A New Method for Predicting CO2 Minimum Miscibility Pressure MMP Based on Deep Learning

Abstract

AbstractThe CO2 enhanced oil recovery (EOR) method is widely used in actual oilfields. It is extremely important to accurately predict the CO2 MMP for CO2-EOR. At present, many studies about MMP prediction are based on empirical, experimental, or numerical simulation methods, but these methods are inefficient in calculation, with low result accuracy and high computation burden. Therefore, more work needs to be done. In this work, a fully connected neural network (FCNN) is developed to facilitate predict the CO2 MMP, based on multiple mixing cell methods (MMCM) and slim-tube experiment. With the results of the slim-tube experiment and the data expansion of the MMCM method, a FCNN model that predicts CO2 MMP by the full composition of the crude oil and temperature is trained. To stabilize the neural network training process, L2 regularization and Dropout are used to address the issue of over-fitting in neural networks. Training results show that FCNN with Dropout possesses higher prediction accuracy. Then, based on the validation sample evaluation, the mean absolute percentage error (MAPE) of the FCNN model is 6.99. Finally, the improved FCNN model is evaluated by six samples obtained from slim-tube experiment results. The FCNN model proposed in this paper has extremely low time cost and high accuracy to predict CO2 MMP, which is of great significance for CO2-EOR.