A Data-Driven Machine Learning Approach to Predict the Natural Gas Density of Pure and Mixed Hydrocarbons

Natural gas is one of the main fossil energy resources, and its density is an effective thermodynamic property, which is required in almost every pressure-volume-temperature (PVT) calculation. Conventionally, the density of natural gas is determined from the gas deviation (Z-) factor using an equation of states (EOS). Several models have been developed to estimate the Z-factor utilizing the EOS approach, however, most of these models involve complex calculations and require many input parameters. In this study, an improved natural gas density prediction model is presented using robust machine learning techniques such as artificial neural networks and functional networks. A total of 3800 data points were collected from different published sources covering a wide range of input parameters. Moreover, explicit empirical correlations are also derived that can be used explicitly without the need for any machine learning-based software. The proposed correlations are a function of molecular weight (Mw) of natural gas, pseudo-reduced pressure (Ppr), and pseudo-reduced temperature (Tpr). The proposed correlations can be applied for the gases having Mw between 16 and 129.7 g, Ppr range of 0.02-29.3, and Tpr range 0.of 5-2.7. The prediction of the new correlation was compared against the most common methods for determining the natural gas density. The developed correlation showed better estimation than the common prediction models. The estimation error was reduced by 2% on average using the new correlations, and the coefficient of determination (R2) was 0.98 using the developed correlation.