An Intelligent Data-Driven Framework to Develop New Correlation to Predict Gas Deviation Factor for High-Temperature and High-Pressure Gas Reservoirs Using Artificial Neural Network

The gas deviation factor (Z-factor) is an effective thermodynamic property required to address the deviation of the real gas behavior from that of an ideal gas. Empirical models and correlations to compute Z-factor based on the equation of states (EOS) are often implicit, because they needed huge number of iterations and thus computationally very expensive. Many explicit empirical correlations are also reported in the literature to improve the simplicity; yet, no individual explicit correlation has been formulated for the complete full range of pseudoreduced temperatures and pseudo-reduced pressures, which demonstrates a significant research gap. The inaccuracy in determining gas deviation factor will lead to huge error in computing subsequent natural gas properties such as gas formation volume factor (Bg), gas compressibility (cg), and original gas in place (OGIP). Previously reported empirical correlations provide better estimation of gas deviation factor at lower pressures but at higher reservoir pressures their accuracies becomes questionable. In this study, a simple and improved Z-factor empirical model is presented in a linear fashion using a robust artificial intelligence (AI) tool, the Artificial Neural Network (ANN). The new model is trained on more than 3000 data points from laboratory experiments obtained from several published sources. The proposed model is only a function of pseudo reduced temperature and pseudo reduced pressure of the gases which makes it simpler than the existing implicit and complicated correlations. The accuracy and generalization capabilities of the proposed ANN based model is also tested against previously published correlations at low and high gas reservoir pressures on an unseen published dataset. The comparative results on a published dataset show that the new model outperformed other methods of predicting Z-factor by giving less average absolute percentage error (AAPE), less root mean square error (RMSE) and high coefficient of determination (R2). The error obtained was less than 3% compared to the measured data.