Abstract: Accurate wind power prediction is critical for enhancing grid stability and optimizing energy dispatch in renewable power systems. To address the challenges of non-stationarity, noise interference, and multi-scale dynamics in wind power sequences, this paper proposes a novel Short-term Wind Power Forecasting Method Integrating TWP and Time-Frequency Dual-channel FxLSTM Architecture. The method integrates adaptive signal decomposition, multi-scale feature extraction, and joint time-frequency modeling to achieve high-precision predictions under complex meteorological conditions.?First, to mitigate mode mixing and noise contamination, Time-Varying Filtering Empirical Mode Decomposition (TVF-EMD) is employed for signal decomposition and reconstruction. This method adaptively designs local cutoff frequencies using instantaneous amplitude and frequency information, followed by B-spline approximation to construct time-varying filters. Effective Intrinsic Mode Functions (IMFs) are selected via correlation coefficient thresholds (e.g., ρ > 0.1), suppressing high-frequency turbulence while preserving ultra-low-frequency (<0.5/day) and diurnal-cycle (0.8–1.2/day) components that dominate wind power dynamics.?Second, a One-Dimensional Multi-level Wavelet Convolutional feature extraction mechanism (1D-WTConv) is designed to capture multi-scale dependencies. Leveraging Haar wavelet bases, it recursively decomposes input sequences into low-frequency (L) and high-frequency (H) subcomponents. Depthwise convolutions with compact kernels are applied hierarchically, followed by inverse wavelet transforms for feature reconstruction. This approach resolves limitations of fixed-receptive-field CNNs and gradient-vanishing RNNs, efficiently modeling both long-term trends and transient fluctuations.?Third, a dual-channel architecture (FxLSTM) synergizes temporal dynamics and frequency-domain enhancements: The FxLSTM module combines scalar-memory sLSTM (with exponential gating and state normalization) and matrix-memory mLSTM (with covariance update rules). sLSTM enhances robustness to abrupt changes (e.g., cold surges), while mLSTM captures spatial correlations among multi-variate meteorological factors (wind speed, temperature, humidity) via fully parallelizable training. The Frequency-Enhanced Channel Attention Mechanism (FECAM) replaces Fourier transforms with Discrete Cosine Transform (DCT) to avoid Gibbs artifacts. It partitions features into subgroups, assigns frequency-specific DCT components (0.375–8 cycles/day), and recalibrates channel-wise weights to amplify dominant frequencies (e.g., diurnal cycles) while suppressing turbulent noise (4–8 cycles/day).?Experiments on real-world datasets from Ningxia and Inner Mongolia wind farms (15-min resolution, cross-seasonal validation) demonstrate proposed method’s superiority: For Ningxia, it achieves RMSE=0.471 MW, MAE=0.331 MW, and MAPE=1.496%, outperforming benchmarks (e.g., xLSTM: MAPE=8.563%; BiLSTM: MAPE=9.321%). In high-turbulence scenarios (Inner Mongolia), prediction errors are reduced by >40% (MAPE from 6.123% to 2.431% in October). The proposed framework provides a new paradigm for high-fluctuation wind power forecasting, with significant implications for grid management and renewable integration.