Session: 17-01-01 Research Posters

Paper Number: 76541

Start Time: Thursday, 02:25 PM

76541 - Machine Learning Application in Predicting the Boosting Pressure of Electrical Submersible Pumps (Esps) Under Various Flow Conditions 

Electrical Submersible Pump (ESP) is one of the most widely used artificial lift methods in the petroleum industry. With its compact structure and high efficiency, it has been applied extensively to various types of oil fields to lift moderate to high flow rates of hydrocarbon fluids. However, its performance and life span are highly sensitive to the flow conditions, including but not limited to viscous fluid flow, gas-liquid two-phase flow, oil-water emulsion flow, solid-contained slurry flow etc. Thus, the accurate prediction of the ESPs’ boosting pressure under varying flow conditions is important but very difficult. A universally validated mechanistic or empirical model to calculate ESP’s boosting pressure under multiple flow conditions is still unavailable in the literature. To better design and optimize the ESP production system, the ESP performance in terms of boosting pressure specifically is necessary. Traditionally, the pump performance can be predicted by the polynomial curve fitting, empirical equations, or mechanistic modeling. Nevertheless, these methods may suffer from request of massive experimental tests, lack of prediction accuracy, or requirement of too much pump geometrical information as the model input, which significantly limit their field applications. Moreover, as pointed out by previous research, the direct application of the existing empirical models to predict the ESP performance under complicated flow conditions is highly questionable.

 

In this study, experimental data of five ESPs under various flow conditions, including high viscosity flow and gas liquid flow, are collected from Tulsa University Artificial Lift Project. Then, the neural network (NN) machine learning algorithm is applied to predict the ESP performance under those flow conditions. Firstly, the modeling structure, activation function, optimizer, etc. are investigated. Secondly, the effect of training data quantity and quality on the NN model is analyzed. Finally, the result of the neural network is compared to other methods. In summary, a neural network requires few experimental data, but predicts pump curves more accurately, especially in gas-liquid flow conditions. Besides, compared to the traditional curve fitting method, it is simpler and faster to use the neural network to fit the pump curves. Since only five pumps were studied, the neural network still needs testing data for each pump to predict its performance in terms of hydraulic boosting pressure. It is suggested to conduct more tests or CFD simulations to extend the pump curve training database. In the future, there is a potential to predict the pump performance with plenty of data.

 

Presenting Author: Haiwen Zhu University of Tulsa

Authors:

Jianjun Zhu China University of Petroleum - Beijing
Haiwen Zhu University of Tulsa
Hong-Quan Zhang University of Tulsa