English abstract
The Electrical Submersible Pump (ESP) is the most efficient and reliable artificial lift
method for medium to high production rates. While the capital cost of ESP is high, it
pales in comparison to the production losses resulting from its failure. Machine learning
(ML) is a branch of artificial intelligence, and it has gained substantial attention in oil
and gas industries recently by virtue of its predictive power. This study aims to develop
an ML model to predict ESP lifespan and identify the key features that influence its
longevity.
A general review is conducted for the ESP failure analysis approach and the previously
applied predictive maintenance to monitor, troubleshoot and predict its failure (PFA).
The study explores various ML algorithms and their applications in predicting the time
and type of ESP failures in different fields worldwide. The aims of this thesis were
directed through ESP static and historical data to be achieved based on the literature
review. The failure history of more than 100 wells from an Omani oilfield was
reviewed, with 132 ESP failures attributed to sand and scale accumulation. The dataset
includes 36 static features related to ESP design, installation, commissioning, failure,
pull-out and teardown. Data were split into training and testing sets by a ratio of 70:30.
Hundreds of tests were performed to five selected algorithms in order to optimize the
parameters and hyperparameters of each algorithm, based on mean absolute error,
average residual and determination coefficient. The algorithms tested were Decision
Tree Regressor (DTR), Least Absolute Shrinkage and Selection Operator (LASSO),
Random Forest Regressor (RFR), Support Vector Regressor (SVR), and Extreme
Gradient Boosting Regressor (XGBR). Outliers were retrained if their predicted values
exceeded a predetermined limit.
The final model was developed in two levels to estimate ESP lifespan after installation
and after the last valid well test. SVR and RFR were chosen for the first and second
models, respectively. The average model predictions yielded a mean absolute error of
25 days for the first level with a 60% determination coefficient and 8 days for the second
level with a 73% determination coefficient. The model`s analysis indicated that certain
factors related to the pump and motor design have the greatest impact on ensuring a
long-lasting ESP before it is commissioned, based on the model's first level. These
include features such as pump discharge pressure, number of pump stages, motor
frequency, and shaft RPM, underscoring the critical role of careful selection and design
of ESP components. On the other hand, during the operational lifespan of the ESP, the
most important features to monitor and control are pump discharge pressure and flow
rates of oil and water, as highlighted by the model's second level.
By scheduling ESP maintenance before failure, these findings can help mitigate capital
costs, while preparing the necessary hoist, rig, and materials for ESP replacement can
avoid deferred operational costs.
