Artificial Intelligence Applications for Predicting Discharge Pressure and Flowrate of Electrical Submersible Pumps for Performance Optimization and Failure Prevention

The natural production of most oil wells inevitably declines over time, therefor, artificial lift methods are essential to maintain oil production. Among these methods, electrical submersible pumps (ESPs) are reliable and demonstrate a high level of productivity. However, ESPs are subjected to degradation or failure during their run life, leading to production losses. This study utilizes artificial intelligence (AI) to optimize the performance of ESPs using real field data from an Arab oil field. The study focuses on dual-parameter prediction for Pump Discharge Pressure (P_d) and Flow Rate (Q). Their impact on the run life of the ESP system was evident through trend analysis when compared to other features. The dataset for this study-comprising over 12,000 datapoints from 44 oil wells- was collected using real-time monitoring and includes 18 parameters such as surface data, reservoir and fluid properties, and pump specifications. During pre-processing, the Pearson Correlation Coefficient (PCC) indicated data non-linearity. Consequently, Support Vector Regression (SVR) and Extreme Gradient Boosting (XGBoost) were employed due to their effectiveness in managing such complexities. Additionally, a Random Forest model was utilized to predict ESP system failures. The performance of each model was evaluated using key metrics, including the coefficient of determination (R²⁾ Root Mean Squared Error (RMSE), and Average Absolute Percentage Error (AAPE). For the SVR model, the testing data yielded R² values of 0.955 for P_d and 0.917 for Q, with corresponding RMSE values of 76.2 and 157.2, and AAPE values of 1.01% for P_d and 5.92% for Q. In comparison, XGBoost demonstrated superior performance, achieving R² values of 0.968 for both P_d and Q, RMSE values of 64.16 and 98.16, and AAPE values of 0.84% for P_d and 3.48% for Q. The classification model successfully predicted failures during testing and was evaluated using precision, recall, and confusion matrix, achieving an overall accuracy of 96%. A comparative analysis of regression models was conducted, and XGBoost was identified as the most accurate model, delivering the best results. Ultimately, the developed models collectively aim to optimize company resources by predicting flowrates during recording challenges and by allocating personnel efforts to ESPs with a high risk of failure. In summary, traditional monitoring methods for ESPs (such as vibration analysis) are reactive and often detect problems only after failures occur. These limitations highlight the need for advanced monitoring solutions. Leveraging AI to optimize ESP performance is critical for ensuring sustainable and cost-effective production. The research findings can be applied across various ESPs, enabling proactive maintenance strategies that minimize downtime, reduce energy consumption, and lower operational costs. This study significantly contributes to the existing body of literature by providing a novel approach to optimizing ESP performance.