Ultra-short-term photovoltaic power forecasting based on two-stage graph convolutional networks for meteorological and geographical information extraction

In recent years, photovoltaic (PV) power generation becomes an essential component of the energy structure, thus making PV power forecasting of significant importance. Artificial Intelligence (AI) based approaches have been widely applied to PV power forecasting due to their strong ability in solving complex nonlinear problems. This paper proposes an ultra-short-term PV power forecasting method based on two-stage Graph Convolutional Networks (GCNs). Firstly, after obtaining the available numerical weather prediction (NWP) data, the stage-1 GCN is utilized to correct the forecasted irradiance by leveraging the correlation among different types of meteorological data, to provide a more accurate irradiance forecasting for PV power forecast. Next, with the multi-station PV data, the stage-2 GCN is adopted to capture the correlation among stations at different geographical locations, to further improve the PV power forecasting accuracy. Finally, the Temporal Convolutional Network (TCN) is introduced to consider temporal features within the PV power forecasting problem. The proposed PV power forecasting approach is validated with practical PV station data, demonstrating improved forecasting accuracy compared to existing approaches.