Short-term power prediction of regional distributed photovoltaic power plants based on spatiotemporal information fusion

Aiming at the complexity and challenges of short-term power prediction of multi-geographic Photovoltaic (PV) power plants, a prediction method based on spatiotemporal graph convolutional network (STGCN) is proposed. The method combines the sampling cross correlation function to analyze the correlation between individual power stations, and effectively captures the temporal dynamic characteristics and spatial correlation in PV power generation data by integrating the time convolution and graph convolution modules. For the target power station, multiple strongly correlated stations are selected, and their historical power generation data are used for cross-station collaborative forecasting. Compared with weakly correlated stations, the prediction accuracy is improved. The effectiveness of the model is verified by utilizing regional distributed PV power plant data in Anhui Province as an example. Compared with the long short-term memory model, convolutional neural network model, backpropagation (BP) neural network model, and genetic algorithm-optimized BP model, the STGCN model achieves improvements in both prediction accuracy and trend capture capability. Therefore, the proposed spatiotemporal information fusion-based cross-station collaborative forecasting model offers high prediction accuracy and can serve as an important reference for spatiotemporal optimization scheduling in smart grids and the efficient utilization of renewable energy.