⚡ electrical engineering

Machine Learning Models for Forecasting Discharge Pressure and Flow Rate of Electrical Submersible Pumps for Performance Enhancement and Failure Prediction

This study demonstrates that machine learning models, particularly XGBoost for regression and Random Forest for classification, can effectively predict electrical submersible pump discharge pressure and flow rate while identifying system failures, thereby enabling proactive maintenance to reduce downtime and enhance operational efficiency.

Original authors: Omama Osman, Mohammed Eissa, Abdalla Abdalla, Amjed Hassan, Mahmoud Jumaa

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Omama Osman, Mohammed Eissa, Abdalla Abdalla, Amjed Hassan, Mahmoud Jumaa

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the earth, in the dark and heavy world of oil reservoirs, machines work tirelessly to bring crude oil to the surface. Among the most vital of these machines is the electrical submersible pump, a massive motor and impeller assembly lowered down a wellbore to push fluid upward against immense pressure. These pumps are the heart of modern oil production, keeping wells flowing when natural pressure fades. Yet, they operate in a brutal environment of high heat, corrosive fluids, and constant vibration. When a pump fails, the well stops producing, leading to significant financial losses and the need for expensive, disruptive repairs. For decades, engineers have monitored these pumps by watching for warning signs like unusual vibrations or changes in electrical current, but these methods are often reactive, identifying a problem only after damage has already begun. The question facing the industry has been whether it is possible to look ahead, to predict how a pump is performing and when it might fail before the machinery actually breaks.

A team of researchers from the University of Khartoum and King Fahd University of Petroleum and Minerals in Saudi Arabia has turned to a powerful tool to answer this question: machine learning. This branch of artificial intelligence allows computers to learn patterns from vast amounts of historical data without being explicitly programmed with specific rules. In this study, the researchers gathered a massive dataset from forty-four oil wells in the Middle East, collecting over twelve thousand records of real-time operations. They focused on two critical measurements that tell the story of a pump's health: the pressure at which the pump pushes fluid out, and the rate at which that fluid flows. By feeding this data into sophisticated computer models, the team aimed to create a system that could forecast these values with high precision and, more importantly, spot the early signs of a breakdown.

The researchers tested several different types of machine learning algorithms to see which one could best mimic the complex behavior of the pumps. They compared a method known as support vector regression, which is good at finding boundaries in data, against a more advanced technique called extreme gradient boosting, which builds a series of decision trees to refine its predictions step by step. They also developed a separate model specifically designed to classify whether a pump was operating normally or heading toward a failure. The data they used came from wells producing heavy oil, a thick, viscous fluid that places extra stress on equipment, making the task of prediction even more challenging. The wells were situated in carbonate rock formations, and the oil had an average gravity that classified it as heavy, meaning it flows less easily than lighter crude.

When the models were put to the test, the results were striking. The extreme gradient boosting model proved to be the most accurate predictor of the pump's performance. It could forecast the discharge pressure with an error margin of less than one percent and the flow rate with an error of just under four percent. In practical terms, this means the computer could predict exactly how much pressure the pump was generating and how much oil it was moving, matching the real-world measurements almost perfectly. The other models performed well, but this specific algorithm handled the messy, non-linear reality of the oil field data with superior consistency. It learned that factors like the temperature of the motor, the pressure at the pump's intake, and the size of the choke valve all interact in complex ways to determine the final output, and it captured these relationships better than the other methods.

Beyond predicting performance, the team also built a system to identify failures before they happened. They trained a classification model to recognize the subtle deviations in pressure and flow that signal a pump is about to trip or break down. The model was tested against seven common types of failures, ranging from electrical trips in the power supply to blockages in the tubing. By adjusting how the computer weighed the evidence, the researchers achieved a system that correctly identified failures ninety-six percent of the time. While the model did occasionally flag a healthy pump as faulty, it successfully caught the vast majority of actual problems, offering a window of time for operators to intervene. This approach shifts the strategy from reacting to a broken pump to proactively maintaining one, potentially saving wells from costly downtime.

The study demonstrates that the era of purely reactive maintenance for these critical machines may be coming to an end. By using real-time data and advanced algorithms, engineers can now anticipate how a pump will behave under varying conditions, such as changes in fluid viscosity or reservoir pressure. The research confirms that discharge pressure and flow rate are the most reliable indicators of a pump's condition, and that machine learning can track them with a level of precision that rivals physical sensors. This does not mean the pumps themselves have changed, but rather that the way we watch them has evolved. The ability to predict performance and failure with such accuracy offers a path toward more efficient operations, reduced energy consumption, and a more reliable supply of oil from the deep earth.

