Modeling apparent viscosity of waxy crude oils doped with polymers using tree-based models, radial basis function neural networks, and Gaussian process regression

Abstract

The probability of wax precipitation increases as temperature decreases and this happens during the flow of crude oil from a high temperature reservoir to the cool surface facilities. The present work aims to evaluate the apparent viscosity of waxy crude oils in two conditions: without presence of polymers and doped with them. To reach this purpose, a comprehensive databank consists of 622 actual values of waxy crude oils’ apparent viscosity were gathered. Various intelligent networks including Radial Basis Function (RBF) coupled with three algorithms, namely Teaching-Learning Based Optimizer (TLBO), Marine Predators Algorithm (MPA), and Barnacles Mating Optimizer (BMO), Decision Tree (DT), Random Forest (RF), Extra Tree (ET), and Gaussian Process Regression (GPR) were trained to predict waxy crude oil viscosity values. According to the findings, the GPR model demonstrates exemplary performance and offers forecasts with the minimum average absolute percent relative error (AAPRE = 0.30 %), the highest coefficient of determination (R2 = 0.999), and the lowest standard deviation (SD = 0.011). The accuracy of the other models was also acceptable and overall, tree-based techniques (except ET) and GPR provided better results than RBF coupled with different optimizers. Ultimately, the Leverage method was applied to identify the suspected dataset utilizing outlier detection. The outlier discovery represented that majority of data points, which were used for modeling in this paper, are statistically valid and trustworthy.