Predicting the performance of constant volume depletion tests for gas condensate reservoirs using artificial intelligence techniques

The development of gas condensate reservoirs relies strongly on the accuracy of pressure–volume–temperature (PVT) data. The produced volumes of gases and liquids can be predicted, under different conditions, using PVT relationships. Also, the design of surface facilities such as storage tanks and fluids separators requires detailed information about the PVT behaviors. However, measuring the PVT properties is costly and time-consuming. Therefore, this work presents a quick and reliable approach to evaluating the PVT properties for gas condensate reservoirs using artificial intelligence techniques. Constant volume depletion (CVD) data were collected from different wells. More than 1000 data sets were collected from 68 wells that cover wide ranges of fluid compositions and reservoir conditions. The collected data contain mole fractions of gas condensate components, pressure relative volume, and reservoir temperature. First, data cleaning was performed to remove the outliers utilizing the standard deviations technique. Thereafter, artificial neural network (ANN), adaptive neuro-fuzzy inference system (ANFIS), and support vector machines (SVM) were used to develop new models to predict the performance of CVD tests. Different evaluation indexes were used to assess the reliability of the developed models; root–mean–square error and correlation coefficient (R²) were used. The obtained results showed that ANN model outperforms ANFIS and SVM techniques in predicting the PVT behavior. The values of coefficient of determinations are 0.995, 0.959, and 0.948 for the testing data using ANN, ANFIS, and SVM models, respectively. The root-mean-square error was reduced from around 1.20 to less than 0.5 using the ANN model, for the testing data set. Additionally, a new correlation was developed using the ANN technique to predict the CVD performance. The new correlation showed a very acceptable accuracy and less estimation errors compared to the available models. Ultimately, this work introduces a new approach to saving time and effort in determining the PVT behavior. A new correlation is presented to provide a direct and simple approach for estimating CVD performance.