Research on Model Predictive Control Strategy of Five-level ANPC Converter Based on Gradient Descent Method

While conventional finite control set model predictive control (FCS-MPC) for five-level active neutral-point-clamped (5L-ANPC) converters demonstrates promising performance, it faces three inherent limitations: (1) computational complexity due to exhaustive vector search, (2) variable switching frequency, and (3) empirical weighting factor dependency for capacitor voltage balancing. To overcome these limitations, this paper proposes a gradient descent-based model predictive control (GD-MPC) strategy. Firstly, during the voltage vector selection phase, the gradient descent algorithm enables iterative dynamic computation, effectively slashing the required vector evaluations from 512 to 2 per control cycle compared to conventional FCS-MPC. Secondly, a three-vector synthesis technique is implemented to enforce fixed-frequency switching operation, achieving a 3.94% reduction in total harmonic distortion (THD) according to simulation data. Furthermore, the redundant vector is used to stabilize the capacitor voltage without requiring empirical weighting factors. Comprehensive simulations under dynamic and steady-state conditions validate the effectiveness of the GD-MPC strategy.