A Time Series Characterization of IGBT Junction Temperature Method Based on LSTM Network

Accurate junction temperature characterization plays a critical role in the performance as well as reliability enhancement design of IGBT modules. However, most current methods are only focused on predicting the temperature at a specific time rather than tracking its temporal variation. To address this challenging issue, we propose a long short term memory (LSTM) neural network-based method for predicting the junction temperature time series of IGBTs, and it can comprehensively characterize the time-dependent junction temperature according to the measured and validated experimental results. As the convolutional neural network (CNN) is limited in its ability to remember temporal information, we integrate the LSTM networks for training, which can fast capture long-time dependencies in the data. Therefore, we exploit the superior performance of LSTM network for predicting the junction temperature of IGBTs here. It is shown that there is a strong correlation between the maximum junction temperature of IGBT and its parameters, such as bus voltage, load current, carrier, output frequencies, as well as its heat sink parameters. By leveraging the LSTM-based method, an accurate time series information of the junction temperature is successfully captured under different operating conditions. Notably, in comparison with the current machine learning method, the proposed method can outperform in the same training dataset.