A Machine Learning Model for Flow Properties Prediction in Mixed Crude Oil

The flow properties of blended crude oil (such as gel point and viscosity) are important fundamental data for the calculation of blended transportation processes. While these properties can be determined experimentally, the blending ratios of different crude oils can vary greatly, and the properties of each crude oil component can also change significantly, making experimental measurements often insufficient to meet engineering needs. In the past, the flow properties of blended crude oil were mainly calculated using empirical equations. This paper establishes a prediction method for the flow properties of blended crude oil based on ensemble learning models. After parameter training with a certain amount of physically related data, the model can accurately predict the viscosity and gel point of any crude oil blend ratio, as well as estimate the viscosity of crude oil at lower temperatures and different shear rates, based on its viscosity at a given temperature and shear rate. Studies based on 13,000 sets of blended crude oil data from 4 types and 37 ratios show that when the training data reaches 1,000 sets, the prediction error for viscosity at 15°C and 10s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup> is less than 20%, with an average absolute percentage deviation of 15.16%; the prediction error for the gel point of blended crude oil is less than 6°C. Furthermore, based on test cases of single crude oils, it was found that with 383 sets of training data, the average absolute percentage deviation of viscosity at low shear rates (e.g., 2s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup>) is 32.36% in the Newtonian region (above 25°C) and 6.56% in the non-Newtonian region.