| Abstract |
This paper presents a novel control approach using a Neural Network and reinforcement learning for Modular Multilevel Series Parallel Converters (MMSPCs). The MMSPC topology allows more _exible control than other multilevel converter topologies. This includes active balancing of temperature, minimizing of power loss due to internal resistance and current ripples and minimizing switching losses as well as harmonic distortion losses [1]. For ideal control, all of these parameters have to be weighted according to signi_cance, which changes during operation of the MMSPC. Due to multiple degrees of freedom which are mostly nonlinear, it is challenging to _nd the optimal scheduling strategy in conventional ways. This approach uses not only Neural Network (NN), but also reinforcement learning to optimize the controller in simulation and in an experimental setup. This leads to even more advantages such as compensating component tolerances and aging processes as well as faster computation time. |
