Connectionist Models for Asphaltene Precipitation Prediction by n-Alkane Titration─Pressure and Crude Oil Properties Considered

Experimental approaches for determining asphaltene precipitation in a laboratory are time-consuming and expensive due to consumption of a large amount of solvents. Development of robust, reliable, fast, and economic predictive tools to forecast the amount of asphaltene precipitation for a wide range of pressures, temperatures, and operational parameters and properties of petroleum fluids is inevitable. The main objective of this research work was to develop machine learning models using experimental data to predict asphaltene precipitation amount due to titration. After collecting 1439 data samples from 27 experimental research works, a quality check was performed for possible logical filling of the missing values and detecting the problematic data samples. Three categories, operational parameters, oil properties, and gas properties, were recognized to be the most influential parameters. The database used in this work is so far the largest ever reported in the literature. In addition, pressure is considered as one of the major parameters in this work, which was not considered in the previously reported models (i.e., all were conducted under ambient pressure). For the first time, 39 different oil samples were considered in the modeling (i.e., the existing works are mostly for one oil sample). We proposed new indices in the modeling to account for different oil types and n-alkanes. Due to the pressure data distribution, the database was split into two clusters. Each cluster went through several statistical preprocessing stages including treating duplicates and zero-variance features, imputing the missing data, assessing the collinearity, feature selection, and data splitting and scaling. Then, five different models, multilayer perceptron (MLP), support vector machine (SVM), decision tree (DT), random forest (RF), and committee machine intelligent system (CMIS), were used for model development. Based on the acquired results, the RF was determined as the best predictor for both clusters, consequently, for the whole database with root-mean-square error (RMSE) and R² values of 0.94 and 0.97, respectively, for the testing data set. The developed models can be used to accurately predict asphaltene precipitation by n-alkane titration for a wide range of pressure and crude oil properties.