Design and Implementation of a Digital Twin for Live Petroleum Production Optimization
a platform was used for Identification of events of interest by subject matter expert review . This method can be fast , effective and multipurpose . Data associated with the timestamps of the label can be easily extracted and utilized for further analysis , and supervised machine learning models can be trained for detecting such complex anomalies . 5 Figure 8 shows an image of such an expert reviewed label to identify abnormal time .
Fig . 8 : Labeling of abnormal events .
Flagging changes : To have a closed loop system for generating set-point change recommendations , it is imperative record historical and life set-point changes in the system to identify and evaluate the response to the stimulus . An illustration of recording flagging changes and their responses is presented in Figure 9 . The Gas Injection Rate displays the set point value , Oil and Water sensors represent the responses measures to a set-point change . In Figure 9 , the periods across set-point change indicated by greyed zones are marked such that tubing and casing pressures are operating in a stable state . Measuring the response of a set point change while the well is operating abnormally leads to incorrect evaluation .
Flagging changes has its similarities and differences when to identification of abnormalities . The difference is that abnormalities are usually not human controlled and are usually unintentional . The similarity is that both set-point change and abnormalities indicate a change in operating state . These can be identified through supervised methods such as labeling followed by machine learning models , and unsupervised methods such as measuring deviation beyond a threshold . The approach in this paper is a supervised “ human-in-the-loop ” approach , where set-point recommendations generated by the system are monitored and reviewed by the operator prior
5
Pennel , Mike , Hsiung , Jeffrey , and V . B . Putcha . " Detecting Failures and Optimizing Performance in Artificial Lift Using Machine Learning Models ." Paper presented at the SPE Western Regional Meeting , Garden Grove , California , USA , April 2018 . doi :
https :// doi . org / 10.2118 / 190090-MS
- 46 - March 2021