Fin-and-tube condenser performance modeling with neural network and response surface methodology

• Fin-and-tube condenser performance is evaluated using neural network.
• Response surface methodology helps collect small dataset for neural network training.
• High accuracy and low over-fitting risk is found in trained neural networks.

This paper presents a new approach of combining response surface methodology and neural network for performance evaluation of fin-and-tube air-cooled condensers which are widely used in refrigeration, air-conditioning and heat pump systems. Box–Behnken design (BBD) and Central Composite design (CCD) are applied to collect a small dataset for neural network training, respectively. It turns out that 41 sets of data are collected for heating capacity and refrigerant pressure drop, and 9 sets of data are collected for air pressure drop. Additional 2000+ sets of data are served as the test data. Compared with the test data, for the heating capacity, the average deviation (A.D.), standard deviation (S.D.) and coefficient of determination (R2) of trained neural network are −0.43%, 0.98% and 0.9996, respectively; for the refrigerant pressure drop, those are −2.09%, 4.98% and 0.996, respectively; and for the air pressure drop, those are 0.11%, 1.96% and 0.992, respectively. Classical quadratic polynomial response surface models were also included for reference. By comparison, the developed neural networks gave much better results. Moreover, the proposed method can remarkably downsize the neural network training dataset and mitigate the over-fitting risk.