This manuscript demonstrates a novel super-resolution (SR) approach to deterministically improve Meteosat Second Generation (MSG) satellite irradiation data, with a focus on qualities for regional photovoltaic (PV) power estimation. The study utilizes enhanced deep residual networks (EDSR) to subdivide MSG observations onto 2 × 2 and 3 × 3 subpixel grids while conserving overall sums, as well as recursively to 4 × 4 and 9 × 9 subgrids. This method works from a single dataset by artificially reducing the MSG resolution and relying on clouds’ self-similarity to learn patterns at larger scales and subsequently apply them to smaller ones. The approach is validated over nearly a decade of resolutions and we show super-resolution data with qualitative improvements similar to true images up to i m 1.45 × the input resolution. We also report spectral signatures from SR irradiation images and demonstrate improved distributions of regional irradiation, e.g., for nonlinear PV modeling. The resulting SR will be used to provide an enhanced, 20-year deterministic irradiation dataset for Germany, though the generalized model is demonstrated on the entire full Earth disc. • Geostationary irradiation super-resolution is possible without secondary target data. • Cloud self-similarity allows training from artificially reduced images. • MSG enhancements are shown to be similar to theoretical true images with 30–45% higher resolution. • Pixel multipliers of 2 or recursively 4 most efficient considering skill and data size. • Deterministic enhancements show some benefits to regional irradiation distributions.
Good et al. (Fri,) studied this question.