ABSTRACT
Oil and gas firms are constantly concerned with determining their reserves. The evaluation of hydrocarbon reserves necessitates a thorough understanding and knowledge of both technical and non-technical variables such as the reservoir's nature, accessible technology, and economic conditions, among other things. The most essential parameter in evaluation is the recovery factor (RF). Several techniques for estimating oil recovery factor are currently available; however, the accuracy of those techniques is heavily influenced by data availability, which is primarily determined by field age. Some of the techniques are extremely accurate, but they necessitate a large amount of production data, limiting their applicability early in the reservoir's life. Others could be used sooner, but they have a low level of precision. Throughout the life cycle of a reservoir field, the methods utilized to evaluate recovery factors change. We often rely on analog fields and empirical methods during evaluation, prior to development, when there are no production data. Because no perfect analog exists, these procedures are usually linked with a wide range of uncertainty. Recovery variables are often connected with simulation and dynamic modeling during summit, but a decline curve analysis is employed later in the field's life, once the field has dropped off the summit.Uncertainty and potential contradictions in recovery estimates result from the employment of different methodologies at different periods of the field life. The production and recovery factor is controlled by a large number of interacting, somewhat linked reservoir and production variables. Machine learning enables more complex multivariate analysis to be used by using a training data set to investigate the roles of these variables and to predict future performance in fields. I used a data set consisting of producing reservoirs to train a series of machine learning algorithms, which could possibly predict the recovery factor with minimal percentage error to investigate that approach. The study database comprises categorical and numerical properties for the United States (U.S) reservoirs. The data sets were divided into training and test sets: approximately 80% of the total information was provided in the training sets, and the remaining 20% were tested. Linear model regression and SVM models were developed with all data set parameters (30 parameters); the performance of such models was compared to the results of two-fold cross validation with the 16 most influential parameters in the data set. The Gaussian Kernel function root-mean-square error of 0.15, mean square error of 0.02 and R-squared of 0.49, is the result of SVM training with the combined 16 geological/engineering parameter models. This model has been tested on 10 reservoirs within the testing set, and the results of the tests were very similar to the cross-validation results during the model training phase.
