To enhance the accuracy of wind power forecasting, a short-term wind power forecasting method is proposed, one that synergistically integrates deep feedback learning with attention mechanisms. First, the historical data of numerical weather prediction (NWP) from the wind farm is taken as the original input. A dual-layer long short-term memory (LSTM)-based learning model is used for the preliminary prediction of wind power. Next, an error estimation model is established based on an extreme gradient boosting (XGBoost) algorithm. This enables fast estimation of the initial prediction errors given the future NWP data. Then, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is used to decompose the initial prediction errors into error sequences of different frequency bands. These serve as an additional feedback input for the secondary prediction of wind power. Also, an attention mechanism is introduced into the secondary prediction model to dynamically allocate weights to the wind power forecasting and error sequences and thereby instructing the prediction model to fully mine and learn the key features related to the prediction errors during the learning process. Finally, the simulation results indicate that the proposed method can remarkably enhance the reliability of short-term wind power forecasting.