A modified semi-empirical correlation for designing two-phase separators

Abstract

Separators play a critical role in oil and gas industry by separating well outflows into natural gas, oil and/or water. A common method to design separators is using semi-empirical correlations. However, separator design is a complicated procedure and these correlations include several simplifying assumptions. Hence, evaluating their capability to design separators is necessary. In this study, a pilot two-phase separation unit is designed and constructed to examine the reliability of a selected correlation. The unit consists of a laboratory-scale horizontal two-phase separator, pumps and compressors to pressurize liquid and gas flows, a static mixer to create a two-phase flow, and a liquid filter. The experimental results from the unit show that the correlation could be modified to yield more accurate results. Therefore, including a correction factor into the correlation to reduce the error between the experimental and theoretical results is necessary. The experiments also show that the correction factor is not a constant value and depends on several parameters. Some of these parameters are gas flow rate, liquid density, liquid droplet diameter, drag coefficient, separator diameter, etc. Conducting experiments using the separation unit is time-consuming and costly. In addition, the sensitivity analysis required to develop a correlation between the correction factor and the parameters affecting it is not experimentally possible for some of the parameters. Thus, an artificial neural network is used to predict the correction factor and conduct sensitivity analysis. Because of the reasons mentioned above, obtaining enough experimental data to be used as input for the network is not possible. Therefore, computational fluid dynamics (CFD) simulation is chosen as the proxy model to generate more data for the neural network. The developed CFD model is validated with the experimental data (with less than 8% relative error) and enough data are then exported from it and added to the experimental data set to be used as input for the neural network. Finally, the network is used to predict the required correction factors. The feature parameters affecting the correction factors are also determined using sensitivity analysis. A new neural network developed using only those parameters predict the correction factors with even better accuracy. The semi-empirical equation modified by adding the correction factor can be implemented in separator design with high accuracy, leading to a considerable reduction in cost and time.