Prediction of Viscosity of Blends of Heavy Oils with Diluents by Empirical Correlations and Artificial Neural Network

Various straight-run and hydrocracked vacuum residual oils mixed with dissimilar light oils were tested for their viscosity using an Engler specific viscometer, generating 158 heavy oil blend viscosity data points. Twenty-one available from the literature empirical correlations were tested for their capability to accurately predict viscosity. It was confirmed that the heavy oil blend viscosity exponentially decreases with the diluent concentration enhancement. The linearized form of double logarithm Walther’s equation, using the concept of viscosity blending index, was found to be suitable to model not only the viscosity of distinct residual oils and bitumen with diverse light oil diluents but also heavy crude oil blend viscosity with an accuracy very close to that of the equation with interaction parameters. The model parameters, however, are found to be specific to the experimental data and must be tuned to the particular viscosity dataset. The results of this work confirm the assertion of other researchers that the artificial neural network (ANN) approach provides a higher accuracy of viscosity prediction of mixtures of heavy oil and diluent, compared with empirical correlations. For a set of 109 viscosity data points of vacuum residual oil–diluent mixtures, the best empirical correlation shows an average absolute deviation percentage (% AAD) of 6.7, while the ANN model shows a value of % AAD = 2.2.