AI-Driven Phase-Shifted Carrier Optimization for Cascaded Bridge Converters, Modular Multilevel Converters, and Reconfigurable Batteries

TL;DR

This paper introduces a neural network-based method to optimize phase-shift angles in power converters, significantly reducing ripple and distortion while enabling real-time, scalable, and efficient control compared to traditional optimization techniques.

Contribution

The paper presents a neural network approach that emulates an optimizer for phase-shift angles, offering real-time performance, scalability, and reduced computational complexity for power converter modulation.

Findings

Reduces current ripple and harmonic distortion by up to 50%.

Achieves 100 to 500 thousand times faster optimization than genetic algorithms.

Demonstrates effectiveness through simulations and experiments.

Abstract

Phase-shifted carrier pulse-width modulation (PSC-PWM) is a widely adopted scheduling algorithm in cascaded bridge converters, modular multilevel converters, and reconfigurable batteries. However, non-uniformed pulse widths for the modules with fixed phase shift angles lead to significant ripple current and output-voltage distortion. Voltage uniformity instead would require optimization of the phase shifts of the individual carriers. However, the computational burden for such optimization is beyond the capabilities of any simple embedded controller. This paper proposes a neural network that emulates the behavior of an instantaneous optimizer with significantly reduced computational burden. The proposed method has the advantages of stable performance in predicting the optimum phase-shift angles under balanced battery modules with non-identical modulation indices without requiring extensive lookup tables, slow numerical optimization, or complex controller tuning. With only one (re)training session for any specified number of modules, the proposed method is readily adaptable to different system sizes. Furthermore, the proposed framework also includes a simple scaling strategy that allows a neural network trained for fewer modules to be reused for larger systems by grouping modules and adjusting their phase shifts. The scaling strategy eliminates the need for retraining. Large-scale assessment, simulations, and experiments demonstrate that, on average, the proposed approach can reduce the current ripple and the weighted total harmonic distortion by up to 50 % in real time and is 100 to 500 thousand times faster than a conventional optimizer (e.g., genetic algorithms), making it the only solution for an online application.