A fluid identification method for caved-fracture reservoirs based on the stacking model

It is difficult to identify the fluid properties with one single logging curve in a caved-fracture reservoir due to multi-stage tectonism, diagenetic evolution, rapid lithology change, developed fractures, and significant heterogeneity. Therefore, a stack generalization-based method is proposed for improving fluid identification accuracy. First, a traditional method with cross-plots is adopted by coring and logging data, but it is found that the identification effect of the caved-fracture reservoir fluid is poor. Second, based on the cross-plot, a correlation analysis is conducted to select sensitive logging parameters such as resistivity, compensated neutron, density, acoustic, and total hydrocarbon, which reflect changes in fluid properties, as feature vectors for the identification of fluid types in caved-fracture reservoirs. Third, support vector machine, k-nearest neighbor algorithm, decision tree, and stochastic gradient descent classification are selected as the base learners. 80% of the data sample points and the remaining 20% are selected as training and test samples for building the stacking models to verify the prediction accuracy of the model. Finally, a fully connected neural network is used as a meta-learner to control the final output. The experimental results show that the prediction accuracy of the stack generalization method can reach 88%. Compared with the cross-plot and single machine learning discriminant model, the fluid identification method based on the stack generalization model has a better effect and can improve the fluid identification effect of caved-fracture reservoirs in the study area.