diagnostic curves (Learning curve and Validation Curve). The dataset is generated using Latin Hypercube sampling towards steamflood screening criteria and several steamflood fields properties to represent worldwide steamflood project. The resulting prediction is presented in an Actual vs Prediction plot using root mean squared error and coefficient of determination as metrics. The proxy model has good performance, applicable for different reservoir flow properties, representative towards worldwide steamflood EOR project, and also a very fast prediction. The trained parameters (weights and biases) are also presented in Appendix.