Prediction of Gas Viscosity of Yemeni Gas Fields Using Machine Learning Techniques

Abstract

Abstract Gas viscosity is an important physical property that controls and influences the flow of gas through porous media and pipe networks. An accurate gas viscosity model is essential for use with reservoir and process simulators. The objective of this study is to assess the predictability of gas viscosity of Yemeni gas fields using machine learning techniques. Performance of some machine learning techniques in the prediction of gas viscosity investigated in this work. The techniques include K-nearest neighbors (KNN), Random Forest (RF), Multiple Linear Regression (MLR), and Decision Tree (DT). About 440 data points were collected from different Yemeni gas fields were used to develop the machine-learning model. The input data used in the training include pressure, temperature, gas density, specific gravity, gas formation volume factor, gas deviation factor, gas molecular weight, pseudo-reduced temperature and pressure, pseudo-critical temperature and pressure, and non-hydrocarbon gas components (N2, CO2, and H2S). Part of the data (75%) was used to train the developed models using the algorithms while another part of the data (25%) was used to predict the viscosity of gas for samples. Trained machine learning models were constructed using the Python programming language. The performance and accuracy of the machine learning models were tested and compared their results based on four different functional input datasets. The result of this study found that that the DT model predicted the gas viscosity with higher accuracy, and gave very good results better than other models based on input parameters of the dataset (A) and (B). This was evidenced by lower the Root mean square error (0.000832), lower mean absolute percent relative error (0.042%), and higher coefficient of determination (R2=0.9465). The proposed approach in the present study provides an accurate and inexpensive model for estimating the viscosity of gases as a function of all input parameters of the dataset (A). Overall, the relative effects of these different input parameters have verified that the gas viscosity has the uppermost relevant to the gas density and specific gravity that have the highest percentage of 51%.