Multilayer Perceptron Modeling for the Prediction of Gas Condensate Composition Changes During Pressure Depletion

Abstract

Abstract The production of the gas condensate in addition to the gas itself can improve the recovery factor of gas condensate reservoirs, as well as increase the economic feasibility of the reservoir. Hence, the prediction of the variation of the fluid properties of gas condensate reservoirs requires an accurate study of reservoir fluid characteristics during various pressure stages. The changes in produced gas composition during production are usually obtained from Pressure-Volume-Temperature (PVT) tests such as the Constant Volume Depletion (CVD) tests, provided the sample is available. Therefore, this study aims to develop an intelligent model to anticipate these changes in gas condensate composition during depletion stages, especially in cases where sample availability or reliability is limited. A large data bank comprising of CVD tests for 141 of gas condensate samples at various pressures, temperatures, and compositions has been used to develop a multilayer perceptron neural network model to estimate the depletion changes in gas condensate composition from its initial composition. Furthermore, in order to demonstrate the effectiveness of the proposed model in cases where the Equation of State (EOS) is not accurately calibrated, a comparison was conducted between the proposed model and the EOS for all gas condensate samples. The statistical error analysis show that model has high degree of accuracy to predict the gas composition within three depleted pressure stages. The model produced an average absolute percent error (AAPE) of 19.55% and 9.27%, and a coefficient of determination (R2) of 0.953 and 0.970, for sweet and sour gas condensates, respectively. Therefore, the model has an excellent agreement with the experimental data. All gas condensate samples in the data bank which were used to develop the model have also been used to study the accuracy of the Peng-Robinson Equation of State (PR-EOS) in comparison with the proposed model. The errors of the PR-EOS prediction of changes in the produced gas condensate composition resulted in an AAPE of 15.68% and 19.18%, and R2 of 0.936 and 0.828, for sweet and sour gas condensates, respectively. The accuracy of the proposed model in predicting the changes in the produced gas condensate composition is favorable in comparison with PR-EOS. In addition, the model eliminates uncertainties related to splitting and characterization of the plus fraction and inclusion of the binary interaction parameter correlations which are required for EOS. The neural network architecture used in the proposed model exhibits a remarkable similarity to the experimental data. Furthermore, the model is highly practical, as it can be implemented within a programming language to rapidly estimate the gas condensate's compositional variation at different pressure stages.