Accurate ultra-short-term power forecasting for distributed photovoltaic (DPV) systems is crucial for the intra-day operation of distribution networks. However, the current method based on a graph network only takes a few DPV sites as forecasted objects; when modeling a large number of DPV objects, the massive graph structure will require multiple instances of information propagation to achieve global correlation extraction. Due to the similar output characteristics of adjacent DPV sites, excessive information aggregation will lead to node features tending towards consistency, making information extraction inefficient and insufficient, which limits the improvement of forecasting accuracy. To address the issues above, this study proposes an ultra-short-term distributed PV power forecasting method considering spatiotemporal correlation. First, the DPV sites are clustered into several sub-regions in different layers considering the spatial location of DPV sites and the temporal characteristics of power output. And a hierarchical architecture is constructed from DPV sites to sub-regions based on subordinate relationship and the order of information transmission. After that, the output mode of every sub-region is dynamically described in refinement by filtering out the noise DPV sites with significant differences in outputs. Finally, by hierarchically and sequentially mining the local and global spatiotemporal correlation among output modes, the hierarchical dynamic graph convolutional network is applied to achieve the regional power forecasting. Experimental results based on data from 166 DPV sites demonstrate that the proposed HDGCN model significantly outperforms the best traditional benchmark model, reducing the Normalized Root Mean Square Error (NRMSE) by approximately 38.56% and the Normalized Mean Absolute Error (NMAE) by 33.79% in a 4 h-advance forecasting scale.
Tong et al. (Sun,) studied this question.