On the Design of NARX-ANNs for the Black-Box Modeling of Power Electronic Converters

Artificial intelligence (AI) has proven to be a game-changer in various fields, even in power electronics, where there is a lack of methodology for its use in the black-box modeling of power electronics dynamic systems. This work suggests a possible workflow for developing a non-linear autoregressive with exogenous input artificial neural network (NARX-ANN) for dynamic modeling of power electronic converters. The proposed workflow deals with the design aspects and the optimization process of the ANN, ensuring adequate accuracy while keeping a low ANN complexity. The design process starts from a given architecture, and the network is optimized step by step until it converges into three final architectures. A boost converter is chosen as a test-case considering, as the principal source of non-linearities, the different behavior in discontinuous and continuous conduction modes (DCM, CCM). A NARX-ANN model is developed and demonstrated its accuracy to replicate the converter behavior in time and for the small-signal analysis even with experimental data.