Prediction of Discrete and Continuous Wellbore Logs Values Using Artificial Neural Network, Decision Trees and Support Vectors Machine Algorithms

Abstract

ABSTRACT: Full characterization of the reservoir formations using a limited number of data is a promising process that can be granted through the integration of machine learning algorithms. Under certain circumstances, the measurement of reservoir properties could be an expensive operation and using empirical correlations to estimate these properties may not solve the problem. In this work, an artificial neural network algorithm was developed using MATLAB software to predict the porosity, the volume of shale and water saturation logs of well F14 from the Volve Field. Through multiple estimations of mean squared error, we induce that the most optimal number of hidden neurons within this input dataset is 10. The test results show that the correlations for porosity, shale volume and water saturation are around 0.997, 0.998 and 0.866, respectively. This indicates the perfect matching of the predictions with the actual data. Besides, supervised classification of the geological layers was done using decision trees and support vector machine algorithms. The optimum number of branches that construct the decision tree is found to be 20. The best quality of fitting was obtained using decision trees algorithm with observed accuracy and actual accuracy of 89.6% and 61.2%, respectively. 1. INTRODUCTION Nowadays, the collection and processing of a large set of data for multiple investigations related to addressing the industrial problems represent the most challenging issues, and applying conventional analyses may not be appropriate for extracting useful information due to the time-consuming at each operation and the higher complexity of the process. For this purpose, a lot of research was devoted to handling these problems through the integration of data mining as a major concept for the treatment and the interpretation of a variety of results in a more accurate way (Sharma and Sharma, 2018; Angra and Ahuja, 2017; Das, Dey, et al., 2015). Thus, machine learning has gotten increasing attention, especially in the field of petroleum engineering. This technique lies in finding the correlations and the rules that can best describe the behavior of the outputs in line with the expected change in inputs properties. However, many algorithms were developed for general purposes and then applied to certain studies related to oil production enhancement and several domains in petroleum engineering (Khan, Alnuaim, et al., 2019; Hegde and Gray, 2017). This includes the prediction of wellbore logs and the composition of earth formation, which is more promising in terms of reducing the number of dispenses that can be generated by the extensive measures that cover a large scale of investigation, while it is efficient and sufficient that the full identification of the formation can be done at a narrower range of depth. Furthermore, assuming a general model that can be parametrized under the most common change in properties of nearby wells in a particular region is highly recommended as an alternative method that can limit the implantation of real logs measurements in the non-exploitable area (Saputro et al., 2016; Anifowose et al., 2017, Ala Eddine et al., 2022). To do this, it is necessary to use specific algorithms that can be valid to perform these types of predictions. Several researchers reported that the utilization of Artificial Neural Network (ANN) in the prediction of wireline logs has resulted in a good fit with what can be measured using the field’s equipment (Mohaghegh, 2000; Gharib, Elsakka, et al., 2018; Baneshi, Behzadijo, et al., 2013). This has granted a new optimized method for the characterization of formations in a more efficient and accessible way. However, the extracted correlation from these approaches is needed for a better understanding of the plausible contribution of each input to the overall wireline logs. The purpose of the present study is to investigate the correlation between particular logs and the distribution of porosity, shale volume, and water saturation using ANN, based on a comparison between the obtained results and the was recorded so far using real-world instruments. The paper includes also the study of the variation in the composition of the formation as a function of depth by applying a classification method with two predicted classes of rocks including sandstone and carbonate. For the sake of this achievement, support vector machine (SVM) and decision trees (DT) algorithms were developed separately for finding the best algorithm that can give higher accuracy in terms of predictions.