To enhance the stability of photovoltaic (PV) grid connection and improve the accuracy of PV power prediction, an ultra-short-term prediction method of PV power based on similar day clustering, swarm decomposition (SWD) and MBI-PBI-ResNet network deep learning network model is proposed. First, the expected frequency of solar irradiance is extracted using a fast Fourier transform (FFT). This is used as a clustering eigenvector, and similar day clustering is achieved using adaptive affinity propagation clustering (AdAP) based on this clustering eigenvector. Second, each class of similar days is decomposed separately using swarm decomposition algorithms to extract the multi-scale fluctuation pattern features of the original data. Finally, MBI-PBI-ResNet is used to realize the mining of temporal features under the influence of multivariate correlation of weather environment. Also it is used for the mining of spatial features of local waveforms and long time-dependent temporal features of multiscale components at the same time, as well as the combined integration of different types of features to realize ultra-short-term prediction of PV power generation. The results show that this method improves the prediction accuracy by more than 3% compared with other deep learning methods in the field of ultra-short-term prediction of photovoltaic power. This indicates that this method has higher prediction accuracy and stronger generalizability in the field of ultra-short-term prediction of photovoltaic power.