Application of Artificial Neural Networks in Predicting Discharge Pressures of Electrical Submersible Pumps for Performance Optimization and Failure Prevention

The pump discharge pressure is a critical parameter that indicates the performance of the electrical submersible pumps (ESPs). Predicting the discharge pressure accurately can help in optimizing the ESP's performance, improving well productivity, and reducing operational costs. This paper presents a novel approach using artificial neural networks (ANNs) to predict the discharge pressure of electrical submersible pumps. The proposed model will enable early detection of possible failures and reduce downtime. Also, the effectiveness of the ANN model will be compared against the performance of different ANN models under various conditions. In this study, a dataset of more than 12000 data points collected from 40 different wells was used to train and test various ANN models with different input parameters. The performance of ANN models was evaluated using a coefficient of determination (R²), root mean squared error(RMSE), average absolute deviation(AAD), and absolute percentage error (APE). The model inputs and ANN structure were adjusted in order to minimize the prediction errors. The results during the training and testing phases were compared to select the most accurate and efficient model. Finally, the performance of the selected model was evaluated using physical analysis of the results and error profile visualization. The results showed that the ANN model with 16 inputs and 1 hidden layer with seven neurons is the most accurate model, with an R² of 0.95 for training data and 0.94 for testing data, and an AAPE of 1.74 for training data and 1.84 for testing data. The model was able to accurately predict the discharge pressure of ESP under different operating conditions, with an average accuracy of 94%. In addition, anomaly detection was also performed on the predicted values to identify any failures or anomalies in the system. This helps in proactive maintenance and troubleshooting of the system to prevent any significant failures. Finally, a new equation was developed utilizing the optimized ANN model. The developed equation provides fast and reliable estimations for the ESP discharge pressure with an error of less than 3%. Overall, the proposed approach provides novel and additive information to the existing literature by demonstrating the effectiveness of using ANNs to predict the discharge pressure of ESPs. Furthermore, the ability to accurately predict discharge pressure can lead to the early detection of possible anomalies, which can prevent costly failures and reduce downtime. Future work may explore the use of other AI techniques to further improve the accuracy of discharge pressure predictions and ESP failure prevention and explore the prediction of other important parameters in ESPs.