Improved prediction of heavy oil viscosity at various conditions utilizing various supervised machine learning regression
Faisal Aladwani, Adel Elsharkawy
Abstract
Fluid viscosity plays a major role in the petroleum industry. It’s required for fluid flow calculations through the reservoir and production systems. In this study, heavy oil viscosity measurements for seventy samples are reported at various temperatures with different API gravity. Three supervised machine learning regression (SMLR) models were developed to predict the viscosity as a function of API gravity, temperature, and density with a 5 K-fold validation technique for a better model representation. Results showed that the newly developed models have superiority over published models. The Gaussian process regression (GPR) and the regression ensembles tree (RET) showed the best performance with a mean absolute percent error of 0.7328%, a root mean square error of 0.6174 and a coefficient of determination of 0.9988 for the GPR where for the RET the mean absolute percent error is 0.5768%, the root mean square error is 3.514, and the coefficient of determination is 0.9971.