To enhance the accuracy of ultra-short-term wind power forecasting and support efficient power system dispatch, this study proposes an integrated multi-algorithm forecasting model. The variational mode decomposition （VMD） is firstly applied to suppress noise and reconstruct the original power sequence. In the modeling stage, a long short-term memory （LSTM） network is adopted to capture temporal dependencies, followed by a multi-layer convolutional neural network （CNN） to extract local features. A Self Attention mechanism is further incorporated to dynamically focus on critical time steps, resulting in a collaborative multi-module forecasting framework. To evaluate the performance of the proposed model, ablation studies, comparative experiments, and seasonal transfer tests were conducted using data from a wind farm in Shandong Province, China. The results show that, compared to baseline models, the proposed model reduces the mean absolute error （MAE） by 23.5% and the root mean square error （RMSE） by 20%, highlighting the advantages of each module in modeling complex temporal patterns. Additional validations in cross-regional （wind farms in Central and Western China） and cross-energy （photovoltaic plants） scenarios further demonstrate the model's strong generalization capability.

convolutional neural network  /  Self-Attention mechanism  /  wind power prediction  /  gradient boosting regression algorithm  /  variational mode decomposition  /  wind power