To solve the problem that traditional ultrashortterm power prediction methods for PV power plants cannot accurately extract both temporal and spatial characteristics of power generation rate, an ultrashortterm prediction method of PV power generation based on spatiotemporal graph convolutional neural networks was proposed. For multiple PV plants in the same area, firstly, graph modeling of the power plants was conducted. The spatiotemporal features of power generation were extracted using graph convolutional networks (GCN) with gated linear units (GLU). Then, based on the extracted spatiotemporal feature information and the historical power generation data of PV plants in the region, the prediction model was trained. Finally, the ultrashortterm prediction of generated power of multiple PV plants was realized. The experimental results show that the method can reduce the RMSE of the ultrashortterm power prediction to 1.122%. It is important for the staff to arrange the dispatch management of the power grid according to the actual situation.