Different Machine Learning Approaches to predict Gas Deviation Factor (Z-factor)

Abstract

The gas compressibility factor indicates the gas deviation from ideal gas behavior. Accurate values of gas compressibility factor affect the estimation of reservoir fluid properties, the initial gas in place, and the natural gas production and transportation process. Gas compressibility factor can be estimated in labs; however, this method is expensive and time-consuming. Due to these challenges, numerous studies created various empirical correlations depending on the results of the equation of state. The Standing and Katz chart is regarded as a standard for estimating gas compressibility factor. Many studies proposed approaches and correlations to fit this chart, however some did not cover the entire range of data, others provided implicit methods taking long time for calculation or faced high errors out of the data range. In this study, Support Vector Machine, Radial Basis Function, and Functional Network as machine learning approaches were implemented to predict the gas compressibility factor, based on 5490 data set of Standing and Katz chart. 70% of the data set was implemented in the training process and 30% in the testing process. The data set included pseudo-reduced pressure and pseudo-reduced temperature as inputs and Zfactor as an output. Different training functions were examined for each method for the best approach optimization. In addition, machine learning best approach was compared with other correlations. The best results in this work were obtained from Radial Basis Function with 0.14 average absolute percentage error and 0.99 correlation coefficient. The developed machine learning approach performed better than the examined correlations.