Black-box modeling of DC-DC converters based on wavelet convolutional neural networks

Abstract

This paper presents an offline deep learning approach focused to model and identify a 270 V-to-28 V DC-DC step-down converter used in on-board distribution systems of more electric aircrafts (MEA). Manufacturers usually do not provide enough information of the converters. Thus, it is difficult to perform design and planning tasks and to check the behavior of the power distribution system without an accurate model. This work considers the converter as a black-box, and trains a wavelet convolutional neural network (WCNN) that is able of accurately reproducing the behavior of the DC-DC converter from a large set of experimental data. The methodology to design a WCNN based on the characteristics of the input and output signals of the converter is also described. The method is validated with experimental data obtained from a setup that replicates the 28 V on-board distribution system of an aircraft. The results presented in this paper show a high correlation between measured and estimated data, robustness and low computational burden. This paper also compares the proposed approach against other techniques presented in the literature. It is possible to extend this method to other DC-DC converters, depending on their requirements.