Application of machine learning for modeling heavy oil viscosity

The viscosity of heavy oil is a critical parameter in designing surface and well production equipment, reservoir simulations, and field development projects. However, direct measurement of heavy crude oil viscosity is often costly and requires representative fluid samples, which can be challenging to obtain. This strong dependence of viscosity on temperature further complicates the process. Consequently, the industry relies on predictive correlations to estimate crude oil viscosity. These correlations, however, have significant limitations, as they tend to oversimplify complex relationships, reducing their practical utility. These models often fall short in capturing the full complexity of dependencies, especially those arising from compositional variations in heavy oil. This in return, limits their accuracy and applicability across a wider range of pressures, temperatures, and fluid compositions. Over the past two decades, machine learning (ML) has gained substantial interest in modeling heavy oil viscosity. However, the results from ML models often show high variability and inherent bias. This work develops and compares the performance of two ML models, specifically, backpropagation neural networks (BPNN) and general regression neural networks (GRNN). The proposed models were trained to predict the dead heavy-oil viscosity using a dataset consisting of 371 records and six input variables: temperature, density, C 7 + , and C 26 mole fractions, and their respective molecular weights. Descriptive statistics and correlation analysis identified temperature and density as the most influential factors in predicting heavy oil viscosity. To enhance model performance, log-transformation of viscosity and temperature were applied. The BPNN, with 15 neurons in the hidden layer, achieved an average absolute relative error (AARE) of 1.27% on the blind testing dataset, significantly outperformed the GRNN, which had an AARE of 15%. Despite this disparity in performance, both neural networks significantly outperformed several empirical correlations reported in the literature by a considerable margin. The BPNN model demonstrated superior prediction accuracy due to several key factors. Its iterative weight optimization via backpropagation allows minimizing errors and improving predictions effectively. The hidden layers of the BPNN automatically learn complex data patterns, making it particularly effective with large, high-dimensional datasets. In contrast, the GRNN, while computationally efficient, faces challenges with memory limitations and generalization. Additionally, the BPNN allows greater flexibility through fine-tuning hyperparameters such as learning rate and activation functions, making it the preferred model for predicting heavy oil viscosity.