Under the background of energy Internet, the effective prediction of renewable energy output by digital means is a key link in the realization of energy Internet technology. To solve this problem, an adaptive ultra-short-term PV power prediction method is proposed. The probability density distribution of the error under different weather conditions is analysed in detail. The difference model is used to predict the PV power in different application scenarios. The decision tree algorithm is used to classify different weather conditions, and the input characteristics under different weather conditions are used differently. For sunny and rainy weather conditions, shallow LSSVM and deep confidence network models are used to predict PV output in the future. Finally, the actual PV data of a place in Henan Province is used to verify that the prediction algorithm combined with deep learning and shallow model can effectively deal with the different prediction scenarios.