Sucker rod pump monitoring is essential to maintaining continuous production. Traditionally, pump abnormalities have been diagnosed by classifying downhole pump cards that are numerically calculated from surface measurements and rod string properties. However, even if the downhole pump card is classified as unhealthy, the pump may continue to operate normally. Given limited personnel and budget constraints, it is more practical to prioritize predicting and preventing actual pump failures rather than attempting to manage all frequent anomalies detected through downhole pump card classification.
In this study, we propose a real-time pump failure prediction method that utilizes deep learning and a new metric called scaled load factor, which is the ratio of the maximum and minimum normalized surface rod loads. The proposed method uses only surface pump load data to accurately predict pump failures and therefore does not require downhole pump card calculations. In predicting pump failures in two shale oil fields in the United States, the proposed method achieved an F1 score of 0.857 and was able to predict failures before they occur with an average lead time of 13.97 days.
This summary is taken from paper SPE 233386 by Y. Jung and Y. Kim, Seoul National University. B. Oh, Seoul National University. J. Jun, SK Innovation; W. Sun, China University of Petroleum (East China). and H. Jeong of Seoul National University. This paper has been peer-reviewed and is available as open access. SPE journal At One Petro.
