A Reduced-Order Temperature Field Prediction Model for Power Devices Based on Proper Orthogonal Decomposition and Deep Learning

Abstract

With the continuous increase in the power density of power devices, the temperature prediction of power devices is becoming increasingly important for reliable operation and thermal management. This paper proposes a temperature field reduced-order prediction model for power devices based on proper orthogonal decomposition (POD) and deep learning. POD is employed to extract the main spatial modes of the temperature field in the sample snapshots obtained from finite element method (FEM) simulation. Long Short-Term Memory (LSTM) is employed to predict the time coefficients of spatial modes. The main spatial modes and the predicted time coefficients are used to construct the reduced-order prediction model. The accuracy and performance of the proposed model are evaluated by comparing it with FEM simulation. The perfect agreement between the results of the two methods confirms the accuracy of the proposed model, and the comparison of calculation time indicates that the proposed model can significantly improve computational efficiency.