Hybrid junction temperature prediction of IGBTs combining detailed electro-thermal modelling and deep learning regression

This paper focuses on predicting the junction temperature of Insulated Gate Bipolar Transistors (IGBTs), a key factor in preventing thermal failure and enhancing system reliability. Traditional RC-based thermal electrical models offer simplified approximations but fail to capture IGBT dynamic behavior during switching. To address this, the electro–thermal Hefner model is employed to generate a detailed dataset. The model was validated against PSIM simulations, showing strong agreement. It captures dynamic parameters such as depletion layer width, charge conductivity, collector-emitter and gate-source voltages, drain current, and switching energy losses. Using this dataset, several machine learning regression models are developed for temperature prediction: Random Forest (RF), Feedforward Neural Networks trained with Scaled Conjugate Gradient (SCG-FNN) and Adam optimizer (Adam-FNN), and a Recurrent Neural Network (RNN). Their performance is evaluated under two scenarios, constant duty cycle with rising temperature, and step changes in both duty cycle and temperature, using Root Mean Square Error (RMSE) and coefficient of determination (R²). Results show that RNN achieves the best accuracy (RMSE <0.5 K, R² ≈ 0.9), followed by Adam-FNN and SCG-FNN (RMSE 0.5–2 K, R² ≈ 0.8), while RF performs well only below 350 K. The combined physics-based and learning approach ensures robust junction temperature prediction.