ESP Data Analytics: Predicting Failures for Improved Production Performance

Abstract

Abstract Failures of Electric Submersible Pumps (ESPs) are common occurrences in the oil industry. The random nature of these failures results in production disruption that amounts to hundreds of millions of barrels of lost or deferred oil production annually. Although important improvements have been made in the last fifteen years on ESP sensors, data collection and communications systems, the industry still lacks a system that can provide ESP health condition monitoring with the capability to accurately predict impending ESP failures. The intent of this study was to evaluate the value of Principal Component Analysis (PCA) as a tool to detect developing ESP failures and predict remaining operating time before failure. Complete historical data of five ESP installations was used. For each installation, a stable region was selected to construct a PCA-based model, which was later used to find projections for the whole data set on the principal component axes. Different techniques were used to correlate projected data with the original data and to draw the conclusions of the study. The analysis showed that PCA data scattered away from the origin before the occurrence of each of the failures. But a more powerful observation was that PCA projections showed distinctive data clusters representing subtle changes in the system that are not apparent by directly examining the actual measured parameters. This study concluded that PCA has potential to be used as a tool to identify dynamic changes in the ESP system and therefore to detect developing ESP problems. PCA can also be used as unsupervised Machine Learning (ML) technique to identify hidden patterns and as an essential pre-processing technique for other ML algorithms. PCA can be used as the foundation for the development of better tools for detection of developing ESP failures and prediction remaining ESP run time.