Multi-step wind speed prediction faces the uncertainty challenge matched with step length, which hampers the safe operation of wind power systems and the efficient utilization of wind energy. The recent divide and conquer prediction method uses decomposition techniques to decompose the original sequence into subsequences in an attempt to address this challenge. However, few studies have considered interaction models with input features and internal parameters of different subsequences to further reduce the wind power uncertainty. To this end, an interaction model based on synchronization learning strategy is designed to reduce uncertainty and realize adaptive features and parameters selection. Particularly, the synchronization learning strategy based on improved bald eagle search algorithm can simultaneously perform input time features selection and prediction model parameters optimization, eliminating the influence of model internal parameters on the feature selection results. Besides, a second-level decomposition technique is designed for improving the predictability of subsequences. Specifically, the original wind speed data are decomposed by exponential smoothing filter into trend and remaining sequences, and the remaining sequence is further processed by variational mode decomposition with compromise decomposition parameters. Finally, all the subsequences are modeled separately, and each subsequence is individually implemented the synchronization learning strategy. For verifying the superiority of the designed model, three experiments are conducted, and 15 models are compared. The results show that the designed model possesses the smallest 1-step and multi-step prediction errors for the four seasonal datasets.