Using Neural Networks for Determining Hydrocarbons Presence from Well Logs: A Case Study for Alagoas Basin

Abstract

This work proposes a method to determine oil, gas and water presence in reservoirs directly from commonly measured well log parameters, such as resistivity, gamma-ray, density and neutron, by using Artificial Neural Networks (ANN). The wells from the oil field approached in this case study require frequent workover jobs to put their multiple zones in production one after another as soon as they become depleted. Once this oil field is segmented in hundreds of hydraulically isolated reservoirs containing small oil volumes, a procedure for identification of the fluids contained in the reservoirs crossed by a well can be of great value to make routine decisions (regarding evaluations and perforations) faster and more accurate. The work started based on the observation that zones with different fluids (or the absence of fluids) show characteristics sufficiently particular in the cited logs that make possible their classification by an artificial intelligence algorithm. The ANN approach is chosen considering its capability of working with a non-linear problem in which some uncertainties are also present, based on algorithms with relatively simple implementation. A characteristic of this method is that the ANN can be trained with data points of a specific zone, block or area, which means adjusting itself to this specific data set, learning its particularities, what leads to more accurate results when used to classify other zones with similar characteristics. The method's applicability increases in mature fields presenting a large number of logged wells with perforated and tested zones allowing validation of the information used in the supervised learning of the ANN. For the training and validation of a Multi Layer Perceptron (MLP) Neural Network, 160 data points were used with depth information and the corresponding Gamma Ray, Resistivity, Density and Neutron log values, along with fluid identifications based on the results of logs and perforations. These data points came from two wells of Alagoas Basin, Brazil. In the tests performed in this work, the ANN presented a generalization capability after extracting information from the training data, leading to a good confidence in the classification results for a testing data set.