Technical Summary: Machine Learning Models for Forecasting Discharge Pressure and Flow Rate of Electrical Submersible Pumps

Problem Statement
Electrical Submersible Pumps (ESPs) are critical for sustaining oil production, particularly in high-rate wells and low-pressure, high-water-cut reservoirs. However, these systems operate under harsh conditions that lead to degradation and failure, resulting in significant production losses, increased workover frequency, and higher operational costs. Traditional monitoring methods, such as vibration analysis and current tracking, are largely reactive; they often identify faults only after failure has occurred or generate false alarms. There is a distinct need for proactive, data-driven solutions that can predict performance degradation and identify failures before they cause downtime. This study addresses the challenge of optimizing ESP performance and enabling predictive maintenance by forecasting two primary indicators of pump behavior: discharge pressure ( $P_d$ ) and flow rate ( $Q$ ).

Methodology
The research utilized a comprehensive dataset comprising approximately 12,000 data points from 44 oil wells in a Middle Eastern carbonate reservoir. The wells produced heavy oil (average API gravity of 19.2°) with varying bottom hole pressures (154–5,576 psi) and water cuts (0–100%). The study employed a structured workflow involving data collection, cleaning, exploratory data analysis (EDA), and model development.

Data Preprocessing: The initial dataset of 24,000 points was cleaned to remove water wells and wells with insufficient data. Missing values were handled using Forward Filling (LOCF) for regression tasks and Backward Filling for classification tasks. Outliers were managed via Winsorization (capping) or removal based on domain knowledge. The data was split into 80% training and 20% testing sets.
Feature Selection: Through trend analysis and Pearson Correlation Coefficient (PCC) evaluation, 18 input parameters were analyzed. Discharge pressure ( $P_d$ ) and flow rate ( $Q$ ) were selected as the primary target variables for regression. For failure classification, additional engineered features such as pressure differential ( $\Delta P$ ), thermal stress ( $\Delta T$ ), electrical power, and flow efficiency were incorporated.
Model Selection and Training:
- Regression: Support Vector Regression (SVR) and Extreme Gradient Boosting (XGBoost) were trained to predict $P_d$ and $Q$ . SVR required extensive hyperparameter tuning (GridSearchCV) and feature standardization. XGBoost was optimized with specific parameters (learning rate 0.3, max depth 6) and did not require feature scaling.
- Classification: A Random Forest classifier was developed to identify ESP system failures (e.g., OHTL trips, VSD trips, tubing blockage). The study addressed class imbalance using techniques like SMOTE, NearMiss, and bagging, ultimately selecting a model with an adjusted classification threshold of 0.75.

Key Results
The study evaluated model performance using $R^2$ , Root Mean Squared Error (RMSE), and Average Absolute Percentage Error (AAPE) for regression, and accuracy, precision, recall, and F1-score for classification.

Regression Performance: XGBoost consistently outperformed SVR. For discharge pressure ( $P_d$ ), XGBoost achieved an $R^2$ of 0.968, an RMSE of 64.16, and an AAPE of 0.84% on the testing set. For flow rate ( $Q$ ), XGBoost achieved an $R^2$ of 0.968, an RMSE of 98.16, and an AAPE of 3.48%. While SVR demonstrated good accuracy, XGBoost showed superior generalization and lower error rates, particularly for the more complex flow rate predictions.
Classification Performance: The optimized Random Forest model achieved an overall accuracy of 96%. By adjusting the classification threshold to 0.75, the model successfully predicted failures with a recall of 0.97 (detecting 60 out of 62 actual failures), though this came with a precision of 0.52 due to some false positives. The model effectively identified non-failure events with a low false-positive rate in specific configurations.

Significance and Contributions
The paper claims to offer an innovative, data-driven approach to enhance ESP operational reliability. The primary contributions include:

Proactive Maintenance: Unlike traditional reactive methods, the developed ML models enable the early identification of performance deviations and potential failures, facilitating condition-based maintenance.
Virtual Flow Metering: The high accuracy of the regression models suggests they can serve as a reliable substitute for physical flowmeters, reducing reliance on expensive production testing and equipment maintenance.
Operational Efficiency: By accurately forecasting $P_d$ and $Q$ , the models support real-time decision-making to reduce downtime, optimize energy usage, and improve cost-efficiency.
Robust Framework: The study validates the application of XGBoost and Random Forest in complex, non-linear ESP environments characterized by heavy oil and varying reservoir conditions, demonstrating their superiority over traditional linear approaches and SVR in this specific context.

The authors conclude that these models provide a reliable framework for improving ESP performance and minimizing operational risks in similar heavy oil production environments.

More like this