Assessing the Capabilities of Bayesian-Optimized Machine Learning Paradigms for Wax Precipitation Prevention

Abstract

Wax precipitation poses a significant challenge in crude oil production and transportation. This undesirable phenomenon increases operational costs and reduces efficiency, making it imperative to accurately determine the wax appearance temperature (WAT) of crude oils to preclude wax precipitation and enhance operational efficiency. However, traditional methods for estimating WAT, including empirical correlations or simple regression models, are limited in their accuracy and generalizability. To address this limitation, this study explores diverse advanced machine learning techniques integrated with Bayesian Optimization for reliable WAT prediction. A database comprising 59 experimental instances (covering density, pour point, freezing point, and wax content) was curated to train and validate the paradigms. Subsequently, Bayesian Optimization was employed to identify optimal hyperparameters of the paradigms: Random Forest (RF), Support Vector Regression, and Gradient Boosting Decision Trees (GBDT). The performance of these models was assessed using statistical numerical metrics. The Leverage technique was also adopted for model reliability and statistical validity assessment. Spearman correlation and model intrinsic methods, including Shapley Additive Explanations and permutation importance, were equally employed to ascertain the influence of independent variables on WAT. Generally, the Bayesian-optimized paradigms enhanced predictive accuracy and averted overfitting compared to default configurations. Among these optimized paradigms, GBDT emerged as the upper-echelon paradigm, registering the highest overall R2 and the least average absolute relative deviation of 0.960 and 0.235%. Moreover, the ensembles, GBDT and RF, ranked ahead of existing underfitted-neural-network-based paradigms. Unanimously, the correlation and model-intrinsic techniques pinpointed pour point as the most significant factor dictating precipitation onset. More than 90% of the instances also fell within the valid data region, affirming model reliability and database statistical validity. This study bridges a critical gap in leveraging artificial intelligence for wax precipitation prevention cost-effectively. Future work will focus on expanding the dataset and exploring hybrid models for improved robustness.