Neural Network-Based Submodule Capacitance Monitoring in Modular Multilevel Converters for Renewable Energy Conversion Systems

The widespread development of medium-voltage and high-voltage direct current transmission systems has highlighted the modular multilevel converter (MMC) as a crucial enabling technology. However, the overall performance and lifetime of the MMC strongly depend on the integrity of its submodules (SMs), making online capacitance condition monitoring a critical requirement. Unlike recent related studies that rely on computationally heavy matrix-based algorithms or “black-box” artificial neural networks requiring massive offline training datasets, this paper proposes a parametric, adaptive linear neuron network. Mapped directly to the physical equations of the MMC, the method simultaneously exploits the arm current, SM switching state, and capacitor voltage to identify online parametric variations caused by aging or harsh conditions. The proposed scheme is fully non-intrusive, requiring no additional hardware sensors or signal injections, thereby reducing implementation complexity. The simulation results obtained in MATLAB/Simulink (vR2024b) demonstrate the method’s fast convergence and a quantified steady-state estimation error within ±1%. Furthermore, the estimator exhibits strong robustness under severe operating conditions, successfully maintaining accuracy during a 20% capacitance reduction, a 100% active power step variation, dc-link voltage fluctuations, measurement noise, grid unbalances, and harmonic perturbations.