Condition-aware averaged ultra-local predictive control with switching-period input–output collaborative reconstruction for NPC three-level power conversion systems

Model-based predictive control for NPC three-level power conversion systems (PCSs) relies on accurate plant parameters and is sensitive to switching ripple, measurement noise, and digital-delay-induced prediction errors. To address these issues, this paper proposes a switching-period input–output collaborative-reconstruction-based averaged ultra-local model (SPIOCR-AULM). In the proposed framework, parameter uncertainties, coordinate-coupling effects, capacitor-voltage influences, and unmodeled dynamics are absorbed into a unified lumped-disturbance channel and reconstructed online through an extended-state-observer-based structure. On the output side, oversampling-assisted multi-point fusion constructs a switching-period averaged state representation to reduce local ripple peaks and noisy instantaneous samples. On the control-input side, subinterval integration reconstructs the fast variation of the equivalent control voltage, while adjacent-period input actions are coordinated to obtain an averaged equivalent input. The noise-suppression mechanism and observer–compensation error dynamics are analyzed to reveal disturbance attenuation and parameter robustness. To further balance steady-state noise rejection and transient response, a condition-aware adaptive-bandwidth configuration is incorporated as an extension of SPIOCR-AULM, forming CAAB-SPIOCR-AULM. Simulations and experiments on a 125 kW prototype verify improved current quality, effective switching-ripple and measurement-noise suppression, and stable control performance under inductance mismatch, noise disturbance, and bidirectional power-flow operation in practical PCS conditions.