Literature Study and Model Example of Machine Learning Application for Plugging Prediction at Hydrate Inhibitor Regeneration System: Study Case

Gas hydrate is one of five solids that commonly cause flow assurance issues, resulting in loss of production opportunity (LPO). As a common problem in the upstream oil and gas industry, gas hydrate formation prevention and mitigation should be considered during the production lifetime. The most popular method is to inject hydrocarbon fluids with antifreeze chemicals called thermodynamic inhibitors such as methanol and ethylene glycol. Compared to other mitigation strategies, this method is often used instead of heating and pigging methods that contribute to non-productive time. Mono ethylene glycol (MEG) is generally chosen as a hydrate inhibitor in gas pipeline transportation due to its capability to be regenerated and reused to reduce operating costs. However, in one of the Indonesian subsea systems, actual problems arise from MEG Regeneration Unit (MRU) such as scaling and fouling which cause plugging in the lean MEG injection system. This paper aims to show a method to predict the possibility of plugging using supervised machine learning by observing the correlation between the total dissolved solids of lean MEG and other process parameters. This discusses several previous studies showing that process parameters in the field may affect the quality of lean MEG injection. Some classification algorithms are compared to evaluate the performance of plugging possibility prediction. The result of this study shows that by applying a Random Forest algorithm, the highest accuracy among other algorithms, to field process parameters, the cleanliness can be determined whether on-spec or off-spec with an average accuracy of 79-92%. Therefore, some benefits might be gained by deploying machine learning to the MRU system and can be used to optimize MRU's operation, monitoring, and maintenance strategy.