(57q) New Viscosity Models Based on Machine Learning and Classical Regression

Several research publications and technical papers have been dedicated to the modeling of crude oil viscosity. While experimental measurements are the most reliable means of determining oil viscosity, they are costly, time-consuming, and sometimes unfeasible due to sample unavailability. To address such challenges, various approaches like correlations, machine learning, artificial neural network models, and data mining methods have been utilized to estimate crude oil viscosity. Nevertheless, some of the published models suffer from inherent limitations. These limitations arise from their reliance on limited data and their tendency to focus on specific types of crude oil, which ultimately leads to reduced accuracy in viscosity predictions. This research aims to overcome these limitations by employing experimentally measured crude oil viscosity data from 66 samples collected at different pressures and temperatures from the Burgan oil field in Kuwait. The study presents multiple methods for predicting the viscosity of dead, saturated, at bubble point pressure, and undersaturated oil.

The most well-known and published models have been used to assess their accuracy in predicting the viscosity data of the 66 samples described in the data bank. It was found that Labedi (1992) model is the most accurate in prediction of the dead oil viscosity with an average absolute relative error (AARE) of 13%. Katoatmodjo et al. (1991) was found to have the best accuracy among all the published correlation for saturated crudes with an AARE of 8%. Abu Khamsin et al (1991) has provide the best accuracy for oil viscosity at bubble point pressure with an AARE of 14% and Abdulmajeed et al (1991) has the best accuracy for undersaturated crude with an AARE of 5.8%.

A 990 measurements has been used to developed machine learning models (MLM) and classical regression models (CRM) for the predictions of dead oil viscosity at atmospheric conditions saturated oil viscosity at pressures below the bubble point pressure, at bubble point pressure and above the bubble point pressure. The proposed CRM was able to predict the dead oil viscosity with AARE of 6.1%, the saturated oil viscosity below bubble point pressure with AARE of 5.2%, the oil viscosity at bubble point pressure with AARE of 6.3%, and the undersaturated oil viscosity above bubble point with AARE of 2.2%.