Abstract: Accurate short-term wind power forecasting is essential for ensuring grid stability, optimizing energy dispatch, and promoting the large-scale integration of renewable energy in the context of global low-carbon development. However, wind power output is inherently nonstationary, characterized by significant noise interference and multiscale fluctuations. Low-frequency components reflect long-term variations, such as seasonal changes and diurnal cycles, whereas high-frequency components mainly result from turbulence, sudden gusts, and measurement errors, severely limiting improvements in prediction accuracy, particularly in complex environments with high turbulence and abrupt wind speed changes. To address these challenges, this study proposes an innovative short-term wind power forecasting method. The method integrates a time-variant wavelet-frequency processing (TWP) module and a time–frequency dual-channel FxLSTM architecture, combining adaptive signal decomposition, multiscale feature extraction, and time–frequency joint modeling to achieve high-precision prediction. First, a time-varying filtering empirical mode decomposition (TVF–EMD) algorithm is employed for adaptive signal decomposition and reconstruction. By utilizing the instantaneous amplitude and frequency of the signal, the algorithm adaptively designs local cutoff frequencies, constructs time-varying filters through B-spline approximation, and selects effective intrinsic mode functions (IMFs) with a correlation coefficient threshold of ρ > 0.1. This approach effectively suppresses high-frequency turbulent noise while retaining key ultra-low-frequency (<0.5 d^–1) and diurnal-cycle (0.8–1.2 d^–1) components, overcoming limitations of traditional decomposition methods such as manual parameter settings and boundary distortion. Second, a one-dimensional multilevel wavelet convolutional (1D–WTConv) mechanism is designed to capture multiscale temporal dependencies. The mechanism decomposes the preprocessed signal into low- and high-frequency subcomponents using one-dimensional Haar wavelet bases, performs hierarchical depthwise convolutions to enhance local feature extraction, and reconstructs comprehensive features via inverse wavelet transform. This approach overcomes the limitations of CNNs with fixed receptive fields and RNNs prone to gradient vanishing. TVF–EMD and 1D–WTConv together form the TWP module, which provides high-quality feature inputs for subsequent modeling. Furthermore, the FxLSTM architecture achieves deep coupling of temporal and frequency information by integrating two complementary memory units: scalar-memory sLSTM with exponential gating and state normalization, and matrix-memory mLSTM based on covariance updates. In addition, sLSTM effectively handles abrupt changes such as cold surges, whereas mLSTM captures complex correlations among multidimensional meteorological variables (e.g., wind speed, temperature, and humidity) and enables parallel training to improve computational efficiency. Additionally, a frequency-enhanced channel attention mechanism (FECAM) is incorporated. This mechanism uses a discrete cosine transform (DCT) to avoid Gibbs artifacts, assigns weights to components in the range of 0.375–8 d^–1), amplifies dominant frequency components, and suppresses high-frequency turbulence noise (4–8 d^–1). Experiments are conducted using 15-min interval cross-seasonal measured data from wind farms in Ningxia and Inner Mongolia, and the results show that the proposed model achieves excellent performance. On the Ningxia test set, the root mean square error (RMSE) is 0.471 MW, the mean absolute error (MAE) is 0.331 MW, and the mean absolute percentage error (MAPE) is 1.496%, significantly outperforming traditional models such as xLSTM and BiLSTM. In the high-turbulence scenario in Inner Mongolia, prediction errors are reduced by more than 40% compared with conventional methods. The model demonstrates strong stability and generalization across different seasons and geographical locations, provides a reliable solution for short-term wind power forecasting in high-volatility scenarios, and offers important technical support for efficient grid operation and renewable energy development.