Temporal Convolutional Network-Based Capacitor Voltage Prediction With Reduced Switching Frequency for Voltage Balancing in MMC

Modular Multilevel Converters (MMC) are widely employed in high- and medium-power applications due to their modularity, scalability, and fault tolerance. Accurate estimation and control of Sub-Module (SM) capacitor voltages are crucial for ensuring voltage balancing, yet direct sensing increases complexity and cost, while high switching frequency leads to excessive losses. This paper proposes a two-stage approach: first, a multi-layer Temporal Convolutional Network (TCN)-based model predicts SM capacitor voltages, eliminating the need for direct sensing; second, a voltage balancing strategy minimizes switching frequency using the predicted voltages. The Pulse-Width Modulation (PWM) operates at the carrier frequency rather than the sampling frequency, reducing unnecessary switching events and lowering losses. Simulation and hardware experiments validate the approach across various scenarios, including input DC voltage variations, SM capacitance deviations, and output load impedance changes. The model achieves a prediction error of 0.32%, effectively reducing capacitance deviation without extensive tuning. Voltage prediction accuracy remains stable as SM count increases, with maximum errors of 1.5% and 0.9% for 10 SMs, and 1.5% and 2.0% for 20 SMs in the unbalanced and balanced regions, respectively. These results demonstrate the methodology's effectiveness in accurately estimating capacitor voltages and ensuring reliable balancing, with efficient scalability for real-time applications. The model achieves near-zero RMSE and MAE values, demonstrating high prediction accuracy. These results confirm the model's robustness, stability, and scalability, making it suitable for large-scale MMC applications.