Abstract: Considering the fact that most of the existing new energy power prediction methods have difficulty in deeply mining the characteristics of time series data of multi-dimensional variables and result in poor prediction accuracy, a new energy power prediction method based on ensemble learning is proposed. Firstly, three correlation coefficients are combined with the Shapley value method to obtain variables with high correlation. Secondly, the dilated causal-convolutional neural network is used to capture the characteristics of historical power time series data, and the bidirectional-gated recurrent unit network is used to extract the variable features from the past and future combined with the temporal pattern attention. Finally, the outputs of different networks are fused according to the Stacking method. Experiments show that the proposed method achieves impressive prediction accuracy in the ultra-short term and the prediction results outperform the other comparison models.