Machine Learning based Reduced Order Model for Full Temperature Prediction of Power Electronics Component Cooling

Power electronics are an essential part of an electrified vehicle (EV). Thus, it’s thermal management and temperature prediction is important to avoid any failure. For better functionality, new active thermal control strategies are under development for predictive health management of power electronics components. Model predictive control is getting more and more important as active sensing during operations is hard to establish because of packaging and costs. New model predictive control strategies rely on digital twins which can predict the electrical and thermal state of the component. Thermal digital twins right now just can predict pointwise temperatures, like the junction temperature of the dies, and are not able to predict the full temperature field of critical regions. This paper uses the machine learning methods to develop a new reduced order model (ROM) approach to be able to efficiently predict the full temperature field of power electronics component cooling system. The approach is demonstrated for a simplified use case of a powertrain inverter. The ROM is based on the different combinations of reduction and approximation methods which is applied on the simplified case of the electric vehicle inverter. The reduction methods, namely proper orthogonal decomposition (POD) or auto encoders (AE), writes the whole evolution of the data in a lower dimension while the approximation methods, i.e. the regression/interpolation methods, finds the optimal coefficient to correlate the lower dimension with higher one. Error analysis had been perform to examine the combinations of reduction and approximation methods to develop a ROM to compute the results fast in a reduced parametric space while retaining it’s accuracy. The resultant model was capable of predicting the temperature field of solid and fluid regions accurately within few minutes.