Accurate determination of water saturation is fundamental for estimating reservoir fluid volumes and better characterization of hydrocarbon reservoirs. There are several methods available to estimate the water saturation in shaly formations but the most commonly used in the industry are those based on petrophysical models, such as Waxman-Smits and Simandoux. These models have limitations and their input parameters are often not readily available. This consequently leads to either underestimated or overestimated fluid saturations.
In this study, a methodology based on artificial neural network (ANN) models was developed and tested to predict water saturation using wire-line logs and core Dean-Stark data. A detailed workflow for developing ANN models for sandstone reservoirs, where conventional wire-line logs are taken as input parameters, is presented. The model used in this study is based on a three-layered neural network with a Resilient Back-propagation (PROP) learning algorithm.
The model was successfully tested on the Haradh sandstone formation (in Oman) yielding a prediction of water saturation with a root mean square error (RMSE) of around 2.5 saturation units (saturation measured in percentage) and a correlation factor of 0.91 on the testing data. Furthermore, the ANN model was shown to be superior to conventional statistical methods such as multiple linear regression. The robustness of the ANN model was demonstrated by using the model to predict other properties such as the volume of shale with an error of 2.
The final optimized neural network model showed that a complexity of 13 hidden neurons is suitable for the class of water saturation prediction problems that we considered in this study.
