Artificial Neural Network Architectures for Predicting Two-Phase and Three-Phase Relative Permeability Characteristics

摘要

Laboratory determination of relative permeabilityrncharacteristics is labor intensive and can be complicated.rnEmpirical models to predict relative permeabilities based onrnrock and fluid properties have experienced relatively limitedrnsuccess. Hence, alternate methodologies for accuraterndetermination of relative permeability characteristics arernalways desirable.rnIn this study, two-phase and three-phase relativernpermeability predictors are developed using backpropagationrnnetworks. In this category of networks, information is passedrnfrom input layer to output layer, and calculated errors arernpropagated back to adjust the connection weights in arnsequential manner to improve the predictive capabilities of thernmodels. In the development of the models, experimentalrnrelative permeability data along with some commonly reportedrnrock and fluid properties obtained from the literature are usedrnduring the training stage, while some other data sets arernpreserved to test the prediction ability of the models. Therntwo-phase relative permeability models are found to performrnin a satisfactory manner within a wide spectrum of basic rockrnand fluid properties. Similarly, three-phase relativernpermeability models are observed to have good predictiverncapability in accurately producing the missing entries of threephaserndata sets for a series of isoperms, and in constructing thernmissing isoperms for a system under consideration.rnFurthermore, they are found to be capable of effectivelyrnpredicting the three-phase relative permeability values atrnvarious saturation combinations for systems with differentrnrock and fluid properties.