Automatic Detection of Poorly Calibrated Models in State Estimation Applied to Oil and Gas Production Systems
MetadataVis full innførsel
In modern oil and gas industry, there is an increasing use of instrumentation. This lead to ahuge flow of information, which typically is not utilized to its full potential. By the use of increasingly more complex Virtual Flow Metering (VFM) solutions, the gap between theamount of data available and the amount of data utilized is reduced. VFM can contribute tooperational awareness and increased efficiency of the operations, which are qualities thatare becoming more and more important for the operators. The performance of a virtualflow meter is, however, highly correlated with the accuracy of the system models. Un-fortunately these system models are infrequently calibrated, and with increasingly morecomplex VFM models, these calibrations become more and more difficult to perform. Tofurther develop the field of VFM, this process of maintaining the models needs to be im-proved. In this thesis, a system that can potentially simplify this process is investigated. This thesis presents an investigation on the use of state of the art fault diagnosis tech-niques, to detect and identify poorly calibrated models used in virtual flow metering. Withthe help of the information gathered from the fault diagnosis, an operator can potentiallypinpoint when and where maintenance of the model is needed. If successful this can leadto a shift from recalibrating the entire system, to focusing on the parts of the model thathave been determined as weak links. By going straight for the weak link, the operator cansave substantial amounts of time and money, while the reliability of the system models issimultaneously increasing. The investigation has been conducted by running test scenarios on a simulator created during this thesis. The tests on this simulator were performed using state of the art virtual flow meters, together with the fault diagnosis tools developed in this thesis. The results show that this idea has good potential, and should be further investigated. Theresults showed this by successfully detecting and identifying poorly calibrated models forsimple test scenarios. That being said, both of the developed fault diagnosis systems, still have some drawbacks and unresolved issues, which makes them less suited for real applications. Several sugges-tions are, however, posted on how these issues can be resolved, and recommendations aregiven regarding the direction of future investigations.