A Reliable Multilayer Perceptron Neural Network Estimator of Gas Condensate Reservoir Initial Composition and Dew Point Pressure

Abstract

Abstract An accurate prediction of the pressure depletion of gas condensate reservoir is important for evaluation, reserve estimation, planning production and future gas cycling to improve recovery. Constant volume depletion (CVD) tests provide various gas condensate fluid compositions at initial pressure and during reservoir depletion behavior. In addition, they provide the dew point pressure, amount of gas and oil recovery, saturation of accumulated retrograde, and compressibility factor. In case of unavailability or unreliability of the PVT test, alternative methods are vital to estimate these properties. Therefore, the objective of this study is to develop a model that can estimate the gas condensate’s initial composition and dew point pressure using neural network models. The model was created by utilizing a large data bank of CVD tests comprising 141 of gas condensate samples at various pressures, temperatures, and compositions. The developed model used a multilayer perceptron neural network model to accurately predict the initial gas condensate composition from composition of produced gas at pressures below dew point, in case the initial composition and dew point is not measured due to sample unavailability. Additionally, a comparison between the proposed model and Equation of State (EOS) is performed for all gas condensate samples to show the advantage of the proposed model when the EOS is not properly tuned. The developed model has a high degree of accuracy to predict initial gas composition with an average absolute percent error (AAPE) of 2.59% and a coefficient of determination (R2) of 0.999 for sweet gas condensate samples, and AAPE of 2.26%, and R2 of 0.982 for sour gas condensate samples. The model also predicts the dew point pressure with an AAPE of 2.45% and R2 of 0.992 for sweet gas condensate samples, and an AAPE of 1.86%, and R2 of 0.984 for sour gas condensate samples. The models’ performance has been compared to the Peng-Robinson EOS (PR-EOS). It was found that the untuned PR-EOS predicted the dew point pressure with AAPE error of 31.01% and a R2 of 0.555. This displays an excellent performance of the developed model’s estimations of the gas condensate’s initial composition and its dew point pressure. The newly proposed model utilizes the neural network architecture and has an excellent agreement with the experimental data. Moreover, the model is quite practical since it can be integrated in any spreadsheet program or implemented within a program language.