Digital Twin Models of Power Electronic Converters Using Dynamic Neural Networks
Andrew Wunderlich, Enrico Santi
Abstract
This paper presents a new approach for creating real-time models of power electronic converters using dynamic neural networks. The models are time-domain, switch-averaged, large-signal, real-time, and embeddable. This combination of characteristics renders the model uniquely suited to create digital twins of the converter which can run on any platform, including locally on the converter’s digital controller. The training dataset for the proposed neural network model can be generated from simulation (rather than measured on a hardware prototype) for convenience. This work will suggest a rationale for choosing the neural network structure and hyperparameters, will discuss best practices for training and testing the model, and will include an analysis of model error. The proposed modeling approach is compared and contrasted with other real-time modeling approaches and shown to be superior to previously proposed machine learning approaches.