Interpretable large-signal black-box models of power electronic converters: a multimodel approach

In this paper, we propose a multimodel approach to derive large-signal black-box models of power converters that are suitable for system-level studies. We introduce an interpretable black-box large-signal model by taking advantage of the inherent interpretability of linear system models and the design of an interpretable weighting function. First, the non-linear model of a power converter is approximated by a set of linear submodels. Next, we consider a neural network-based weighting function trained to combine the linear submodels' responses. A post-analysis of the trained neural network is used to speed up the partitioning of the operating space by restricting the number of new experiments that have to be carried out. A single-machine infinite bus system is used to illustrate the rationale behind the proposed multimodel approach and to illustrate some of its inherent limitations. The overall methodology is illustrated using a voltage-regulated DC-DC boost converter. Finally, the approach is validated using a small system including a battery, a voltage-regulated DC-DC boost converter, and a DC motor.