The Use of Artificial Neural Networks for Prediction of Water in Oil Emulsions' Viscosity from Brazilian Light Oils

Abstract

Brazilian offshore activity has increased substantially in recent years, with many new oil fields being developed, and there is also a significant investment in the maintenance and optimization of existing ones. In all cases, the presence of water-in-oil emulsions during oil production is a critical issue, causing pressure drops in subsea lines and adding complexity to petroleum processing, resulting in a loss of productivity and quality of the produced oil. The factors mentioned can determine the technical and economic viability of offshore oil production, so predicting this property is crucial for both the project and operational stages, although it is not an easy task to accomplish. Several empirical correlations are present in the open literature to predict the viscosity of emulsions, but usually, they are not accurate enough to be directly applied to Brazilian oils. In this paper, a machine learning approach based on the review of the literature and good practices used in the oil and gas industry and other engineering fields is proposed to predict water in oil emulsions viscosity. Was utilized 726 data points of light oil from different Brazilian fields to train an Artificial Neural Network (ANN). The input variables for the regression problem were temperature, water cut, shear rate, and °API, while the output was the relative viscosity of the emulsion. The Python programming language was used for statistical treatment, data processing, mathematical modeling, and resolution of the presented problem. After training the ANN, the resulting model demonstrated good performance, with a coefficient of determination (R2) above 0.99 for the data used for testing. The final model obtained underwent cross-validation and the mean value for R2 was also above 0.99, proving the methodology's capability to create generic models for the presented problem.