Satellite-Based High-Precision Clear-Sky Irradiance Estimation Using Machine Learning and Physical Model Harmonization

Accurate short-term estimation of clear-sky Global Horizontal Irradiance (GHI) is vital for solar resource assessment and grid operations, yet existing methods rely on sparse radiometers and coarse global weather reanalysis (e.g., MERRA-2 at 50–70 km spatial resolution with 1 month latency). To achieve scalability in high-precision estimation, we propose a framework that removes dependence on ground measurements by combining multi-satellite observations with reanalysis variables in a physics-supervised machine-learning paradigm. We developed a multi-source-fused high-resolution environmental dataset with 5 min granularity and 1 km spatial precision, incorporating Geostationary Operational Environmental Satellite (GOES-16) observations, polar-orbiting satellite (AURA) data, and MERRA-2 reanalysis. As supervisory physics, we harmonized two complementary parameterized radiative transfer models (MAC2 and REST2V5). The harmonized GHI estimates are used as training labels for a Multilayer Perceptron (MLP) and a Residual Long Short-Term Memory (LSTM) network model. The trained MLP model achieved a root mean square error (RMSE) of 66.67 W/m2, representing a 7.50% reduction over the conventional MERRA-2-driven baseline. For 30-min-ahead forecasting, the LSTM model reduced RMSE by 29.37% over the persistence baseline. Evaluated at four climatically diverse U.S. sites, the system achieves ground-sensor-like accuracy and is deployable anywhere within GOES-16 coverage.