Abstract: A prediction model based on gated recurrent unit is developed to address the problem of low prediction accuracy of photovoltaic power generation systems. A social network algorithm and an attention mechanism are used to optimize the parameters of the gated recurrent unit,K-means is used to classify the weather types,and a material generation algorithm is proposed to find the best combination of the number of mode decompositions and penalty factors in the variational mode decomposition to realize the decomposition of the initial data. The gated recurrent unit after hyperparameter optimization of social network search algorithm is used to extract the temporal features,and an attention mechanism is introduced to enhance the attention to important information in the temporal input. The operation data of a PV plant in south Xinjiang in 2021 is selected for analysis. The simulation results show that the proposed MGA-VMD-SNS-Attention-GRU prediction model can effectively improve the PV output power prediction accuracy. The average MAPE is decreased by 8.14% and 8.59% compared with SVR and Elman models,respectively.