The main objective of this work is to identify and enable automatic online optimization actions in wells with plunger lift systems using machine learning.

For this purpose, a fault classification model has been developed for plunger lift systems using neural networks focused on production loss detection.

With the use of this tool, the aim is to reduce 10% of undetected production losses and increase production by 10%, improve the allocation of resources in the field, reduce the intervention of slickline equipment and improve the opportunities for multidisciplinary work between production engineers, data science specialists and automation specialists, so that an integrated solution can be obtained using data physics, prescriptive analysis and automation models.

The conceptualization of this proposal started by defining the different types of failure that are responsible for production losses or optimization opportunities applicable to wells with Plunger Lift. Time series techniques were used to compare and isolate failure patterns available in historical head pressure data. The patterns were subsequently converted to images and those images corresponding to failures later classified using neural networks. Finally, an interdisciplinary team of data scientists, production engineers, and optimizers, identified and proposed the actions to solve each group of failures.

3000 images corresponding to failures were collected and used to train the classifier, obtaining an accuracy of 80% in the identification of new events.

Due to the large number of wells in the field (around 400), this classifier has been built on the principle of wells performance control by handling through exception, as otherwise it would unpractical to analyze the totality of events occurring daily. Thus, this tool has proved to be of great help in reducing study times by focusing primary on those wells that are impairing production and later on those which might still be optimized, aiming at the end of the process achieving operational excellence by maximizing production across all wells to their expected potential.

The project also highlights YPF′s leading business approach by promoting the concept of multidisciplinary collaborative spaces by merging specialists from Experience Center Data Scientist and production Engineering Sector. This convergence provides a considerable improvement to an economical and profitable extraction system for unconventional wells.

The identification of failures in the Plunger Lift extraction system presented in this work consists of an extension of information that is already in the literature, however, the use of machine data driven models coupled with a classifier tool which optimizes actions using a risk approach based on production losses to identify and solve these types of failures is considered a novel instrument. Also, this work is the initial part of a slickline schedule automation system.

This content is only available via PDF.
You can access this article if you purchase or spend a download.