Improving the Simulation of Waterflood Performance With the Use of Neural Networks

Abstract

Abstract Secondary recovery performance of most oil reservoirs is impacted by formation heterogeneity. Therefore, accurate description of the formation attributes is necessary for appraising the economic success of the secondary recovery operations in a complex reservoir. Porosity and permeability distributions are the key formation attributes, which are required for realistic simulation of the secondary recovery performance. The permeability and porosity are usually determined from the analysis of the core samples. Generally, core samples are only available from limited number of wells in the reservoir because of the expenses associated with obtaining and analyzing the core samples. At the same time, geophysical logs are commonly run in most of the wells in the reservoir. Well log data are then utilized to identify the productive intervals and to estimate the formation porosity. A correlation between core-derived permeability and log-derived porosity is often developed and utilized for permeability prediction. However, a reliable correlation between permeability and porosity generally cannot be established in heterogeneous formations. This is mainly due to presence of multiple "Flow Units" within the formation. The Flow Unit is defined according to geological and petrophysical properties that influence the flow of fluids. Prediction of Flow Units is a difficult and complex problem because typically the available information from cores is inadequate. The use of well log data to identify the Flow Units and predict permeability distribution therefore, represent a significant technical as well as economic advantage. In this study, several artificial neural networks were successfully developed to predict the Flow Units, permeability and porosity from the available well log data. The study was performed in an oil filed located in West Virginia. Well log data and core analysis results from seven wells in this field were utilized to train and test the neural networks. The results of the networks' prediction in conjunction with the relevant secondary recovery data were then utilized to simulate the secondary recovery performance. The simulation results indicated that the neural network predications significantly improved the simulation of the secondary recovery performance.