Aiming at the problems of high stochasticity of time series data and low prediction accuracy of a single model for wind power, a short-term wind power prediction model based on ICEEMDAN-TCN-lightGBM is proposed. Firstly, the original sequence data are decomposed into multiple subsequences using the improved empirical modal decomposition of fully adaptive noise ensemble (ICEEMDAN). Then the sample entropy is used to evaluate the complexity of each component, and the TCN-lightGBM model predicts each subsequence of the decomposition. TCN algorithm performs multi-feature prediction of high-frequency subsequence, and lightGBM algorithm performs multi-feature prediction of low-frequency subsequence. Finally, the predictions of each sequence are evaluated and the predictions of each subsequence are evaluated. The TCN algorithm performs multi-feature prediction for high-frequency sub-sequences, and the lightGBM algorithm performs multi-feature prediction for low-frequency sequences. The simulation results show that the model suggested in this paper has greater prediction accuracy and better generalization ability compared to other models.