Application of Various Machine Learning Techniques in Predicting Water Saturation in Tight Gas Sandstone Formation

Abstract

Water saturation (Sw) is a vital factor for the hydrocarbon in-place calculations. Sw is usually calculated using different equations

however, its values have been inconsistent with the experimental results due to often incorrectness of their underlying assumptions. Moreover, the main hindrance remains in these approaches due to their strong reliance on experimental analysis which are expensive and time-consuming. This study introduces the application of different machine learning (ML) methods to predict Sw from the conventional well logs. Function networks (FNs), support vector machine (SVM), and random forests (RFs) were implemented to calculate the Sw using gamma-ray log, neutron porosity log, and resistivity (Rt) log. A dataset of 782 points from two wells (well-1 and well-2) in tight gas sandstone formation was used to build and then validate the different ML models. The dataset from well-1 was applied for the ML models training and testing, then the unseen data from well-2 were used to validate the developed models. The results from FN, SVM, and RF models showed their capability of accurately predicting the Sw from the conventional well logging data. The correlation coefficient (R) values between actual and estimated Sw from the FN model were found to be 0.85 and 0.83 compared to 0.98 and 0.95 from the RF model in the case of training and testing sets, respectively. SVM model shows an R-value of 0.95 and 0.85 in the different datasets. The average absolute percentage error (AAPE) was less than 8% in the three ML models. The ML models outperform the empirical correlations that have AAPE greater than 19%. This study provides ML applications to accurately forecast the water saturation using the readily available conventional well logs without additional core analysis or well site interventions.