Comparative study of machine learning algorithms in predicting asphaltene precipitation with a novel validation technique

Several thermodynamic models have been proposed in literature to estimate the amount of asphaltene precipitation (AAP); however, they usually need several many inputs, and the characterization of the samples is often not accurate enough. This paper compares the performance of four data-driven methods, including Extreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), Least Squares – Support Vector Machine coupled with a Particle Swarm Optimization (LSSVM-PSO), and Multilinear Regression (MLR), to predict the AAP as a function of oil composition, API, SARA fractions, solvent molar mass, dilution ratio, pressure, and temperature. The dataset includes 1703 samples, 20% of which is used for testing. An innovative nested K-fold cross-validation is also proposed to tune the hyperparameters of the data-driven methods. The contributions of this work include a comparison of the performance of different data-driven methods to estimate the AAP and introducing a novel cross-validation technique. Data-driven results are compared with those of the Perturbed Chain – Statistical Associating Fluid Theory Equation of State. The results reveal the superiority of the data-driven methods over the thermodynamic model, except for the MLR. Meanwhile, XGBoost showed the best performance among other data-driven methods. Coefficients of determination of 99.57%, 98.96%, 98.17%, 85.23%, and 90.40% were achieved by the XGBoost, LSSVM-PSO, MLP, MLR, and the thermodynamic model, respectively. Finally, it is shown that the proposed nested K-fold cross-validation positively affected the generalization of the data-driven methods. The findings of this study can help engineers select reliable methods to estimate the AAP and improve their generalization using the nested K-fold cross-validation.