PREDICTION OF AVERAGE VOID FRACTION AND PDF OF VOID FRACTION IN VERTICAL 90O BEND FOR AIR–SILICONE OIL FLOW USING MULTILAYER PERCEPTRON (MLP) CODES

Abstract

Multilayer Perceptron (MLP) models have been developed to predict two- phase average void fraction and probability density function (PDF) of void fraction in 90o bends. The Artificial Neural Network (ANN) methodology was reported using MLP trained with 2 algorithms. Logarithmic sigmoid transfer function was used in a single hidden layer for both algorithms (Gradient descent (GDMV) and Levenberg-Marquardt (LM) algorithms). Both MLP models were optimised by varying the number of neurons in the hidden layer while monitoring the Mean Square Error (MSE). The performance of the models was evaluated using the Average Absolute Relative Error (AARE) and Cross Correlation Coefficient (R). Both MLP models developed for the prediction of average void faction before the bend performed excellently well. However, the MLP model trained with LM algorithm having 3 neurons in the hidden layer gave better performance. Similarly, the MLP model trained with LM algorithm, having 11 neurons in the hidden layer for the prediction of PDF of void fraction before the bend gave excellent prediction. Model performance for the MLP models after the bend gave poor generalisation property. However, the MLP model based on GDMV algorithm gave better prediction for predicting average void fraction and PDF of void fraction after the bend. It was concluded that MLP models may with some confidence be used to predict the average void fraction and the PDFs of void fraction observed before a vertical 90o bend.