Enhanced neural network-based polytopic model for large-signal black-box modeling of power electronic converters

We propose a large-signal black-box model of power electronic converters inspired by polytopic models. Small-signal models are identified around different operating points to mimic the converter's local dynamics. The linear models' responses are then weighted using a trained neural network to create a large-signal model. The traditional trial and error weighting function tuning of polytopic models can result in a suboptimal combination of linear models. In this work, we use neural networks to approach an optimal combination. The analysis of the trained neural network can enhance the model's accuracy by suggesting new small-signal models. It also permits removing linear models that do not significantly improve the global model's accuracy while reducing complexity. The methodology is applied to a voltage-regulated DC-DC boost converter and provides accurate models of converter dynamics.