Solar Irradiance Intraday Forecasting Using LTSM Networks, Ground Measurements and Satellite Information
Abstract
Accurate forecasting of solar irradiance at photovoltaic (PV) sites enhances grid reliability and supports key operational tasks such as reserve planning and energy trading. This study presents a hybrid, dual-input recurrent neural network for intraday forecasting of global horizontal irradiance (GHI) in the Argentine Pampas incorporating past ground measurements and satellite images. The model combines sequential information from recent observations with exogenous predictors representing spatial and geometric characteristics. Sequential inputs include the clear-sky index ( kc ), reflectance indices derived from visiblechannel satellite imagery at multiple spatial scales, and meteorological variables. The exogenous branch encodes static features such as station coordinates, station identifiers, and solar zenith angle. The network predicts the hourly kc for lead times from 1 to 6 hours, which is then converted to GHI using clear-sky estimates. Results show that integrating satellite-derived reflectance and meteorological predictors substantially improves forecast accuracy compared with models based solely on ground-based measurements, with the largest gains observed at longer horizons and for reflectance indices aggregated over wider spatial areas.
Keywords
GHI Forecast, LSTM Networks, Satellite Indices, Meteorological Information