Perpustakaan Digital ITB

Tubing leaks are one of the production problematics that is causing a decline in production and causing the company millions of dollars in workover cost. It is caused by corrosive environment and enhanced by erosion due to sand production. To mitigate this problem, a proper tubing material selection must be conducted. The tubing design needs to be just right, it cannot be underestimated because it may cause leaks and compromise the well barrier, but it cannot be overestimated because it can cause high capital and operational expenditure. The tubing design just needs to exceed the electrical submersible pump (ESP) run life to be economically effective. To evaluate the companys tubing design, a machine learning algorithm was created with several models to be able to classify whether the tubing design will be successful or failed. The model is built using production data, PVT data, tubing leak database, and well completion guide. There are 106 tubing leak cases, with 77 failed and 25 successful classes. The parameters are selected using a heatmap diagram to filter similar parameters. Using data preprocessing and synthetic minority oversampling technique (SMOTE), the data imbalance was processed to be split and trained. The machine learning algorithms that were developed are Random Forest, Support Vector Machine, Gradient Boosting Tree, AdaBoost, and Xtreme Gradient Boosting. From the selection of algorithms, XGBoost comes out superior with accuracy of 82.98%, precision of 90.91%, sensitivity recall of 76.92%, specificity of 76.92%, and F1 score of 83.33%. The model was implemented to create an application as a method that was easy to use. From the application, two wells that is used as a blind test, RMA-03 and IMA-23, are tested. The model was able to predict the durability result of the tubing design correctly. The well IMA-23 that is predicted to fail are redesign to mitigate tubing leak problem. The redesign well uses 13% Cr tubing with casing and completion string configuration similar to that of the actual well. The model created was able to classify the success or failure of the designed tubing that is more suitable for the field actual conditions. The implementation of the machine learning algorithm used a data driven process to learn and predict tubing durability. Thus, it can help to address the complexity of corrosion phenomenon that the semi empirical model was struggling to do. The further improvement of the data in the future will eliminate the possibilities of the model limititations such as the imbalance of dataset and confusion in the model prediction.