témavezető: Fodor Dénes
helyszín (magyar oldal): Széchenyi István Egyetem helyszín rövidítés: SZE
A kutatási téma leírása:
With increasing computing power becoming ubiquitous, the need for increasingly performant power converters is huge: the maximally supported currents are exploding and several thousands of Amperes are to be expected within the next years. At the same time the converter needs to be efficient for these very high currents (e.g. during complex computing processes) and also very low currents (e.g. while the processor is idling). Lastly, very fast transitions need between low and high currents and/or required changes in the output voltage need to be supported. Lastly, the converter needs to track the required output target voltage and provide a stable voltage response for a wide range of operation conditions (e.g. temperatures, load profiles and alike). Recently, new technologies (e.g. GaN or SiC) are being used in order to address some of these topics. In parallel, new topologies have been developed. The basic control schemes, based on PID regulation, have remained large unchanged. While the properties of PID regulations are well understood, it becomes more and more apparent that the nonlinearities in the converter cannot be addressed properly by linear control. Approaches like gain scheduling are the result. Looking at applications which feature a highly nonlinear plant (e.g. chemical or nuclear reactors) and the need for high performant control, the de facto standard of control has become model predictive control. Here, a model of the plant is combined with an optimization approach to determine the best possible regulation strategy. The regulation quality can be extremely good even for highly nonlinear systems, but it is also computationally very expensive and therefore cannot be easily applied to problems with very fast dynamics, like power converters. In recent publications, the structural similarities between model predictive control and reinforcement learning, a training technique well known in the field of AI, is highlighted. These approaches may combine the benefits of model predictive control over PID with a computational effort that still may be manageable. (Commonly with industrial partner: Infineon)