A Breakthrough Approach for Predicting ESP Wells Virtual Flow Rate by Using Supervised Machine Learning Method

Abstract

For field that deployed clusters of Electrical Submersible Pump (ESP) wells, implementing a robust surveillance system that could serve as early warning detection, real-time monitoring, and optimizing could give significant advantages. In establishing that system, real-time fluid rate becomes one important aspect but it cannot be obtained yet due to the limited frequency of conventional well test. However, ESP wells which have downhole sensor give valuable benefit to engineers as downhole sensor generates non-stop streaming data that represent ESP condition. Therefore, an idea to convert this nonstop streaming data into real-time fluid rate which could serve as an alternative to the conventional well test arises. An innovation that is discussed in this study is proposed to predict a virtual flow rate by utilizing the collection of data from ESP real-time sensor and wells information which simultaneously train and run a selected machine learning model.

In this study, the dataset collected from formation layers, ESP specification, tubing property, ESP realtime sensor, wellhead pressure, casing pressure, and historical well test data have been cleaned up before it is used to train model and predict the result once it is deployed. Afterwards, feature engineering is conducted to reduce the dimensionally of data. With the value of R-squared as indicator, six regression models comprised of K-Nearest Neighbor (KNN), Support Vector Machine Regression (SVR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGBoost), Linear Regression, and Elastic Net are compared to choose the best predictive model after parameter optimization for each model is applied.

This study used 14,915 data points from 12 mature wells in the Offshore Southeast Sumatera field to train and test the model. The sensitivity study done yielded SVR with the penalty parameters (C) value of 1000 and gamma (γ) value of 0.1 as the best algorithm and parameters for this case. The model reaches 96.05% level of accuracy when it is evaluated with 176 point of historical roduction test data. This study also shows that the model has succeeded to estimate the value of unknown fluid rate when the wells are not being tested.

The novelty of this paper is associated with the application of new machine learning model that can estimate ESP wells virtual flow rate in Offshore Southeast Sumatera. This study also shows the importance of data preparation, parameter optimization and feature engineering in achieving the proper model for prediction. Post-deployment, the model must be continuously updated its data especially when it is unable to approximate fluid rate properly.