Hybrid Model Temperature Field Prediction Method Based on Dynamic Mode Decomposition and Deep Learning for IGBT Modules

The detailed temperature field distribution of high-power Insulated Gate Bipolar Transistor (IGBT) modules is important information for the reliability analysis and thermal design of power electronic systems and it is difficult to obtain quickly. This paper proposes a hybrid model for temperature field prediction based on Dynamic Mode Decomposition (DMD) and deep learning for IGBT modules. Firstly, the IGBT temperature field (ITF) snapshot is obtained through finite element simulation. Secondly, the DMD is used to extract stable and unstable trends of the ITF snapshot, and the future stable trend is predicted by recursion. Thirdly, a deep learning model autoencoder-long short-term memory (AE-LSTM) is proposed to predict the future unstable trend. Finally, the ITF prediction snapshots are obtained by adding the future stable and unstable trends. The proposed hybrid model prediction method realizes precise ITF prediction and significantly reduces the computation time. Experimental and simulation results validate the viability of the proposed method.