Empirical and Data Driven Approaches for Optimizing Water and Oil Separation in Offshore Desalting Systems in Early-Stage Production Fields

This study investigates a combined approach of empirical field testing and data-driven modeling using machine learning to optimize water and oil separation in offshore desalting systems during early-stage production. The goal is to evaluate the influence of key variables and derive actionable insights for process improvement. Before conducting field tests aimed at optimizing dehydration and desalting efficiency, preliminary data analysis was performed to investigate potential causes of non-compliance with the treated oil's water content specification (BS&W). Although this initial phase could not clearly determine the influence of all relevant variables, three main hypotheses were identified: (1) increasing demulsifier dosages could improve water separation efficiency; (2) the volume of freshwater used during desalting, while necessary to reduce salinity, might impact BS&W if demulsifier dosages are inadequate; and (3) high shear forces at the mixing valve upstream of the second electrostatic treater could cause emulsification that resists treatment at this stage. Field tests validated the preliminary hypotheses and mapped the influence of additional variables through well-defined steps, such as the importance of the treater's produced water pump recycling flow (minimal flow control) into the vessel's inlet. This recycling, combined with freshwater flow, contributes to the total water content at the second electrostatic treater's inlet. The main operational finding highlights the direct impact of increased water content at the desalting stage inlet (second electrostatic treater) on the treated oil's BS&W. Contrary to industry practices and equipment manufacturer recommendations that require a minimum water content of approximately 5% at the vessel's inlet, no performance improvements were observed when adhering to this criterion. Additionally, reducing the pressure differential at the mixing valve—where freshwater is blended with the oil stream before the second electrostatic treatment—proved critical. A relatively small reduction of 15 kPa, achieved by opening the valve from 70% to 100%, significantly decreased emulsification and consequently lowered the BS&W of treated oil. This outcome, while theoretically possible, was not expected and demonstrated the sensitivity of freshwater emulsification within the unit. Following field testing, the data model was retrained to enhance predictions and refine the assessment of variable importance. This time, both regression and classification models were evaluated. This step resulted in models with improved accuracy and greater resolution regarding the impact of variables, particularly those less clearly understood during the initial phase. The findings underscore the value of conducting well-structured tests to explore variable influences, even when using advanced machine learning models such as Random Forest. The approach presented in this article aimed to reduce subjectivity in interpreting the impact of process variables on test results, especially those with less evident effects.