Abstract: The volatility and randomness of photovoltaic power pose a challenge to the planning and operation of the power grid, and the improvement in the accuracy of photovoltaic power prediction is of great significance to maintain the stability of the new power system operation. In this paper, an ultra-short-term photovoltaic combination forecasting model based on the Informer is proposed which combined with modal decomposition, multi-dimensional feature modeling, Informer and bidirectional long short-term memory network (BiLSTM). Firstly, the photovoltaic power signals is decomposed into intrinsic mode functions (IMF) of different frequencies by variational mode decomposition to reduce the non-stationarity and complexity of signals. Then, the discrete wavelet transform is used to extract the detail components of the weather factors to realize the complementary advantages of different decomposition algorithms, and the random forest algorithm is used to screen the redundant features for each IMF. Furthermore, the feature matrix is sent to the Informer for modeling, and the critical moment information in different time steps is extracted to improve the prediction efficiency of long time series. Finally, in order to further improve the prediction accuracy of the model, the prediction error is compensated with the BiLSTM after analyzing the characteristics of the error sequence. The model is validated using the actual data, and the results show that the proposed model improves the accuracy of ultra-short-term photovoltaic power prediction.