Ultra-short-term wind power predictions have a guiding effect on the operation control and energy scheduling of the unit. In order to weaken the effects of the wind speed of the numerical weather prediction (NWP) on ultra-short-term prediction accuracy, an ultra-short-term NWP wind speed correction method considering multiple error scenarios set division is proposed. The bidirectional short-term memory network (BILSTM) predicts the prediction error of the NWP wind speed in the next 4 hours, and the error scenes set is divided into the wind speed error prediction value based on error scene different BILSTM networks are trained to match error and wind speed forcast error prediction and correct the wind speed. Based on the revision results, some modes have been used to predict the ultra-short-term wind power. The method is applied to a wind farm in Inner Mongolia in China for example. The results show that the method of this article effectively reduces the NWP wind speed error. On the basis of the original data, compared with the non -fixed wind speed of NWP, the RMSE value has reduced 1.859, and the MAE value has decreased by 1.464, and MAPE reduces 26.01%. Among them, the BP neural network ultra-short-term power prediction accuracy increased by 7.5%, and the GRU deep network is increased by 8.7%, the multi-linear regression model is increased by 9.6%, the results verify the effectiveness of the method.