Machine Learning-Based Failure Prediction in Electric Submersible Pumps

Abstract

Electric Submersible Pumps (ESP) are among the most widely used artificial lift methods in oil production wells. Despite their effectiveness, ESP systems are prone to unexpected failures that lead to significant production losses, costly workover operations, and extended downtime. Traditional maintenance strategies, such as corrective or time-based maintenance, are often insufficient in predicting early signs of ESP degradation. This paper presents a machine learning–based approach for predicting potential ESP failures using historical SCADA operational data. The framework integrates statistical feature engineering via a sliding window, SMOTE-based class balancing, and a comparative evaluation of four algorithms: Random Forest, XGBoost (XGB), Extra Trees, and Logistic Regression. The results demonstrate that tree-based ensemble models achieve exceptional classification accuracy up to 99.96%, significantly outperforming the baseline Logistic Regression model (99.83%) and effectively identifying early indicators of ESP performance degradation with minimal misclassifications. Feature importance evaluation revealed that thermal and hydraulic parameters serve as the primary diagnostic anchors. Integrating this predictive framework into automated systems can successfully transition field operations from reactive to proactive maintenance, optimizing the economic lifecycle of ESP systems.