Machine Learning Estimators for Power Electronics Design and Optimization

In the design and optimization of power electronics systems consisting of multiple converters, there is a need to characterize performance of a large number of available converter modules in order to determine the system performance characteristics such as efficiency or size. This characterization process can be computationally expensive, requiring multiple module-level design and optimization steps. Machine learning (ML) based estimators can be used to reduce the computational expense associated with dc-dc converter characterization. This paper compares ML techniques for generating dc-dc converter performance models and shows that random forest and gradient boosting techniques are capable of accurately predicting the performance of commercially available dc-dc converters. Test cases show that system optimization approaches using ML estimators can generate power electronics systems with Figure of Merit (FOM) expressed in terms of efficiency and size within 15% of systems designed using high fidelity but computationally expensive models.