Development of a neural network model for the prediction of liquid holdup in two-phase horizontal flow

Abstract

Accurate prediction of liquid holdup associated with multiphase flow is a critical element in the design and operation of modern production systems. This prediction is complicated by the complexity in the distribution of the phases and the wide range of fluid properties encountered in production operations. Consequently, the performance of the numerous existing empirical correlations and the more recently developed mechanistic models has been inadequate in terms of desired accuracy and range of application. This investigation focuses on the development of neural network models, a relatively new approach that has been successfully applied to a variety of complex engineering problems. Data from five independent studies were used to develop two predictive models applicable to horizontal two-phase flow. One model is based on discrete parameters including flow rates, pipe diameter, and fluid properties, while the other model uses dimensionless numbers incorporating these parameters. A detailed comparison with existing empirical correlations (Beggs-Brill, Minami-Brill, Mukherjee-Brill, Abdul) and mechanistic models (Taitel-Dukler, Xiao) reveals that both neural network models show an improvement in overall accuracy and perform more consistently across the range of liquid holdup and flow patterns. Additionally, the neural network models do not require prior estimation of flow pattern, as do some of the other approaches. Therefore, the models are not affected by errors in flow pattern prediction and are not subject to discontinuities in predictions made across transition boundaries. Furthermore, the neural network models can be readily adapted to a broader range of flow variables with the introduction of additional training data.