Evaluating neural control with optimal architecture for DC/DC converter

Abstract

Controlling DC/DC converters (topologies widely used in the active reduction of harmonic content for single-phase nonlinear low-power equipment) raises great design challenges due to the mathematical model's complexity and its highly nonlinear dynamic characteristics. Artificial intelligence techniques, such as neuronal networks, suppose great improvements in design and final performance, given their capacity for learning complex dynamics and generalising their behaviour. This work was aimed at proposing (and evaluating dynamic response later on) direct control link with neuronal networks which also allowed eliminating test elements and error in its design. Artificial neuronal network-based direct control was designed as well as possible using bioinspired search models. This simultaneously optimized two different but fundamental aspects of the network: architecture and the weight of the connections. The control was applied to a boost converter. The results led to observing the scheme's dynamic performance; response time and exit voltage delta led to concluding that the criteria selected for designing the control were appropriate and represented a contribution towards developing control applications of DC/DC switchmode systems.