Interpretable AI for Power Conversion Control

Control approaches based on artificial intelligence (AI) have recently received attention in academic research. Yet, they are hardly used in power conversion. This lack of adoption could be explained by the fact that neural network decisions can often not be understood by humans. Interpretable AI models provide this transparency, but they are much harder to design: our approach is to reformulate the proportional integral differential (PID) regulation as a neural network first and then add several extensions based on typical neural network architectures. This way, the algorithm stays comprehensible no matter how the parameter values are selected. A learning approach is then proposed replacing the time-consuming manual parameter tuning. Finally, Simulation results of this new class of regulation schemes are provided for a single-phase buck converter.