Publications

Electric Submersible Pumps (ESP) have been widely used in Sumatera Light Oil (SLO) operation units with more than 2500 production wells. Approximately 1200 wells are equipped with online devices that measure dynamic parameters such as ampere and voltage. Gross production data is required to analyze ESP performance and make production optimization decisions. However, the production data is obtained by acquiring fluid rate tests, which have a limited frequency of around once a month. Therefore, only relying on actual fluid rate measurement would lead to late actions, which cause the decisions to be sub-optimized. This paper demonstrates a Machine Learning (ML) approach to providing a virtual daily gross production rate using real-time ESP measurement and pumps information. In addition, the paper also describes the practical utilization of virtual tests to detect ESP problems. By using Random Forest (RF) as a final model, the Coefficient of Determination (R2) on the blind test is 0.98, and the Mean Absolute Percentage Error (MAPE) is 10%. The response time, which is a required time to replace the pump since it starts to deteriorate, became extremely faster, from 45 to 25 days. Finally, based on the calculation after one year of implementation, the financial benefit is equivalent to USD 1.5MM approximately.