WAX DISAPPEARANCE TEMPERATURE MODELING WITH MACHINE LEARNING TECHNIQUES

Abstract

Wax precipitation along with asphaltene and scale deposition present major problem for flow assurance in the oil production. Mitigation and prevention measures for wax precipitation requires understanding of wax forming conditions. Key parameter to define wax forming borderline is Wax Disappearance Temperature (WDT). The available methods to determine WDT include laboratory measurements, empirical correlations, thermodynamic modeling, and data-driven approaches. Laboratory measurements are costly and may logistically be challenging, thermodynamic approaches require detailed and accurate characterization of production fluid, and some may experience convergence issues. Data driven approaches provide a good alternative to earlier methods with almost no loss to accuracy of data. Use of machine learning (ML) techniques to determine WDT is reported in more recent literature, though limited studies are conducted. Use of more recent developments in ML for WDT determination are not well documented. Hence, an attempt was made to develop intelligent models using decision tree (DT) with boosting algorithms for WDT prediction: AdaBoost, Gradient Boosting Machines, XGBoost and CatBoost. Conventional Linear Regression (LR) and K-Nearest Neighbor (KNN) methods were also used for comparison with DT approaches. A detailed analysis of the input data from published WDT experimental studies in literature was performed that includes selection of input features, building dataset, validating data sources, and the input data analysis with statistical tools and graphical analysis. This research work resulted in building a database with 380 data points of experimental WDT. Overall, all DT boosting algorithms have performed better than the conventional LR and KNN techniques, with XGBoost and CatBoost being the top performers (R2 = 0.996 and RMSE = 0.791 and 0.7916, respectively). AdaBoost could be a model of choice if simplicity is preferred with negligible difference in performance (R2 = 0.9945 and RMSE = 0.9272) compared to the top performer. Model performance evaluation was based on both statistical assessment and graphical analysis such as parity plots and error distribution plots. Further trend analysis was assessed where all models indicate WDT increase with MW increase that validate models’ generalization and predicting capacity. The developed XGBoost and CatBoost models are superior to the existing models reported in literature, both thermodynamic and data-driven methods. The developed XGBoost and CatBoost models offer higher accuracy and better generalization with possible implications in flow assurance schemes involving wax.