Neural network approach to prediction of the liquid petroleum products viscosity

Abstract

An approach to predicting the viscosity and density of petroleum products using artificial neural networks was proposed. Based on reliable data on the thermophysical properties, viscosity and density of petroleum products, models of multi-layer neural networks with a different number of hidden layers were developed, which were trained on an array of experimental data using the Levenberg-Marquardt and Bayesian regularization algorithms. The test results showed that models of neural networks trained by the Levenberg-Marquardt algorithm are unable to predict the viscosity with sufficient accuracy for practical purposes for those values of input signals that were not involved in their training. The best predictive capabilities in terms of the ratio of accuracy and computational costs were provided by neural networks with three hidden layers, trained by the Bayesian regularization algorithm, for which the average relative deviation of the calculated deviations from the experimental data was 0.8%