Data-driven modeling of a forced convection system for super-real-time transient thermal performance prediction

Abstract

Abstract Transient thermal performances have always been of great interest in various thermal and energy engineering areas. Generally, there are three common methods to gain transient thermal performances: experimental, theoretical, and numerical. However, there are limitations in gaining the transient thermal performance using these three methods. This paper discloses a novel data-driven approach to attain the transient thermal data from certain operating cases (acquired from those three methods above) to predict the transient thermal performance of other cases. A forced convection system is selected as a demonstrating example. Transient thermal performances of 39 operating cases with various inlet velocities and heat loads are attained from numerical method first. Then, data of a randomly selected 32 operating cases falls into the training data set while data of other 7 cases into the testing set. With the training data set, a neural network model is trained and the trained network gives a high accuracy estimate for the testing set, where an averaged accuracy of 91.5% is obtained. Additionally, the processing time for predicting the transient performance from 0 to 300\xa0s can be reduced to 13\xa0s, suggesting a super-real-time prediction. The proposed approach is expected to model complex thermal-fluid system precisely as well.