Traditional DC-DC converter topology derivation methods typically depend on human expertise, which is largely learned from existing topologies. This reliance may restrict the diversity of the results due to the limitations inherent in human learning capacity. In this letter, artificial intelligence (AI) is employed to compensate for human learning capabilities, and a novel hybrid-learning topology derivation method is proposed. In the method, graph variational auto-encoders (GraphVAE) is utilized to learn topology generative rules from existing topologies directly. Besides, human expertise is further integrated into GraphVAE to refine these generative rulers, so that the proposed method can derive topologies by combining AI and human insights. To demonstrate the effectiveness of the proposed method, we applied it to the discovery of single-switch converters (SSCs). The GraphVAE model is trained by 135 existing SSCs and successfully derives 78 new SSC topologies with different performance characteristics. These results highlight the method’s ability to enhance both the efficiency and creativity in deriving topologies.
Categories