Abstract: In order to improve the forecast accuracy of ultra-short-term wind power,an improved ultra-short-term wind power forecast model based on variational mode decomposition convolutional neural network （AVMD-CNN）,gated recurrent unit （GRU） and attention mechanism （Attention） is proposed.Firstly,the wind power sequence is decomposed into K sub-modes by using the improved VMD.Then,each sub-mode is classified by sample entropy （SE） and center frequency.According to the classification results,each sub-mode is given a normalization method,and input into GRU-Attention and CNN-GRU-Attention models for training and forecasting according to SE values.Finally,the final results are obtained by superimposing the forecast results of each sub-mode,so as to complete the ultrashort-term wind power forecast.Using the determination coefficient（R~2）,mean absolute error （MAE）,root mean square error （RMSE）,and mean absolute percentage error （MAPE） as the accuracy assessment indexes,the actual arithmetic examples show that the R~2 of the proposed model is improved by 12.06%on average compared with other methods,and the MAE,RMSE,and MAPE are reduced by59.36%,62.49%,and 48.34%respectively,with high prediction accuracy.