A Case Study of Multiobjective Automatic Design for Active Neutral Point Clamped Inverter Based on Machine Learning

The traditional converter design follows the sequential process that usually contains design of topology, modulation, components, performances and their internal iterations. This method relies heavily on manual experience and are time-consuming, especially for multiobjective design (MOO) emphasized by using wide band gap devices that exacerbate the conflict between design objectives. Computer-aided virtual-prototype method, simplified as modeling and optimization, still hindered by long calculation time of numerical calculations in the modeling, and not responding to changed design requirements in the optimization. To address these challenges, the article proposes a premium concept of automatic design of power converters based on machine learning, combining ANN-based modeling and DRL-based optimization. The proposed method can not only reduce computational burden, improving various design objectives, e.g., efficiency, but also quickly obtain the optimal design variables with any design requirements, e.g., power level. The proposed method is validated in the active neutral point clamped (ANPC) inverter, and a case study of maximizing efficiency, minimizing component volume, cost and current ripple, while meeting the common-mode noise standard is presented. Finally, a prototype is built and the experimental results validate the proposed method. The method can significantly reduce human-dependence and develop further for automatic design of power converters.