High pressure experiments and data-driven flow regime identification in gas-liquid two-phase pipeline-riser systems

An accurate identification of undesired flows in pipeline-riser systems is imperative for flow assurance in offshore oil fields. Gas-liquid two-phase flow experiments ranging from 1 bar to 50 bar in a pipeline-riser system were carried out and used as data sets to train a prediction model. A method based on the Levenberg-Marquardt algorithm using a back-propagation neural network and differential pressure signals was developed to identify the flow regimes. Prediction models trained using the statistical features of the signal at two locations clearly outperformed those trained at one single location. By comparing the recognition rate under different feature sets and sample lengths, a flow regime prediction model involving a sample duration of 20 s and an experimental database generated at standard pressure level was found best to identify flow regimes at higher pressures. The prediction model shows a good recognition rate for the most important unstable flow/severe slugging flow instability types, even under high pressures. This study serves as an exploratory investigation of flow regime identification in high-pressure pipeline-riser systems, with potential relevance for deepwater production scenarios.