Abstract
Oil and gas wells often need artificial lift technologies to help extract reservoir fluids as the wells age and the reservoir pressure decreases–among those technologies is plunger lift. Plungers are metal cylinders that fit snug inside the tubing in which they are still able to slide freely. Plungers are often used with gas lift: they get pushed up and down the tubing like a piston to unload all the fluids by periodically shut in and open up the well. Many plungers contain a one-way valve that enables them to fall through flow easily and rise to the surface with a seal to prevent fluid slippage. There are many styles of plungers based on their weight, fall speed, and embedded one-way valves’ mechanism, and they should be used in different kinds of wells. Plungers can be grouped into five major different categories. This project utilizes machine learning and data analytics to predict what the most optimal type of plungers is for a given well in order to reduce gas injection and maximize liquid production.
To accomplish this project, more than two million rows of raw plunger lift production data were queried using SQL, a database query language, then pivoted, cleaned, and turned into a data table with approximately two hundred twenty thousand rows. The data came from about 900 wells with a time span of more than 400 days of production. Utilizing the python package Scikit Learn's random forest regressor, five separate machine learning models were trained, tuned, and cross-validated to predict the daily revenue and/or efficiency of a well if it were to run with the corresponding style of the plunger. The models were able to achieve about .85-.90 accuracy scores. On the other hand, a data visualization guide was built to visualize all the past plunger operation history to analyze the efficiency of a type of plunger; it also acts as an educational tool for operators to study the behaviors of different plungers in various wells.
The results of this project were achieved by field testing. Given that there was a time constraint, approximately ten wells were tested with eight of them showing revenue and/or efficiency improvements of 5-10%. More testing, tuning, and data gathering need to be done in the future to improve the project to apply at a larger scale.