APPLICATION AND COMPARISON OF ENSEMBLE MODELS TO PREDICT CORROSION RATE FOR AMINE REGENERATION UNIT IN REFINERY

Abstract

Monitoring the corrosion activity of Amine Regeneration Units (ARU) in oil and gas refneries is crucial to avoid unexpected failures and unplanned shutdowns. Various corrosion monitoring devices are installed to provide real time metal loss readings translated into corrosion rates. However, existing corrosion monitoring devices do not ofer prediction capabilities, and a huge amount of supervised historical data is usually ignored due to its complex characteristics and incompleteness. By utilising ensemble models, machine learning can be adopted to generate value from the unattended historical data and provide prediction capabilities. This research work is mainly intended to process the historical data in Amine Regeneration Unit (ARU), implement Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting (LightGBM) ensemble models for corrosion rate prediction, and compare the two to identify the most accurate model. An experiment was conducted via simulation through programming software that utilises Python as the programming language to process data from an undisclosed oil and gas refnery in Malaysia. Evaluation of accuracy was performed using statistical methods, which include R-squared (R2 ), Mean Average Error (MAE), Mean Squared Error (MSE), and Root-Mean-Square Error (RMSE). Upon model hyperparameter tuning and comparison between eight diferent models, the model utilising the XGBoost algorithm with K-Fold cross-validation was reported to provide the best accuracy in terms of R2 score at 65.4%. Future improvement for this research work includes exploration of various other corrosion groups in ARU among multiple oil and gas refneries and experimentation on various boosting tree hyperparameter tuning.
 Keywords: Amine, corrosion rate, LightGBM, oil and gas, predictive analytics, refinery, XGBoost