Prediction of water saturation in a tight gas sandstone reservoir by using four intelligent methods: a comparative study

Reservoir water saturation is an important property of tight gas reservoirs. Improper calculation of water saturation leads to remarkable errors in following studies for development and production from reservoir. There are conventional methods to determine water saturation, but these methods suffer from poor generalization and cannot be applicable for various conditions of reservoirs. These methods also depend on core measurements. On the other hand, well log data are usually accessible for all the wells and provide continuous information across the well. Customary techniques are not fully capable to prepare meaningful results for predicting petrophysical properties, especially in presence of small data sets. In this regard, soft computing approaches have been used here. In this research, Support Vector Machine, Multilayer Perceptron Neural Network, Decision Tree Forest and Tree Boost methods have been employed to predict water saturation of Mesaverde tight gas sandstones located in Uinta Basin. Tree Boost and Decision Tree Forest are powerful predictors which have been applied in many research fields. Multilayer Perceptron is the most common neural network, and Support Vector Machine has been used in many petrophysical and reservoir studies. In this research, by using a small data set, the ability of these methods in predicting water saturation has been studied. Based on the data from four wells, two data set patterns were designed to evaluate training and generalization capabilities of methods. In each pattern, different combinations of well data were used. Three error indexes including correlation coefficient, average absolute error and root-mean-square error were used to compare the methods results. Results show that Support Vector Machine models perform better than other models across data sets, but there are some exceptions exhibiting better performance of Multilayer Perceptron Neural Network and Decision Tree Forest models. Correlation coefficient values vary from 0.6 to 0.8 for support vector machine, which exhibits better performance in comparison with other methods.