Machine learning based prediction of heat transfer deterioration of supercritical fluid in upward vertical tubes

Heat transfer deterioration (HTD) of supercritical fluid is critical to the safe operation of power systems. Current criteria of HTD have difficulties in the identification of working conditions of supercritical fluid. This paper proposed a new definition of HTD and a machine learning based method to predict the onset of HTD of su-percritical water and CO2 in upward vertical tubes. Eight criteria of HTD were compared and analyzed based on two traditional definitions and the proposed one. Results show that missing HTD points are inevitable due to the discontinuity of the experimental data. The accuracy of traditional methods falls in a range from 55% to 82%. In the two traditional definitions, selected cases used to fit the correlation are over-HTD instead of the exact onset of HTD, leading to the misjudgment of HTD cases as no-HTD cases frequently occur under traditional definitions. Machine learning based methods are used to solve the binary classification problem of HTD and no-HTD. Several machine learning models have been demonstrated to have a high prediction accuracy up to 95% for all the three definitions of HTD. The proposed definition of HTD and machine learning method can improve the identification and prediction of HTD of supercritical fluid.

Automated summary

The heat transfer deterioration (HTD) of supercritical fluids is crucial for the safe operation of power systems. Current criteria for HTD face difficulties in identifying the working conditions of supercritical fluids. This paper proposes a new definition of HTD and a machine learning-based method to predict HTD in upward vertical tubes with supercritical water and CO2. Eight criteria for HTD were compared and analyzed based on two traditional definitions and the proposed new one. Results showed that the discontinuity of experimental data leads to missing HTD points. Traditional methods achieved accuracy ranging from 55% to 82%. The machine learning-based method demonstrated high prediction accuracy of up to 95% for all three definitions of HTD. The proposed definition and the machine learning method can improve the identification and prediction of HTD in supercritical fluids.