With the massive integration of distributed photovoltaic (PV) systems into the power grid, accurate forecasting of distributed PV power generation is of great significance for the stableoperationof power systems.Existing methods often neglect the spatiotemporal correlations among distributedPVsystems,making it difficult to effectively utilize spatiotemporal featureinformationtoimprovepower forecasting accuracy. Considering the strong spatiotemporal correlations among distributed PVoutputs and the difficulty in obtaining meteorological data, thispaperproposesadistributedPVpower forecasting method based on Spatial-Time-Attention, Graph Convolutional Network (GCN), and Temporal Convolutional Network (TCN). This method achieves accurate PV power prediction under weather-free conditions based solely on historicalpowerdata.First, thespatial-timeattentionmechanism is used to extract the spatiotemporal dependencies of the distributedPVclusteroutput.Next,GCNandTCNareutilizedto construct a spatiotemporal graph convolutional network for prediction training, obtaining the forecastingresults.Finally, a simulationisconductedbasedonrealdatafromadistributedPV powerplant inacertainregionofChina.Thecasestudyresults showthat theproposedmodelhashigheraccuracycomparedto traditionaldeeplearningmodels.

Distributed Photovoltaics;Power Forecasting; Spatial-Time Attention Mechanism; Spatiotemporal Graph Convolutional Network;Weather-free Conditions