Prediction of the volume fraction of liquid-liquid two-phase flow in horizontal pipes using Long-Short Term Memory Networks

Abstract

This paper presents the development of a Long Short-Term Memory neural network designed to predict the volume fraction of liquid-liquid two-phase flows flowing through horizontal pipes. For this purpose, a comprehensive database was compiled using information sourced from existing research, comprising 2156 experimental data points utilized for model construction. The input of the algorithm consists of a vector containing the superficial velocities of the substances (oil and water), the mixture velocity, internal pipe diameter, and oil viscosity, while the output is the volume fraction of oil. Training and validation procedures involved preparing and segmenting the data, using 80% of the total information for training and the remaining 20% for validation. Model selection, based on performance evaluation, was conducted through 216 experiments. The predictive model with the best performance had a Mean Squared Error (MSE) of 3.5651E-05, a Mean Absolute Error (MAE) of 0.0045, and a Mean Absolute Percentage Error (MAPE) of 3.0250%. This performance was obtained by structuring the model with a ReLu transfer function, 20 epochs, a learning rate of 0.1, a sigmoid transfer function, a batch size of 1, ADAM optimizer, and 150 neurons in the hidden layer.

Keywords

• Artificial neural network
• LSTM
• machine learning
• two-phase flow
• volume fraction