An artificial-neural-network based prediction of heat transfer behaviors for in-tube supercritical CO₂ flow

Supercritical CO₂ flowing in tubes has been researched intensively to confirm their potential applications in energy conversion systems. However its performance prediction is always damaged by the non-linear variable thermo-physical properties and conventional correlation methods. To this context, this paper considers GA-BP model performed in MATLAB software to improve prediction accuracy. Firstly, three heat-transfer regimes are defined and parametrical effects are evaluated. Then, an architecture of 2-5-5-1 network is well-trained for accurately redistributed reconstruction of non-linear density of sCO₂ and achieves a reasonable root-mean-square error of 1.112 kg/m³ and a regression coefficient of 0.99. Finally, using 5895 sets of reliable experimental data with a verified architecture of 5-150-1 network makes the ANN model more adaptive and precise in interval predictions of heat-transfer behaviors, with a mean-absolute-percent error and as root-mean-square error far less than 2.97% and 3.11 °C, respectively. Results also suggest that the ANN model is technically superior to those correlations by I. Pioro, T. Preda, H. Kim, J.D. Jackson and Bringer-Smith under established test 1–5 groups. This study can provide a methodological guidance and shed a new insight for the further prediction of heat-transfer problems with supercritical fluids.