![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
ABSTRAK Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
COVER Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
BAB 1 Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
BAB 2 Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
BAB 3 Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
BAB 4 Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
BAB 5 Muhammad Adhim Mulia
PUBLIC Suharsiyah ![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
DAFTAR Muhammad Adhim Mulia
PUBLIC Suharsiyah
A gas lift well requires precise tuning of its operating parameters to give its production
performance, which is a complex process. Conventional approaches encountered significant
challenges in resolving the well production prediction with numerous parameters and multiple
wells, which could give a considerable load to the production engineer. Machine Learning is
an application focused on improving programs that could learn from various operational
experiences. Implementing Machine Learning in gas lift operation to predict oil production
could provide significant working time reduction since they enable a simulation of the behavior
of a gas-lifted well using historical data without modeling the well in commercial software.
This paper presents a prediction of gas lift oil production by utilizing some machine learning
algorithms, which use the existing data from the Society of Petroleum Engineers (SPE)
repository as time series data. Four algorithms, Auto-Regressive Integrated Moving Averages
(ARIMAX), Vector Auto Regression (VAR), Long-Short-Term Memory (LSTM), and Prophet
were studied, and their results were compared to the actual gas lift well data and commercial
software results. The study shows that LSTM was the best model among the other three models
indicated by the lowest evaluation metrics: Mean Absolute Error (MAE) and Root Mean
Square Error (RMSE) of 36.33 and 54.36, respectively.
The average difference between the LSTM model result from the actual data was 4.8% while
the well model resulted in a 9.7% average difference from the actual data. This machine
learning model was then used to predict the oil production of another well in the range of the
first well oil production data, resulting in a relatively low average difference of 4% from the
actual produced oil data.