Ultra-short-term Probabilistic Forecasting of Distributed Photovoltaic Power Generation Based on Hierarchical Correlation Modeling

Accurate probabilistic forecasting of regional distributed photovoltaic (PV) power can provide more comprehensive information support for the optimal operation of active distribution networks. When meteorological measurement or forecasting data is lacking, mining and utilizing spatio-temporal correlation information of distributed PV can effectively improve power forecasting accuracy. However, existing research either struggles to specifically mine spatio-temporal correlation information or loses a significant amount of valuable information during the modeling process. To address this, a method for ultra-short-term probabilistic forecasting of regional distributed PV power based on hierarchical correlation modeling is proposed. Firstly, a clustering method based on deep consistency is employed to divide the distributed PV clusters into subregions, which supports targeted modeling of the spatio-temporal correlations within the subregions. On this basis, a hierarchical graph structure is constructed to simultaneously model the intra-subregion and inter-subregion spatio-temporal correlations, enabling effective utilization of correlation information across different hierarchical levels. Then, a probabilistic forecasting model based on hierarchical graph convolutional neural networks (GCNs) is proposed to mine deep spatio-temporal correlation features between PV power stations, thereby enhancing the accuracy of ultra-short-term probabilistic forecasting of regional distributed PV power. Finally, the effectiveness of the proposed method is validated using actual distributed PV power data sets.