Real-Time ESP Management Framework Using Hybrid Physics-Based and ML Models

Abstract Water injection is a crucial field development strategy to meet targeted production requirements and improve the oil recovery from carbonate reservoirs in the Middle East. Electrical Submersible Pump (ESP) failures cause disruptions both in the producers and water supply wells, leading to an increase in downtime and a decrease in well availability. This study introduces hybrid physics-based and ML models to identify and predict various specific failure modes, enabling proactive maintenance and improved system reliability. ESP datasets were collected from surface and downhole monitoring equipment over 5 years from 120 wells from producers and water source wells. The datasets, including high-frequency pump data, production data, failure cause, downtime, and reservoir information, are applied to design the real-time failure prediction framework. 13 key parameters were implemented as the foundation layer for the water injection system consisting of oil production wells and water source wells. Descriptive analytics is the next layer for using data mining methods to provide insight into past failure root cause analysis. Hybrid physics-based and ML models were implemented to predict the potential ESP failures in the water injection system as the predictive analytics layer. Most ESP operational failures are characterized as electrical failures and pump failures. The proposed integrated prediction framework evaluated 18 water supply wells and 40 producers, and the model successfully predicted 40% of historical failures. Moreover, if the solutions had been deployed in the real-time system and could have forecast failures 15 minutes to 30 days before actual failures. The limitations in ESP real-time data availability and dismantling, inspection, and failure analysis report issues impacted model accuracy. Also, several fault types were inherently unpredictable. The proposed ESP failure prediction framework enhanced the water source wells availability by 25% and the oil production well availability by 10%, significantly increasing water injection system capacity. These physics-based and ML models will assist operators in avoiding undesirable events, reducing downtime, and extending the lifespan for several specific ESP failure types. The presented framework integrates essential elements of ESP surveillance and prediction analysis into fully digitized intelligent software for water injection systems, allowing engineers to monitor early signs, diagnose potential causes, and take preventive actions.