Regional DPV Short-term Forecasting Method based on Sub-clusters Division and GCN-GRU

Accurate forecasting of distributed photovoltaic (DPV) power at a regional level is significant for enhancing the management of distribution networks and optimizing efficiency. However, DPV are characterized by their high quantity, relatively low capacities, and the challenge of obtaining meteorological data, setting them apart from centralized PV. Many existing studies tend to overlook the importance of effectively segmenting DPV sub-clusters with similar output characteristics and often fail to uncover the temporal and spatial correlations between different sub-clusters. Therefore, this paper proposes a short-term DPV power forecasting method based on sub-clusters division and graph convolution gate recurrent unit neural network (GCN-GRU). Firstly, a subcluster division method based on affinity propagation (AP) clustering is proposed to identify typical benchmark DPV plants. Secondly, the evolution characteristics of the dynamic spatial correlation is extracted by GCN and GRU, then the DPV power forecasting models of typical benchmarks DPV are established. Lastly, we determine the scaling weight by evaluating the coherence and similarity of sub-clusters, ultimately enabling us to forecast DPV power for the entire region. A simulation was conducted using actual DPV power generation data from northern China. The results demonstrate that the proposed method outperforms traditional methods.