MACHINE-LEARNING ALGORITHM ASSIST DATA-DRIVEN EARLY DETECTION PROBLEMS ON CONTINUOUS GAS LIFT WELLS

• Main Author: Dwi Mustaqim, Sri
• Format: Final Project
• Language: Indonesia
• Subjects: Pertambangan dan operasi berkaitan
• Online Access: https://digilib.itb.ac.id/gdl/view/73298
• Institution: Institut Teknologi Bandung

The world's need for oil is increasing along with the world's population. On the other hand, oil production is decreasing over time, while exploration and development of new oil wells usually take at least 10 years. To increase oil production, several methods can be used based on the oil well's production phase. In the primary recovery phase, oil can be produced by natural flow and artificial lift. When oil cannot be produced by natural flow, then oil needs to be lifted to the surface by using artificial lift. Artificial lift is divided into two methods, namely the pumping method and the gas lift method. In the era of digitalization, the development of Machine Learning (ML) for the pumping method is a common thing such as ML on Sucker Rod Pump, Electric Submersible Pump, etc. However, the development of ML for the gas lift method is still rarely applied. Conventional monitoring and identification of problems in continuous gas lift wells takes a long time. This is because the engineers need to identify problems that might occur in the well after obtaining data from monitoring activities. Based on these conditions, it is necessary to develop machine learning to monitor the continuous operation of gas lift wells and predict well problems. In this study, ML is developed that can monitor gas lift systems based on digitalized two pen recorder chart in the form of casing pressure, tubing pressure, the pressure difference between casing pressure and tubing pressure, the pressure changes of casing pressure, and the pressure changes of tubing pressure and troubles The training data used for the ML algorithm is 70% of the total data and 30% of the total data is test data. The training data used consists of the digitalized two pen recorder chart data and the identified problem data. Moreover, various ML algorithms such as Gradient Boosted Tree (GBT), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM) are used to predict possible problems that occur in the current continuous gas lift well. After obtaining the ML algorithms with accuracy more than 90%, the ML algorithms were tested with 52 actual data to ensure the effectiveness of the designed ML. The performance evaluation of the ML algorithms was done by evaluating the value of accuracy, recall, precision, and F1-measure. According to the results of the performance evaluation, RF can be obtained as the ML algorithm with the smallest error to predict well problems in continuous gas lift wells with an accuracy of 97.984%. With the implementation of ML on continuous gas lift wells monitoring, the time needed to determine the possibility of problems that occur in continuous gas lift wells is reduced from days to hours. Therefore, this study is the answer to increase the effectiveness of oil production in continuous gas lift wells.