Gustavo A. Carvalho, Eduardo Pereyra and Cem Sarica, University of Tulsa
Raphael Viggiano, Baker Hughes
Mike Micozzi and Wrangler Pankrantz, Ovintiv
The plunger lift process can be divided into four distinct cycles: buildup, upstroke, after flow, and liquid discharge. One key parameter that can be measured for optimizing oil production is the total gas flow rate produced during the liquid discharge cycle. Typically, the only known parameters are the controller’s on and off time, so post-processing is required to identify the liquid discharge period and quantify the observed flow rate.
Human analysis is enough to identify when the liquid discharge happens, which is characterized by the sudden increase in the gas flow rate. Analyzing one single well is feasible, however, the evaluation of tens or hundreds of wells becomes an unfeasible task.
This work proposes a machine-learning approach based on neural networks to automatically split plunger lift cycles. The model employs a long-short term memory (LSTM) neural-network, commonly used for time series data, with a classification head to identify and classify each stage. The model’s input is a time window containing casing, flowline, and tubing pressures, along with gas flow rate data; its output consists of probabilities corresponding to each plunger cycle. After the cycle automatic splitting, the cumulative gas flow rate produced during the liquid discharge period is quantified and recorded.
To train the model, field data must be acquired and manually labeled by a subject-matter expert. To automatize this part, a graphical-user interface (GUI) was developed to load well data and interactively select the correspondent plunger stage. The model was trained using data from five different wells and tested on a different well, achieving an accuracy of 98% for the cycle’s prediction.
This study presents an efficient and automated method to address a common challenge in production monitoring - quantifying well performance. Once trained, the proposed neural network can rapidly classify real-time data, enabling improved troubleshooting, production optimization, and performance tracking.