Application of Machine Learning to Evaluate the Performances of Various Downhole Centrifugal Separator Types in Oil and Gas Production Systems

Abstract

Pumping artificial lift techniques, such as rod pumps and ESPs, are applied for gassy wells more than ever before. This has made the downhole separators a critical part of most such installations. There are multiple categories of downhole separators, with various techniques developed to assess and improve their performances, but no general guidelines are established for their application. This paper aims to classify the separator types and review their performances in the open literature. In addition, various data sets are collected and put together to evaluate and rank downhole centrifugal separators using data analysis and machine learning (ML) techniques. A comprehensive literature review is conducted to collect the available downhole separator performance data. Experiments and Computational Fluid Dynamic (CFD) simulations are the techniques used by the researchers. This information is collected to identify the optimum conditions for each separator type, considering the effects of liquid and gas rates and other flow parameters. The data collected from various research projects over the last 20 years are combined to make a comprehensive downhole separation databank. These data are analyzed using various machine learning algorithms to rank the performances of downhole separators at various operating conditions. Various downhole separators have been tested in the open literature, including poor-boy separators, two-stage separators, packer-type separators, rotary and spiral separators with different designs, etc. A critical factor that adds to the uncertainty is the separator's control system, which significantly affects its efficiency. The available data show that most separators provide separation efficiencies higher than 80% if the downstream casing valve is adequately controlled. The separation efficiencies decline as the liquid and gas rates increase past an upper limit. The collected data from multiple previous studies form a broad dataset. Data analysis is used to compare the performances of different downhole separator classes, and machine learning is applied to identify a robust prediction model. This paper gathers, interconnects, and examines several available research works through data analytics. The results provide a fundamental source and a valuable guideline for downhole liquid-gas separation, particularly in artificial lift applications.