A predictive model for steady-state multiphase pipe flow: Machine learning on lab data

Abstract

Engineering simulators used for steady-state multiphase flows in oil and gas wells and pipelines are commonly utilized to predict pressure drop and phase velocities. Such simulators are typically based on either empirical correlations (e.g., Beggs and Brill, Mukherjee and Brill, Duns and Ros) or first-principles mechanistic models (e.g., Ansari, Xiao, TUFFP Unified, Leda Flow Point model, OLGAS). The simulators allow one to evaluate the pressure drop in a multiphase pipe flow with acceptable accuracy. However, the only shortcoming of these correlations and mechanistic models is their applicability (besides steady-state versions of transient simulators such as Leda Flow and OLGA). Empirical correlations are commonly applicable in their respective ranges of data fitting; and mechanistic models are limited by the applicability of the empirically based closure relations that are a part of such models. In order to extend the applicability and the accuracy of the existing accessible methods, a method of pressure drop calculation in the pipeline is proposed. The method is based on well segmentation and calculation of the pressure gradient in each segment using three surrogate models based on Machine Learning (ML) algorithms trained on a representative lab data set from the open literature. The first model predicts the value of a liquid holdup in the segment, the second one determines the flow pattern, and the third one is used to estimate the pressure gradient. To build these models, several ML algorithms are trained such as Random Forest, Gradient Boosting Decision Trees, Support Vector Machine, and Artificial Neural Network, and their predictive abilities are cross-compared. The proposed method for pressure gradient calculation yields R2=0.95 by using the Gradient Boosting algorithm as compared with R2=0.92 in case of Mukherjee and Brill correlation and R2=0.91 when a combination of Ansari and Xiao mechanistic models is utilized. The application of the above-mentioned ML algorithms and the larger database used for their training will allow extending the proposed methodology to a wider applicability range of input parameters as compared to standard accessible techniques. The method for pressure drop prediction based on ML algorithms trained on lab data is also validated on three real field cases. Validation indicates that the proposed model yields the following coefficients of determination: R2=0.806,0.815 and 0.99 as compared with the highest values obtained by commonly used techniques: R2=0.82 (Beggs and Brill correlation), R2=0.823 (Mukherjee and Brill correlation) and R2=0.98 (Beggs and Brill correlation). Hence, the method for calculating the pressure distribution could give comparable or even higher scores on field data by contrast to correlations and mechanistic models. This fact is an indicator that the model can be scalable from the lab to the field conditions without any additional retraining of ML algorithms.