Abstract: Short-term day-ahead wind power forecasts are important for power system dispatch planning. In order to improve the accuracy of wind power prediction, this paper proposes a Transformer-based prediction model Powerformer. This model mines the temporal dependence of sequences through the causal attention mechanism, which is optimized to improve the predictability of the data itself through the De-stationary module. The predictive power of the model is improved by designing the trend enhancement and the cycle enhancement modules. By improving the multi-headed attention layer of the decoder, the model extracts the cycle features and the trend features. In this paper, we first preprocess the wind power data. We decompose the wind power data series into the eigenmodal functions with different frequencies and calculate their sample entropy by using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) so that the wind power series are reorganized into the cycle and the trend series. Then, the series are put into the Powerformer model to achieve the accurate prediction of the wind power in a short term. The results show that, although the training time is longer than the existing prediction models, the prediction accuracy of the Poweformer model is better improved. The necessity and effectiveness of each of the modules in this model are verified by the ablation experiments, which has a certain application value.