Robust Estimation Method for Power Angle and Stator Voltage of Pumped-Storage Units Based on Res–Autoformer and Unscented Particle Filter

Accurate estimation of power angle and stator voltage is essential for the safe operation of pumped-storage units and the stable control of the emerging power system, where these units play key roles in peak shaving, frequency regulation, and renewable energy accommodation. However, pumped-storage units experience frequent operating-mode switching, missing and noisy measurements, and strongly nonlinear/non-Gaussian dynamic behaviors, which pose serious challenges to conventional dynamic state estimation methods. To address these issues, this paper proposes a robust estimation approach based on Res–Autoformer and an unscented particle filter. First, to handle sparse and noisy raw measurements, cubic spline interpolation is employed to reconstruct missing entries and improve temporal resolution, and the Savitzky–Golay filter is applied for smoothing and denoising, thereby improving data quality for subsequent model training. Second, a Res–Autoformer prediction model is constructed by integrating a residual network for local temporal feature extraction and Autoformer for capturing long-term global dependencies, achieving high-accuracy prediction of future trajectories of power angle and stator voltage. Finally, the predicted values from Res–Autoformer are used as prior information for unscented particle filter-based nonlinear filtering, which effectively mitigates the effects of non-Gaussian noise and time-varying model parameters. The predictor is trained offline on historical SCADA data (minute-level resolution) and deployed online with PMU measurements (sub-second resolution), enabling the framework to operate at multiple temporal scales without compromising either training coverage or tracking fidelity. Simulation results on an IEEE standard test system integrated with a pumped-storage unit and on field measurement data from a large reversible pumped-storage unit demonstrate that the proposed method achieves excellent estimation accuracy and robustness under operating-mode switching, sensor bad data, and strong non-Gaussian noise. Compared with existing schemes, the proposed method significantly improves the estimation accuracy of power angle and stator voltage, confirming its effectiveness under the tested real-world operating conditions.