Abstract: Accurate wind speed prediction is crucial for ensuring the stability of the power grid and enhancing operational efficiency. To improve the accuracy of predictions, a hybrid model for ultra-short-term wind speed prediction is proposed which integrates wind power dynamic features and channel attention. Firstly, considering the impact of meteorological factors on wind speed variations, a feature matrix is constructed by integrating static and dynamic features of meteorological data, thereby exploring the key underlying factors influencing wind speed. Then, the time-varying filtering empirical modal decomposition is employed to preliminarily decompose the original wind speed, followed by variational mode decomposition to further decompose high-frequency components to reduce the instability of data and enhance the predictability of the model. Subsequently, bidirectional long and short-term memory network prediction models are constructed for each subsequence separately, and an efficient channel attention mechanism is incorporated to assign weights to the multi-channel feature information adaptively, so that the model can focus on the key feature information and thereby improves the prediction accuracy. Finally, the final wind speed prediction values is obtained by integrating the output of all sub-models. Case studies demonstrate that the proposed model achieves superior prediction accuracy and robustness.