Deriving a thermal model from data by machine learning based on physical laws

Abstract

We propose a method for creating a reduced-order model (ROM) for temperature prediction that is applicable to nonlinear time-variant systems, able to be created using a very small amount of data, and easy to interpret. Our nonlinear time-variant ROM is an extension of sparse identification of nonlinear dynamical systems, which was first proposed in 2016. Three machine learning methods are developed for automatically deriving a thermal network model from time-series data. Link relationships between temperature nodes and parameter settings, such as thermal resistance and heat capacity, are automatically inferred by machine learning. The effectiveness of the proposed method is demonstrated using the results of thermo-fluid analysis for a natural air-cooled power module in which thermal resistance is a function of temperature. The ROM was created using the results of 15 thermo-fluid analyses, with is fewer than the 25 independent variables in the ROM. The computational efficiency of the created ROM is very high, and unknown data, including extrapolated data, were predicted with error of less than 1 K. By using the proposed method, it is expected that design parameters can be fine-tuned and actual loads can be taken into account, and also that condition-based maintenance can be realized through real-time simulation.