In order to solve the problem of new energy power prediction modeling and low precision caused by the randomness, volatility and intermittence of wind power generation and photovoltaic power generation, based on the excellent performance of variational autoencoder (VAE) deep learning model in the aspect of time series modeling and nonlinear approximation feature extraction, the VAE model was used to carry out the shortterm power prediction research of new energy power station. The prediction result of VAE was compared with the prediction results of the recurrent neural network (RNN) and long shortterm memory (LSTM) deep learning methods and the support vector regression (SVR) machine learning method. The independent power prediction results of photovoltaic power plants and wind farms show that the deep learning model has better prediction performance than the baseline machine learning model, and the VAEbased prediction method can learn higherlevel features and its prediction performance is better. The RMSE,MAE and R2 values of photovoltaic power prediction model are 1.593,1.098 and 0.973 respectively. The windsolar integrated power prediction results show that the VAE and RNN models can improve the accuracy of power prediction,and R2 values of integrated power prediction model are 0.96 and 0.91 respectively.