Machine learning algorithms for improving the prediction of air injection effect on the thermohydraulic performance of shell and tube heat exchanger

Abstract

In this study, improved prediction methods based on supervised machine-learning algorithms is proposed to predict the effect of the application of air injection and transverse baffles into shell and tube heat exchanger on the thermohydraulic performance. The injection process is accomplished by injecting air into the shell with different flow rates to obtain the optimal thermohydraulic performance. Four different machine-learning algorithms have been employed to predict the thermohydraulic performance of the heat exchanger to avoid mathematical modeling or carrying out costly experiments. These algorithms are random vector functional link, support vector machine, social media optimization, and k-nearest neighbors algorithm. The algorithms were trained and tested using experimental data. The inputs of the algorithms were the cold fluid and injected air volume flow rates; while the outputs were the outlet temperature of hot and cold fluids, in addition to pressure drop across the heat exchanger. The inlet temperatures of inlet hot and cold fluids and volume mass flow rate of hot fluid are considered as constants. The obtained results demonstrate the high ability of the random vector functional link model to find out the nonlinear relationship between the operating conditions and process responses. Moreover, it provides better prediction capabilities of the outlet temperature of hot and cold fluids and pressure drop values compared with the other three investigated models in terms of performance statistical measures. The root mean square error and mean relative error for RVFL results is approximately one-third and one-fourth of that of SMO, SVM, or k-NN, respectively. The root mean square error was, 0.719167, 2.477069, 1.741808, and 1.855635 for RVFL, SMO, SVM, and KNN, respectively; while mean relative error was 0.016167, 0.061746, 0.043366, and 0.041383 for RVFL, SMO, SVM, and k-NN, respectively.