DEVELOPMENT OF SOFTWARE TO SELECT THE BEST SET OF CORRELATIONS FOR MULTIPHASE FLOW IN PRODUCING WELLS USING SUPPORT VECTOR MACHINE

Abstract

Abstract The prediction of flowing pressure gradient is crucial for many oil and gas field applications. Completion design, artificial lift systems optimization, and determination of surface and downhole equipment specifications are examples of applications that require accurate prediction of flowing pressure gradient in the wellbore. Over the years, significant research efforts were directed for developing predictive methods for multiphase flow flowing pressure gradient. The previous comparative studies performed between available predictive correlations and models showed that no single predictive method can be generalized for all flow conditions with reasonable accuracy. In absence of flowing pressure gradient surveys, it is difficult to select appropriate correlation(s) for a wide range of operating conditions. In this work, seven of the most implemented correlations and semi-mechanistic flowing pressure prediction models were chosen. The fully mechanistic models were investigated in a different research work. Modifications were introduced to the chosen flow regime transition rules suggested by the original authors. The original and modified correlations were programmed and coupled with a thermodynamic model to allow simultaneous prediction of flowing pressure and temperature. The prediction accuracy of the modified and original correlations was evaluated using a massive database with 4175 survey points from 1283 wells covering a wide range of well configurations and flow conditions. The modification introduced to predictive correlations improved overall prediction performance for some of them when tested against the large database. The Orkiszewski model showed improvement in terms of average absolute percent error from 13.4% to 11.7%. The results of this comparative study were used to create a new selection model for predictive correlations depending on flow conditions. The selection model was developed using support vector machine learning algorithm. The developed selection model was coupled with programmed correlations to create software that chooses the best correlation(s) to be implemented depending on well geometry and flow conditions. Two different support vector machine models were developed (e.g. SVM model and ECOC model). The developed selection models were tested against a test dataset (127 wells) that was not incorporated in its development. The SVM model showed a superior performance when compared to the ECOC model. The overall classification accuracy of the proposed SVM model was 75%. The software pressure prediction accuracy was validated against the test dataset and compared with other programmed correlations. The software yielded the minimum average absolute percent error 7.39%. Modified Duns and Ros correlation occupied the second place with 10.48% error. Thus, the proposed system can reduce the overall prediction error and improve flowing pressure traverse modeling by providing robust selection rules for multiphase flow correlations, which can be easily applied by petroleum engineers whenever the downhole flowing pressure traverse data is not available.