Asphaltene Stability Prediction Using Hybrid Artificial Neural Network Modeling Approach

Abstract

Abstract Accurate asphaltene stability estimation is imperative in oil and gas operations in preventing or mitigating the aggravation of problems associated with asphaltene precipitation and deposition such as flow assurance challenges and production halt or decline. However, experimental techniques for assessing asphaltene stability are time-consuming and expensive. Extant-developed models provide quick stability outcomes, but high accuracy remains a major drawback of these models. Considering these limitations, this study developed a hybrid supervised machine learning (ML) model to accurately predict the asphaltene stability honoring 129 (79 stable and 50 unstable) crude oil density and SARA fractions data points extracted from the literature. Specifically, the predictive prowess of three stability parameters (colloidal stability index (CSI), colloidal instability index (CII), and stability index (SI)) and an artificial neural network (ANN) were coupled. The collected data was preprocessed and subsequently explored for its statistical features. The data was split in a 70:30 ratio for model training and testing. The model performance was optimized via hyperparameter tuning. Classification evaluation metrics such as accuracy, precision, recall, specificity, and F1 score were utilized to assess the hybrid model's performance. Subsequently, the hybrid model's predictive performance was compared with other ML models (decision tree (DT), logistic regression (LR), and random forest (RF)) and empirical correlations (CSI, CII, SI, and Abdus, Nimra, Javed, Imran & Shaine (ANJIS) asphaltene stability predicting model). Based on the Spearman correlation output, asphaltene stability negatively correlated with CSI, CII, and SI. Thus, high CSI, CII, and SI would promote asphaltene precipitation and deposition. The hybrid ANN model exhibited remarkable asphaltene stability prediction performance with accuracy, precision, recall, specificity, and an F1 score of 100% for the training set. Similarly, the model achieved accuracy, precision, recall, specificity, and F1 scores of 97%, 95%, 100%, 95%, and 97% respectively for the test set. Data proportions caused slight variations in model testing performance while the training performance remained unaffected, which signifies the hybrid model's robustness. The hybrid model also outperformed DT, LR, RF, CSI, CII, SI, and ANJIS predictors, demonstrating the novel hybrid ML model's accuracy, reliability, and generalization capability. CSI and CII traded their true positive (stable crude oil) prediction rates (40% and 10% respectively) for high precision and true negative (unstable crude oil) prediction rates. SI also accurately classified 18/20 of the stable crude oils and misclassified 13/19 of the unstable crude oils. However, the ANJIS model exhibited moderate performance in asphaltene stability prediction, achieving ~70% accuracy, precision, recall, specificity, and F1 score. Novel/Additive Information The hybrid ML model would significantly reduce experimental time, minimize cost, and reduce uncertainties surrounding the previously developed models in the prevention and mitigation of asphaltene precipitation and deposition.