Predictive Data Analytics for Effective Electric Submersible Pump Management

Abstract Electrical Submersible Pump (ESP) failures cause disruptions that lead to production deferement besides the cost of interventions/workovers post failure. The service life of an ESP is difficult to predict as it is affected by several factors which include reservoir characteristics, pump operating conditions and even the installation procedure. Measurement and monitoring of both dynamic and static ESP parameters play a critical role in extending the run-life. However, due to the complex nature of ESP failures, it can be difficult to identify anomalies by simply trending data. Notable progress has been made in the past years with respect to the development of systems for monitoring but most operators are yet to fully leverage on a system that will allow for proactive ESP health condition monitoring. In this paper, generalized machine-learning techniques and information acquired through real-time streaming was used to predict impending failures. This study applies Principal Component Analysis (PCA) on ESP intallations for a marginal field in the Niger Delta where production optimization and cost reduction are key to sustenance. Python was used for the data processing/statistical analysis and the algorithm development. The major objective of the PCA was to identify correlations in the dynamic ESP parameters: Intake Pressure, Intake Temperature, Discharge Pressure, Vibrations, Motor Temperature, Motor Current, Systems Current and Frequency recorded by the Variable Speed Drive (VSD) at regular intervals. Once the correlation/pattern was identified, the PCA approach found the directions of maximum variance in the high-dimensional data (in this study eight-dimensional) and projected it onto a smaller dimensional subspace while retaining most of the information. For each installation, a stable region for the operating frequency was identified and failed ESPs showed a clear drift from the stable region months before the failure occured, which was not apparent in the recorded parameters from the VSD. The paper describes how to use Machine Learning (ML) algorithms to predict an ESP's runlife bringing the industry a step closer to proactive ESP monitoring as opposed to the current reactive methods.