Self-Similar Super-Resolution of MSG Irradiation Images for PV Estimation

Satellite-derived solar irradiation data from Meteosat Second Generation (MSG) offerextensive coverage but at a limited coarse spatial resolution (∼ 5km), limiting theireffectiveness for precise solar energy assessments. These coarse-resolution pixels inherentlyaverage out local cloud-induced variability, causing inaccuracies in photovoltaic(PV) applications sensitive to small-scale fluctuations in irradiance. To overcome thislimitation, this study introduces a deep learning-based super-resolution method thatleverages self-similarity inherent in cloud structures. By artificially downscaling originalsatellite images (e.g., from 220×220 to 55×55 pixels) to create low-resolution inputs,a deep convolutional neural network (CNN) is trained to reconstruct lost details andrecover the original resolution. This process preserves total pixel-integrated energywhile accurately restoring realistic local gradients and cloud-induced spatial variability.Once trained, the super-resolution model is applied recursively to native-resolutionsatellite images, achieving finer spatial resolutions (approximately 2.5 km per pixel).These enhanced irradiation maps maintain physical consistency in energy distributionand provide sharper detail critical for PV applications, such as accurately depictingrapid sunlight transitions under patchy cloud conditions. Accurate distribution ofirradiance at smaller scales is crucial due to nonlinear effects, including the directionalityof cloud-scattered light, tilted irradiation impacts, inverter performance limits, andimplications for self-consumption. The reliability and realism of the super-resolvedoutputs are validated against original satellite imagery and independent ground stationmeasurements. The resulting self-similarity-based super-resolution approach offers arobust solution for bridging the gap between coarse geostationary satellite data and thehigh-resolution requirements of distributed solar energy analyses, without necessitatingadditional high-resolution observational data.