

2022 TA PP MOHAMMAD FADILLAH 1.pdf
Terbatas Suharsiyah
» ITB
Terbatas Suharsiyah
» ITB
Electrical Submersible Pumps are one of the most common and cost-effective secondary techniques in the industry (ESP). This technology can pump enormous volumes of fluid by decreasing the bottom hole pressure, allowing oil to flow from the reservoir, and adding energy from the pump and electric motor to bring the fluid to the surface. However, due to high gas volume, high temperature, and corrosive environments, ESP performance frequently degrades and reaches the point of service interruption without much warning. Maintenance of ESP is a highly capital-, resource- and labor-intensive task traditionally accomplished through reactive process monitoring of multivariate sensor data. The financial impact of an interruption in ESP service is substantial due to production losses and replacement costs. Consequently, developing technology capable of predicting ESP failures is crucial for the oil industry.
This case study utilized downhole sensor ESP data in real-time to develop an analytical method for detecting ESP failures. A slope-shaped classification will be performed on one-hour interval data for three-month (2,675 data) forecasting performance. Utilizing Long-Short-Term Memory to Forecast Performance. Long Short-Term Memory can model problems with multiple input parameters downhole sensor of ESP almost seamlessly. However, LSTM is typically a time series problem in Machine Learning. The crucial difference between time series and other machine learning problems is that the data samples in time series occur in a sequence. Therefore, LSTM can classify sequential data, but the prediction failure ESP dataset is not sequential. For this case study, classification using the Supervised Learning Technique was utilized (Decision Tree).
The downhole sensor ESP forecasting model using LSTM gets an average accuracy of over 90% and less than 10 errors%. It makes prediction failure ESP gets for long until 5 days. The models will be constructed based on the individual distinct parameter characteristics of nine statuses: closed valve, open choke, low PI, higher PI, increase in water, tubing leak, higher PI, increase in frequency, and sand ingestion, with a 95 percent accuracy rate. Besides that, to be more accurate with field conditions, this case study uses the latest matrix troubleshooting, which has 27 trips for failure prediction of ESP. Artificial Intelligence can be utilized as an effective technology in monitoring ESP systems. A human operator must constantly monitor these automation and control systems to ensure that all processes usually operate. In addition, abnormal behavior is identified in advance, allowing operators to quickly determine the best corrective action to avoid ESP failure based on the attached recommendations. Moreover, an oil company can generate billions in revenue from ESP failure prevention measures carried out by engineers.
Keywords: ESP, LSTM, Artificial Intelligence, machine learning, forecasting, early failure prediction