Abstract
Asphaltene precipitation during natural depletion can lead to substantial issues in oil transport and production systems, requiring close assessment of crude oil stability through thermodynamic and structural factors. This precipitation can alter wettability, reduce permeability, and block flow, causing increased pressure drops in wells, upstream facilities, and pipelines. Determining asphaltene precipitation experimentally is time-consuming and costly. Thus, it is essential to find fast and precise techniques for determining asphaltene precipitation. This research aims to accurately predict asphaltene precipitation values using an extensive databank containing 380 experimental data points. Pressure, bubble point pressure, temperature, oil °API gravity, and saturated-aromatic-resin-asphaltene (SARA) fractions of the crude oil samples were treated as input variables, while asphaltene precipitation was the output of the models. Four smart models, namely, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), cascade forward neural network (CFNN), and generalized regression neural network (GRNN) were constructed. To train the CFNN, the Levenberg-Marquardt (LM) algorithm was implemented. Various statistical and visual assessments were conducted to validate the reliability and precision of constructed models. Findings revealed that the developed LightGBM technique could outperform other models in estimating output values with the lowest absolute error value of 8.57% and the highest correlation coefficient (R(2)) of 0.9904. Besides, all developed models in this study could forecast asphaltene precipitation values with an absolute error of less than 10%. Also, trend assessment indicated that predicted output values by the LightGBM paradigm have a satisfactory agreement with actual asphaltene precipitation values. Sensitivity evaluation indicated that pressure and the amount of asphaltene exert the most significant negative and positive effects on the output, respectively. Lately, outlier discovery utilizing the Leverage technique illustrated that almost 96% of the used data points seem to be valid and reliable, statistically.