Details

Autor(en) / Beteiligte

• Giannetti, Niccolo
• Redo, Mark Anthony
• Sholahudin
• Jeong, Jongsoo
• Yamaguchi, Seiichi
• Saito, Kiyoshi
• Kim, Hyunyoung

Titel

Prediction of two-phase flow distribution in microchannel heat exchangers using artificial neural network

Ist Teil von

• International journal of refrigeration, 2020-03, Vol.111, p.53-62

Ort / Verlag

Paris: Elsevier Ltd

Erscheinungsjahr

2020

Quelle

Access via ScienceDirect (Elsevier)

Beschreibungen/Notizen

• •ANN is implemented for the prediction of two-phase refrigerant distribution.•The possibility for substantially improved prediction accuracy is demonstrated.•The optimised ANN model mostly achieves a ± 5% deviation without overfitting.•Reverse ANN could define optimal values of design parameters to achieve a target take-off-ratio. Due to the intrinsic complexity of two-phase flow distribution and the limited mathematical flexibility of conventional formulations of the phenomenon, previous attempts generally fall short in the accuracy and applicability of their prediction. To address these issues, this study focuses on methods with higher mathematical flexibility. Specifically, the construction and training of Artificial Neural Network (ANN) is presented for the identification of this complex phenomenon. The interaction of the numerous physical phenomena, occurring at different scales, is thus represented by the network structure, offering a formulation capable of achieving higher accuracy. Experimental data from a full-scale heat exchanger of an air-conditioning system operating over a wide range of conditions are used to train and test the ANN. The network optimisation with Bayesian regularisation against experimental data leads to a structure featuring 4 inputs, 3 hidden layers, and 3 neurons for each layer, which demonstrates deviations on the single output mostly lower than ± 10% and a correlation index higher than 98%, when the whole data set is used for training the ANN. The analysis of the network optimisation for different shares of data used for the network testing, shows higher training and testing accuracy as the number of training data increases, along with no apparent overfitting.