A New Approach to Choke Flow Models Using Machine Learning Algorithms

Abstract

Abstract Computer Science Technology has been widely used for simulation of Gas and Petroleum Networks. Wellhead chokes or Pressure Control Valves are specialized equipment used extensively in the Hydrocarbon Industry for two purposes; to maintain stable downstream pressure from the wells, and to provide necessary backpressure to balance gas well productivity while controlling downhole drawdown. Use of multiphase choke flow models and empirical choke flow equations have been developed in the past half-century to improve gas estimation at different fluid, flow regime, flow types and pressure drop scenarios. All these have carried over certain measurement errors which make it difficult to predict well performance parameters with the mentioned methods. Traditional models use sonic flow equation and Gilbert-type formulae for critical flow of multiphase choke cases as a base line. Evolution of new models capture further regression refinements, constrains values and multiple regression studies at different pressure drop, PVT properties, gas-liquid ratios, and choke sizes. The new Algorithm has been developed by using Random Forests Regression (RFR) and has applied the help and learn method to data classification by constructing a multitude of decision trees for stored measurements of multiple gas production variables. A decade ago a second generation of choke equation models was developed, consolidating multiple databases from production operations. This choke equation has been used extensively showing single digit errors in most of gas estimations when compared against conventional well testing physical equipment readings. The use of this 2010 Choke Gas Equation (Ref. 9) has been valuable on reducing use of conventional testing equipment without jeopardizing data quality. However, the prediction error of these models starts to increase in deviating conditions such as low gas rates or increased water and condensate ratios. New data collection has been taking place considering multiple different scenarios, different time laps and additional variables. These new and enhanced databases help evaluate new models and better data-driven analytics. The application of this algorithm improves the prediction accuracy compared to traditional regression methods as it captures more of the variance in the data, thus the implementation of RFR and enables more accurate prediction of the Separator rate for the overall gas wells in the field. This paper, explains and applies the machine learning algorithm known as RFR (Random Forest Regression) and compare with GPR (Gaussian Process Regression) to this particular request on Gas Production Engineering metering. The algorithm allows the computer to understand underlying patterns in the data and make better predictions based on different regression trees and their use for nonlinear multiple regressions. This paper explains the application of RFR and GPR methods to the separator gas rate estimation, and shows better prediction results. This paper also explains and application of those two-machine learning algorithm (Random Forest Regression and Gaussian Process Regression) helping us to predict gas volume, using choke size, upstream and downstream flowing pressures, condensate to gas ratio (CGR) and upstream temperatures. These approaches are benchmarked against the first (back 2005) and second models (Ref. 9) and demonstrate a drastic reduction in prediction error and a more robust ability to manage high variability in the data in comparison previous models using single variable statistics tools.