Abstract: As the global energy landscape transitions toward low-carbon sources, high-precision short-term power forecasting for distributed photovoltaic (PV) systems has become increasingly critical for optimal station dispatch and grid integration. However, initial PV power forecasts inevitably contain errors arising from input data noise and model imperfections. These error sequences are typically characterized by strong non-stationarity and multi-scale feature coupling, where systematic deviations, periodic fluctuations, and high-frequency stochastic disturbances are deeply intertwined. Direct modeling of such raw error sequences often yields limited predictive accuracy because the physically meaningful and predictable components are obscured by strong stochastic noise. To address this challenge, this paper proposes a novel error-correction method based on successive variational mode decomposition (SVMD) and the Mamba model to effectively enhance the accuracy of short-term distributed PV power forecasting. The proposed approach first constructs the initial forecast error sequence by subtracting the preliminary predicted power from the actual measured power. Since different weather scenarios give rise to distinct error characteristics, the historical data are classified into three typical generation scenarios: sunny, cloudy, and overcast. For each scenario, the raw error sequence is adaptively decomposed using SVMD, which successively extracts a series of intrinsic mode functions (IMFs) with different center frequencies. Unlike conventional variational mode decomposition, SVMD does not require prespecifying the number of modes; it automatically determines the optimal decomposition by adding constraints that minimize spectral overlap between the newly extracted mode and the residual signal. Low-frequency IMFs capture the slow-varying systematic bias of the initial model; mid-frequency IMFs reflect periodic error patterns linked to diurnal solar variations; and high-frequency IMFs mainly contain random noise caused by transient meteorological disturbances. To avoid subjective manual tuning, the red-billed blue magpie optimizer (RBMO) is employed to automatically identify the optimal SVMD balance parameter by minimizing the envelope entropy of the decomposed modes. After decomposition, all IMF components together with the historical error data are fed into the Mamba model for error prediction. The Mamba architecture, built upon a selective state-space model framework, offers linear computational complexity with respect to sequence length and features an input-dependent selection mechanism that dynamically adjusts state update rules according to the input content. This mechanism suppresses irrelevant noise injection during high-frequency fluctuation periods while enhancing the memory of systematic and periodic patterns, making it particularly suited for error sequences with low signal-to-noise ratios. The final corrected power forecast is obtained by superimposing the predicted error compensation value onto the initial forecast. The proposed SVMD-Mamba framework is validated using real operational data from a 90 kW distributed PV station in southern China. Experimental results across all three weather scenarios demonstrate that the method significantly enhances forecast accuracy, with the root mean square error (RMSE) and mean absolute error decreasing substantially and the coefficient of determination increasing after correction. Specifically, the RMSE reduction reaches 9.95% for sunny conditions and 10.02% for overcast conditions, with a corresponding improvement of 1.58% for cloudy conditions. The relatively smaller gain under cloudy skies is consistent with the inherently lower predictability of error sequences in that scenario, where random cloud-driven fluctuations dominate. Ablation experiments further verify that both the SVMD decomposition and the RBMO-based parameter optimization contribute independently and cumulatively to the overall performance gains. These findings indicate that the proposed method possesses strong engineering applicability as a complementary error correction module for enhancing distributed PV power predictions.