Predicting Failure in Electric Submersible Pump by Utilizing Machine Learning Based on Real-Time Sensor Data

Abstrak

Abstract. Electric Submersible Pump has been used widely for oil wells around the world. The investment of ESP installation is costly, therefore the high return is expected from the implementation. This research has the objective to build a machine learning model to predict the possibilities of failure in electric submersible pump installation. By making an early prediction, the user can prepare to do the action needed to overcome the problem. Therefore, it will reduce the loss of fluid production.

The artificial intelligence will use data from a real-time sensor installed on the Electric Submersible Pump. There are six parameters that can be retrieved: ampere, intake pressure, discharge pressure, intake temperature, motor temperature, and vibration. The liquid rate then to be calculated using an analytical equation. The data are prepared by picking a set of data for every three hours from the raw data. This is decided because the interval of the data recorded by the sensor are not always at the same interval. The prepared data then to be transferred into a database. The evaluation will be conducted by developing the slope of the three days data and called as set data. Every data set then labelled according to the type of ESP’s failure. The method that is used to develop a proxy model came from comparing several methods, such as logistic regression, decision tree analysis, etc.

There are thousands of data set that already filtered from over 400,000 raw data. The data came from four types of pumps that are commonly used in the observed fields. The proxy model resulting in several kinds of events that can be read, such as low PI, pump wear, tubing leak, increase in frequency, increase in water cut, etc. Those events have individual distinct parameter’s characteristics. The model has an accuracy rate of over 70% with the proportion of 75% training set and 25% test set. The  implementation of this proxy model will tell the user the most possible event that will happen according to the latest data set.

The utilization of a machine learning to ease the work of a user will be the most efficient way in this era of digitalization. This is the first machine learning approach that is implemented to predict Electric Submersible Pump failure application, and the outcome is to be able to predict the problem in the early stage so the fluid loss due to the problem is decreased and more importantly help prevent the well to be shut-in.