Recurrent neural networks model based reliability assessment of power semiconductors in PMSG converter

Abstract To obtain accurate lifetime evaluation with acceptable simulation time for fulfilling the total life cycle design criteria, this paper proposes a Recurrent Neural Networks (RNN) based model with the replacement of the Simulink model. It starts with the establishment of the Averaged Switch (AS) model and Averaged Fundamental (AF) model of the Permanent Magnet Synchronous Generators (PMSG) to calculate accumulated damage. Then, under the same mission profile, the junction temperature, rainflow counting and accumulated damage of the AS and AF model are calculated and compared. It can be noted that the AS model is more accurate to calculate the reliability of components, since both the big thermal cycles caused by load variations and the small thermal cycles due to the fundamental AC current are considered. However, it consumes more time compared to the AF model. To this end, the RNN model is proposed to substitute the most time consuming part of the system reliability evaluation procedure. With aid of proposed model, the consumed time can be greatly reduced compared with the Simulink model. In the end, a 1-hour case study is applied to verify the efficiency of the RNN model. The Mean Absolute Percentage Error (MAPE) of the testing case is 0.51% and the time for getting results in the RNN model is less than 1 s. Besides, an annual case is also implemented to verify the RNN model, which has an averaged 0.78% MAPE for the whole year.