Wind power output exhibits significant volatility and randomness, posing challenges to grid scheduling and wind power integration. To enhance forecasting accuracy, this paper presents CESF-Net, an ultra-short-term wind power prediction model that combines channel-wise EMA decomposition with a hybrid neural network. First, the model employs the densitybased spatial clustering of applications with noise (DBSCAN) algorithm to identify and remove outliers in the wind speed-power relationship, and uses linear interpolation to impute missing values. Second, CW-EMA is applied to decompose multivariate time series into trend and seasonal components along the channel dimension. Then, a time-series segmentation mechanism is introduced to enhance the model′s ability to capture local temporal structures, and a dual-stream network based on fast fourier transform attention and gated recurrent units is constructed to separately extract seasonal and trend features. Finally, the outputs of the dual streams are concatenated to generate the prediction through CESF-Net. Experiments conducted on real wind farm datasets demonstrate that, for the 15-minute forecasting task, the CESF-Net model outperforms commonly used models by 18.78%, 11.11% and 0.26% in terms of MAE, RMSE and R2, respectively. For the 60-minute forecasting task, although the prediction accuracy of all models decreases, CESF-Net still achieves improvements of 2.98%, 2.74% and 0.61% in the respective metrics.