Ultra-short-term prediction of regional photovoltaic power based on dynamic graph convolutional neural network

With the increasing scale of grid-connected photovoltaic (PV) power generation, the accurate prediction of regional PV power is of great significance for the safe and stable operation of power systems. However, the temporal spatial correlation of different PV stations is changing with the movement of clouds. This paper proposes a dynamic spatial temporal graph convolutional neural network (DSTGCN) to model the time-varying correlation for improving the accuracy of power prediction. Firstly, according to the spatial distribution of clouds and the location of PV stations, a graph structure considering the weather type of each PV station is constructed. Then, the dynamic graph structure of the PV stations in the region is predicted based on cloud displacement calculation and trajectory prediction. Finally, a regional PV power prediction model is proposed based on the dynamic graph convolutional layer and bidirectional long short-term memory (LSTM) module. By adaptively integrating the information from satellite cloud images, numerical weather prediction (NWP), and historical power, the proposed model achieves higher and more stable prediction accuracy. Case from a real-world large PV region shows that the proposed method can improve the accuracy of 1.5–4 h ahead prediction by 1–3 percent, and the standard deviation of prediction error is smaller, which verifies the superior and stable performance of the proposed model.