RT22 | Model Predictive Control and Neural Networks Tools Applied to Multilevel Converters

Standard control strategies for modular multilevel converters are mainly based on multiple loops of proportional-integral controllers working in multiple transformed references to achieve decoupling. However, the use of multiple SISO controllers for a highly coupled system, which has more variable than degrees of freedom, reduces the robutness of the control in comparison with MIMO controllers. This work proposes a sequential phase-shifted model predictive control (sPS-MPC) to govern modular multilevel matrix converter (M3C), which considers the coupling of the states and symplifies the control structure. The novelty and merit of applying this MPC in a M3C a new and non-trivial formulation of the M3C model required and a very particular dynamic and differences compared to other topologies, since it is connected to two ac-ports and presents a very particular behavior at different operating frequencies. Experimental results on a 36-cell M3C have been performed to verify the proposal performance. The predictive control scheme was implemented on an OPAL-RT OP4510 and OP4520 with a sampling time of 62 us. As evidenced by the esperimental results the sPS-MPC formulation results in a fast closed-loop dynamic response and effective tracking of the input, output and circulating currents. Additionally, the M3C governed by the proposed PS-MPC operates with a fixed switching frequency, low harmonic distortion, and distributed power losses among power switches. Sequential Phase-Shift MPC has also been applied to a DC-AC MMC, however, standard and MPC based control schemes use one voltage measurement per submodule to balance the capacitor voltages and achieve MMC control. Consequently, the control system requires a significant amount of sensors and the effective communication of sensitive data under relevant EMI conditions, impacting converter reliability and cost. In order to overcome these issues, a distributed neural network observer for estimating the capacitor voltages at each SM in real time was implemented, The observer predicts each SM capacitor voltage using a standard average model. Then, each prediction is corrected and denoised by a neural network of reduced computational complexity. As a result, the observer reduces the number of required voltage sensors per arm to only one and filters the high-frequency noise in the estimated SM capacitor voltages. The real-time neural network deployment was made in a OPAL-RT controller and experimentally validated on a MMC with 24 cells, scalability and computational cost on OP4510 simulator has also been evaluated.