Abstract：
Plunger lift has been widely used in unconventional gas wells to remove liquid accumulated at the well bottom.
Production surveillance provides large amount of data of production processes and abnormal operations, which can be used in
machine learning (ML) to develop algorithms for anomaly diagnosis and operation optimization. However, in the surveillance
data the majority is daily operation and the data of failure cases are rare. Also, the failure cases may not be repeatable, and many
failure case signatures are not available until they happen. Large data of anomaly cases are needed to improve the ML model
accuracy. Dynamic simulation of the plunger lift process offers an alternative way to generate synthetic data on the specified
anomalies to be used to train the ML model. It also helps better understand the trends reflected in the surveillance data and their
root causes.
From the available surveillance data of gas wells with plunger lift, the simultaneous measurements of different parameters
at different points in a production system with normal and abnormal occurrences can be analyzed and the corresponding trends/
signatures can be identified. The typical signatures that conform to pre-determined anomalous patterns can be obtained. Using
a commercial transient multiphase flow simulator, the actual field data of tubing/casing pressures can be matched through a
tuning process. Trial-and-error is needed to improve the dynamic plunger lift model so that a good agreement with the production
data can be achieved by adjusting the reservoir performance, plunger parameters or surface pipeline boundary conditions. After
validation under different flow conditions, synthetic datasets for various operational and flow conditions can be generated by
performing parametric studies. Unlike the field data, the synthetic data from the dynamic simulations mainly comprise anomaly
signatures (e.g. tubing rupture, missed arrival of plunger, etc.), which can be added to the ML data pool to reduce the data
covariance and increase independency.
The dynamic multiphase simulator OLGA has been applied to gas wells with artificial lift to simulate the parameter trends
in plunger lift systems under different flow and operational conditions. The preliminary comparison of simulation results against
field data shows good agreement in predicted tubing/casing/line pressure as well as production rates. Assuming an abnormal
occurrence, the parameter trends at different locations versus time can be characterized. Compared to the conventional data
munging techniques based on surveillance data only, the proposed data preparation method by generating synthetic datasets
from dynamic simulations is an efficient and economical solution towards better ML models to detect/predict the anomalies in
plunger-lift operations.
