Robust asphaltene onset pressure prediction using ensemble learning

Abstract

Most works on asphaltene onset pressure (AOP) prediction rely on a single model without making them robust against noise. This paper adopts a robust approach to training three machine learning models—Multi-Layer Perceptron (MLP), CatBoost, and Random Forest (RF)—to predict AOP as a function of oil composition, SARA fractions, saturation pressure, and temperature. Moreover, a Power-Law Ensemble Model (PLEM) is used to integrate the predictions obtained from the individual experts. Robustness was achieved by tuning the hyperparameters using 5-fold cross-validation with multiple sets of noisy data. Results revealed that the robust approach enabled models to outperform standard ones on noisy data by >10 %, while keeping a good performance on original data. The PLEM could increase the coefficient of determination by a minimum of 3.4 % relative to the best individual data-driven model, regardless of the tuning strategy. Additionally, the proposed approach outperformed thermodynamic and mathematical models. In the end, Sobol's sensitivity analysis showed that the concentration of C1C7+, asphaltenes, saturates, CO2, and the saturation pressure had a positive impact on AOP, while the temperature and the concentration of N2, H2S, resins, and aromatics had negative effects. The value of the work is leveraging three distinct models to predict AOP and employing a robust hyperparameter tuning approach to make the models robust against measurement errors. The findings indicate that robust hyperparameter tuning increases the stability of the models, while combining the outputs of the models improves the prediction accuracy.

Keywords

• Asphaltene onset pressure
• Machine learning
• Ensemble learning
• Robust optimization
• Digital oilfield