Application of artificial neural network to predict slug liquid holdup

This work demonstrates the artificial neural network (ANN) ability for predicting slug liquid holdup (H-LS), using 2525 measured points from 20 experimental studies. Six variables, including superficial gas velocity (V-SG), superficial liquid velocity, liquid viscosity, pipe diameter, pipe inclination (empty set), and surface tension, are selected as inputs to the ANN. The optimum ANN structure obtained is 6-11-1, with tangent sigmoid as an activation function. The developed ANN performs best and outperforms 12 existing H-LS models compared with present data and independent data. A sensitivity analysis shows that empty set has the lowest impact, whereas V-SG is the most significant variable on ANN-H-LS. To demonstrate the impact of ANN-H-LS on pressure drop in pipes, a mathematical model is derived and combined with the slug mechanistic models of Zhang et al. (2003) and Abdul-Majeed and Al-Mashat (2000). Based on Tulsa University Fluid Flow Project field measured data (1712 well cases), the new mathematical model's incorporation results in better predictions than using the individual H-LS correlations of these two models. The statistical results indicate that the slug model of Zhang et al. (2003) with ANN-H-LS and modified Barnea map gives the best performance compared to the existing pressure models.