Machine Learning Models for Predicting Asphaltene Stability Based on Saturates-Aromatics-Resins-Asphaltenes

Abstract

Summary Asphaltene precipitation is one of the challenging flow assurance problems as it can cause permeability impairment and pipeline blockages by depositing on the surface of well tubing, flowlines, and heat exchangers. The cost of remediating an unexpected asphaltene problem is expensive and time-consuming wherever offshore or on land. Thus, the provision of asphaltene precipitation is vital and an effective approach is stability screening for monitoring asphaltene precipitation problems. In this study, asphaltene stability performance in crude oil was evaluated using six machine learning (ML) techniques, namely decision tree (DT), Naïve Bayes (NB), support vector machine (SVM), artificial neural networks (ANN), random forest (RF), and k-nearest neighbor (KNN). A large stability data containing 186 crude oil samples of known stability were used to design the classification models for predicting asphaltene stability. The inputs to the models were the content of saturates, aromatics, resins, and asphaltenes (SARA); and the output was stability. The classification results showed that the best classification model is the KNN classifier, and it has an accuracy of 82%, area under the curve (AUC) of 83%, precision of 75%, and f1-score of 83%. Also, three empirical correlations with high accuracy including stability index (SI), stability crossplot (SCP), and asphaltene stability predicting model (ANJIS) were utilized comparatively with the ML models to evaluate asphaltene stability. Results revealed that the KNN classifier has superior performance in this work with an accuracy of 80%, a precision of 82%, and an f1-score of 79%. Results of this study showed that ML is effective for asphaltene stability, providing potential in asphaltene management to reduce asphaltene deposition risk in production.