Method of lithology identification in carbonates reservoirs using well logs based on deep forest

Abstract

Lithology identification of complex carbonate reservoirs is very important for fine characterization and quantitative evaluation of reservoirs. In order to solve the problem of complex carbonate reservoir lithologic logging response with strong multi solutions, this paper introduces the deep forest algorithm to deeply mine lithologic information from the conventional logging calibrated by core descriptions, so as to improve the lithology identification accuracy of complex carbonate reservoir. Deep forest is a combination of random forest algorithm and neural network. It avoids the shortage of long training time of deep neural network through the cascade structure of feedforward and can automatically adjust the model structure. Taking the complex carbonate reservoir of Paleogene Oligocene-Neogene Miocene AS formation in B oilfield of Zagros basin in the Middle East as an example, the deep forest is used to identify and analyze 8 kinds of lithology. On the basis of sensitivity analysis of lithologic logging curves, 9 curves such as gamma ray, compensative density and photoelectric absorption cross-section index are selected to build a lithology identification model based on deep forest. The results of the posterior well analysis show that the deep forest algorithm is better than the traditional random forest algorithm for thin layer lithology and transition lithology identification, and the lithology identification accuracy is 94.4%. The influence of the number of training samples and the setting of model parameters on the lithology identification results shows that the deep forest method can improve nearly up to 2% compared with the random forest method under different training samples, and its default parameters can achieve a lithology identification accuracy of more than 91%.