Effective Management of Electric Submersible Pumps: A Predictive Failure Analytics Approach

Abstract Electric submersible pumps (ESPs) are considered one of the most widely used artificial lift methods. ESPs exhibit failures due to scale deposition, broken shafts and electrical causes. These failures leads to production interruptions that amount to thousands or millions of barrels of deferred oil production annually. The logistics of replacing an ESP is a complex and lengthy process and can affect well production and overall field economics. The ability to accurately detect the damaging events and anomalies that limit ESP run life is very helpful in making proactive operational decisions, workover scheduling and priotization, and reduce revenue loss. Although, ESP sensors, data collection and communication systems have improved in recent years, the industry still lacks a system that can monitor ESP health condition with the capability to accurately predict impending ESP failures. This paper presents a methodology using advanced machine learning to deploy neural network and extreme gradient boosting tree (XGBoost) in order to analyze real time sensor data to predict failures in ESPs. It provides a high degree of accuracy on historical ESP failures and contributes to an efficiency increase by cutting down on the time required to dismantle the pump system, inspect it and conduct failure analysis. The objective is achieved by applying the neural network model and its output is pipelined with an XG Boosting model for further prediction of the system status. This model is built on the previous ML approach used to detect emerging ESP failures, the Principal Component Analysis (PCA). PCA is used as an unsupervised Machine learning technique to identify hidden patterns and as a pre-processor for other algorithms. This PFA model is able to identify deeper functional relationship, dynamic changes and longer-term trends inferred from historical data with 80% precision and 60% recall on failures, and 15% false alarms. This paper describes how ESP monitoring and an enhanced predictive failure analytics approach increase production efficiency and eliminate deferments caused by ESP failures.