Classificação de litofácies através da análise automática de perfis elétricos de poços de petróleo da Amazônia

Abstract

Among several steps which are necessary for the commercialization of oil, the analysis of well logs plays an important role to estimate the capacity of a well. Traditionally, this analysis is conducted in a semi-automated process which generates graphs of curves used by human experts to analyze and make the reservoir characterization. One goal of this analysis is to classify lithofacies. Lithofacies are lithological units(rocks) that characterize the environment and compositional aspects of the rocks. In order to characterize an oil reservoir, a set of classes of sedimentary rocks occur. This is which is the major reason for the classification of lithofacies. This master thesis investigates the use of automatic classification techniques applied to the problem of classification of lithofacies. The following five classification methods are investigated: Support Vector Machines, k-Nearest Neighbor, Multilayer Perceptron and Logistic Regression. The database investigated consists of samples from three oil wells of the same reservoir in the Amazon State. In addition, the performance of individual classifiers are compared to the combination of the same five classifiers through majority voting. Finally, we will verify whether or not individual classifiers, or ensemble of classifiers, may train using data obtained from one well and accurately classify data from other wells. In order to get these answers, we have run two series of experiments. First, we trained classifiers and test classifiers individually and combined within the same oil well. The obtained results show that Support Vector Machines achieved the best results in two of the three wells, while Multilayer Perceptron ouperformed the other methods in the third well. In the second series of experiments, we trained classifiers with data from a well and them with data from another well, simulating a situation closer to a real application, since we may use a manually classified database to train a classifier, or ensemble of classifiers, in orde to learn the pattern of the reservoir. Then, data from other wells of the same reservoir may be automatically classified. In this test, the ensemble of classifiers outperformed individual classifiers in 4 of the 6 possible combinations. In the two other combinations, the combination by majority vote was the second best. It is also worth saying that in average, ensemble of classifiers was the best option to classify lithofacies. Our results indicate that combining classifiers in a system of majority voting, shows a better performance and better stability of the results.