Due to the volatility of wind energy tends to lead to training sample diversity and the lack of historical data, both of which greatly impact wind power prediction. In response to this problem,  we propose an ultra-short-term wind power prediction method for wind power clusters based on weather feature selection and spatial density clustering. First, a feature selection method based on complete kernel fisher discrimination is used to subject the NWP(Numerical Weather Prediction) information to principal component analysis and extract the most critical wind speed features of each wind turbine. After that, an improved clustering method based on spatial density is used to classify the clusters of each wind turbine in the wind farm based on the above features. Finally, the GRU-D(Gated Recurrent Unit with Decay) method is used to predict the power of each wind turbine cluster and sum it to get the predicted power. The results using historical data of onshore wind power output in a region of Spain show that the root-mean-square error prediction accuracy of the proposed method improves by 0.25% and 2.02% compared to the prediction methods based on traditional DBSCAN(Density-Based Spatial Clustering of Applications with Noise) and K-means clustering methods, and the root-mean-square error prediction accuracy of the model improves by 0.82% compared to the GRU(Gated Recurrent Unit) that cannot handle missing values.