Methods and algorithms for power devices losses behavioral modeling

Abstract

Power electronics is since decades in the focus of very important technology innovations, as the characteristics and the performances of power supplies can severely condition and limit the performances of the system to be fed. In almost all the applications there is the demand to increase as much as possible the ratio between the maximum power the power supplies can deliver and their volume, defined as the power-density while, at the same time, the cost must be as reduced as possible. For this reason, electronic system designers have the task of finding, in a reasonable time, ever better performing solutions, choosing the best semiconductor devices and magnetic components. The attention of this thesis has been on the modeling of power losses of magnetic components and semiconductor devices, considering that they have the biggest impact on the system efficiency. The model classically adopted are usually calculated in different conditions from the operative conditions, require long time simulations, need the knowledge input variables that cannot be easily measured and have coefficients difficult to be identified. For this reason, the aim of this thesis has been to investigate a general approach to identify power losses models of devices, obtained from experimental data. In particular, sufficiently accurate and at the same time simple and intelligible loss model are desired. The approach adopted is based on Genetic Programming (GP), that is an evolutionary method able to return output models, in order to minimize a given fitness function that is a metric of the quality of the solution. The goal of the algorithm has been to obtain models accurate, but at the same time simple and intelligible for the user. These two desired conditions are often conflicting, being complicated models usually accurate and simple models usually inaccurate. For this reason, a Multi Objective (MO) approach, returning a Pareto Front composed non-dominated solutions, has been adopted. Moreover, the GP has been modified to return parametric functions, having the same structure, but different coefficients for all the devices to characterize. In this way, it is supposed to have a more general model, that is sufficiently good for all the devices. ... [edited by Author]