Novel application of support vector machines to model the two phase boiling heat transfer coefficient in a vertical tube thermosiphon reboiler

Abstract

The prediction of the rate of heat transfer is one of the primary requirements for the design of thermosiphon reboilers. For the first time, a unified data-driven model for the prediction of boiling heat transfer coefficient in a thermosiphon reboiler, using support vector regression (SVR) as the modeling method, has been proposed. Different single component liquids with wide variation in thermophysical properties and operating parameters have been employed for the purpose. The SVR model for boiling heat transfer coefficient (in the form of Nusselt number) has been developed in terms of dimensionless parameters, namely PeB, Xtt, PrB and σwater/σL. 300 experimental runs from literature including the author's own work for the boiling of acetone, benzene, ethanol, ethyl acetate, ethylene glycol, propan-2-ol, toluene and distilled water have been utilized for the development and validation of the SVR model. The data was randomly divided into a training set and a test set in the ratio of 8:2. The former was used for training the SVM while the latter was used to test the generalization ability of the model. The optimal values of the SVR model hyper parameters viz., C and ɛ, and the RBF kernel parameter γ, were obtained by applying the grid search methodology with 10-fold cross-validation on the training data set. The values were: C=256, ɛ=0.145, and γ=0.35, respectively. An AARE of 8.43%, R of 0.9836, RMSE of 0.1103, SD of 0.0649, Q2LOO of 0.9672, and MRE of 0.0852 on the training data were obtained while the corresponding values for the test data were: 9.18% AARE, 0.9775 R, 0.1136 RMSE, 0.0581 SD, 0.9555 Q2ext, and 0.0946 MRE.The said SVR-based model has shown high prediction accuracy in terms of model evaluation indices, both for the training data as well as for the virgin test data. This model has scored highly over the other models from literature in terms of a remarkably enhanced prediction and generalization performance with an AARE of 8.43% on the training data and 9.18% on the test data.