Reinforcement learning-based weighting factors auto-tuning for thermal management in assembled inverters

For three phase inverters, power semiconductor device MOSFET is the most sensitive component to failure which affects the reliability of the inverters directly. Through previous research on failure physical models, the junction temperature swing will mainly affect the remaining useful life of the MOSFET. Active thermal management is an effective way to control the temperature swing of control components. This paper proposes an optimized model predictive control for thermal management, which is achieved by suppressing overall system losses and focusing on vulnerable components, effectively suppressing the maximum junction temperature swing of the system. Afterwards, the reinforcement learning is used for weighting factors auto-tuning. The proposed method verified through experiment can effectively achieve balanced optimization of inverter life and performance and improve the reliability of the inverter system.