Bayesian Networks Applied to ESP Performance Monitoring and Forecasting

Abstract

Predicting the degradation of the Electrical Submersible Pumps (ESP) performance well in advance to avoid its failure is a rewarding yet challenging task. The complex interdependencies with well-reservoir properties and between machine components influence ESP performance and ultimately can lead to different types of failure. Predictions based on typical data-driven approaches are hampered by data quality and lack of variations in the operating conditions. In this study, Bayesian Networks (BNs), where expert knowledge can be incorporated and prediction uncertainty can be easily assessed, were tested as a potential method for performance monitoring and forecasting of a single ESP based on condition monitoring and production data. Different BNs based on auto-learned structures (relations being learned by the machine learning model) and expert specified structures were evaluated in the project. Network variables were selected from a dataset containing time-series sensors data (Pressure, Temperature, Frequency, Voltage, Current, etc.) for ESPs which degraded due to different reasons, including electrical failures and well conditions. The pump hydraulic efficiency, an indicator for the machine health, was used as a target variable. We report the results for a selected case where the ESP failed due to electrical failure (downhole ground fault). The hydraulic efficiency showed a noisy unsteady decreasing trend for several months before the failure. A sliding-window forecasting of the pump efficiency was performed with time horizons varying from few hours to several weeks. Based on only 4 sensors (current, two pressures, vibrations), the results between BNs with different structures were compared. The effects of adding additional variables (such as motor temperature or flowline pressure) to the network were also studied. A small user-defined BN was able to predict the pump hydraulic efficiency with an average absolute error ranging from 1.1% for the next 48h to 2.5% for the next 10 days. The novelty of this study is the application of BNs to ESP performance monitoring and forecasting. Within this framework, expert knowledge can be included by explicitly defining the causality between variables, suggesting a way to enhance data-driven methods. Since the prediction is probabilistic, the confidence in the predicted value can be straightforwardly assessed.