A deep learning algorithm with smart-sized training data for transient thermal performance prediction

Abstract

There is an impulsion to introduce machine-learning algorithms into the thermo-fluids areas. Machine learning modeling, especially in high-precision regression problems, relies on large amounts of randomly-selected training data. Such randomness has little effect on final results because a big database (millions or billions of data points) can swallow such randomness. However, training data in the heat transfer area is not so huge and in fact it should be kept as small as possible because data acquisition is not convenient. Therefore, training data need to be shrunk without sacrificing prediction accuracy. Based on our previous work, we herein detail the selection process for a smart-sized training dataset. The smart-sized training dataset was slimmed by 50% with improved prediction accuracy. The prediction accuracy for the worst condition was still above 90%, an increase of 3.5% on previous results. This is a great improvement on our previous work because the training data scale was substantially reduced, demonstrating better results with half of the effort. The developed small-sample modeling using the deep learning algorithm provides a powerful information-processing tool to augment the understanding of the nonlinear systems in thermo-fluids mechanics, and to even transform current lines of thermo-fluids-related and more extensive applications.