Abstract
Abstract
Multiphase flow modelling is of major importance in the design of pipelines, separation plants, and many other systems found in the chemical and petroleum industries. Many multiphase flow models apply a number of closure relationships; one of such is the drift velocity. Empirical correlations, with varying range of applicability and predictive capability, are typically relied upon by researchers to predict this parameter.
This work therefore presents the development of a machine learning approach for predicting drift velocity in horizontal and non-horizontal pipelines. Python programming language environment using Tensorflow and Keras deep learning libraries with early stopping to prevent overfitting and inaccuracy was implemented. The approach was based on a three-layered supervised deep neural network developed as a function of pipe diameter, pipe inclination, fluid density, fluid viscosity and surface tension.
The developed model was validated by comparing with experimental data and existing correlations by statistics and error analysis. The developed deep learning model, in general, performed better than the investigated existing correlations under their respective range of validity. Over 90 % of the calculated Froude numbers from the DNN model were within the fixed ±20 % average relative error bandwidth considered for this work.