Petroleum viscosity modeling using least squares and ANN methods

Abstract

274 crude oils pertaining to the groups of extra light (gas condensates), light, medium, heavy, and extra heavy crude oils were characterized by true boiling point distillation, specific gravity and kinematic viscosity at 21.11 and 37.78 °C. Eight published regression empirical methods were examined for their capability of accurately predicting the crude oil viscosity. Among them the model of Kotzakoulakis and George was found to provide the lowest average absolute relative error (AARE) of 24.0% with AARE of 21.5% for the crude oils containing <30 wt% vacuum residue (VR) and AARE of 37.2% for the crude oils having >30 wt% VR. The model of Aboul-Seoud and Moharam exhibited the lowest AARE (16.3%) for the crude oils with <30 wt% VR. A new nonlinear regression model was developed that predicted the viscosity of the 274 crude oils with AARE of 19.5%, with AARE of 14.9% for the crude oils containing <30 wt% VR, and AARE of 42.0% for the crude oils having >30 wt% VR. Another model based on the artificial neural network (ANN) technique was developed. The ANN model predicted the viscosity of the 274 crude oils with AARE of 44.3%, with AARE of 50.2% for the crude oils with <30 wt% VR, and AARE of 13.9% for the crude oils containing >30 wt% VR. The combination of predicting the viscosity of crude oils having <30 wt% VR by the new nonlinear regression model with the predicting of viscosity of crude oils with >30 wt% VR by the ANN model provides of AARE of 14.9% of viscosity prediction for the entire data base of 274 crude oils.