Perpustakaan Digital ITB

Electric Submersible Pump (ESP) is one of artificial lift method that is commonly used in mature fields. Installation of ESP increases the cost of production. High economic return is expected from the installation of this artificial lift. In ESP operation, the run life of the ESP is critical. The asset team was able to benchmark equipment and vendors using this information. Understanding pump failure also can help operators prevent common causes of failure and extend the life of their wells. Operators recognize the necessity to achieve better ESP performance and longer run life to maximize economic value. The creation of a precise system for predict ESP run life and understand pump failure will enable proactive solutions such as reviewing operating conditions in order to increase the life of the pump, workover scheduling that is more efficient, reduce maintenance cost. This research show how we can choose the most optimum pump type of ESP and also predicting run life ESP using machine learning. This research use ESP installation data and production data in Offshore Southeast Sumatera. This is primarily achieved using Supervised Learning Technique; Decision Trees, K-Nearest Neighbors, XGBoost, Support Vector Machine, and Random Forest. From the prediction of ESP run life it will be know when the pump will fail. Then, predict liquid rate using Time-Series Analysis to obtain the desired liquid rate on downhole problem date to select ESP pump type. The selection of pump type is based on longest ESP run life and the highest pump efficiency. This study used 1242 clean data points from 285 wells in the Offshore Southeast Sumatera Field to train the model. This research uses Extreme Gradient Boosting to train the model and have 55.6% accuracy on test dataset with 5 cross-validation. The production rate data was subjected to a thorough investigation and data cleansing. Then, for forecasting purposes, a certain segment of the entire production rate data was chosen. The time series model was then trained using this section. The trained model was used to create a production rate in the future at downhole problem date. When compared the results achieved with actual production rate data, the Time Series model can anticipate rate with Root Mean Square Error less than Standard Deviation of the production rate data. After that, the forecasted rate can be used for ESP design and choose ESP. This program will help the workload of engineer to speed up data processing time, better decision making, improve the performance of ESP, and this program has potential for further development.