With the ever‑increasing grid integration of large‑scale wind power

its pronounced nonlinearity

multi‑scale oscillations

and issues of missing or anomalous data pose serious challenges to ultra‑short‑term power forecasting. To address these issues

this paper presents an ultra‑short‑term wind power forecasting framework that integrates data cleansing and imputation

hybrid signal decomposition

and deep learning. First

missing values are recovered via cubic Hermite interpolation; anomalies are detected using the minimum covariance determinant and Mahalanobis distance and subsequently corrected with a random forest regressor to ensure data consistency. Then

the maximal information coefficient (MIC) and Spearman’s correlation analysis are employed to select key meteorological variables and construct an environmental feature matrix

meanwhile

the jump‑plus AM‑FM mode decomposition (JMD) and variational mode decomposition (VMD) are used to extract multi‑scale power components to thoroughly dissect the signals’ spectral structure and form a power feature matrix. These two feature matrices are then fed into a Dual‑MTSA deep learning model. Within its encoder

TKAN is embedded into Mamba’s selective state‑space framework

and the Mamba’s linear‑time complexity is used to efficiently capture long‑range dependencies across both time steps and variables. Moreover

a B‑spline kernel with gated fusion is utilized to finely represent high‑frequency oscillations and low‑frequency trends. A self‑attention module is then used to globally recalibrate these multi‑source

multi‑scale features

and a final fully connected mapping layer generates ultra‑short‑term forecasts spanning 15  minutes to 3  hours. Empirical results demonstrate that the proposed method can be used to effectively reduce the prediction errors and significantly improves the prediction accuracy and robustness when compared with the benchmark models.