Solar: a month of sky images is enough for a Transformer Vision to estimate diffuse radiation almost as well as with a year of data

Researchers propose estimating diffuse solar radiation without instruments on the ground, using fisheye images of the sky and calculated atmospheric parameters.

Diffuse horizontal radiation (DHI) is a key factor in the design and control of photovoltaic systems, but it requires expensive instruments that are lacking in many regions. Vision Transformer models can already estimate it from photos of the sky, provided that long series of images are available, which are rarely available. A study published on arXiv tackles this case of data scarcity.

The method replaces the measurements with calculated parameters (extraterrestrial radiation, cyclic encoding of the time and season), first predicts synthetic global radiation to use as an index, then processes in parallel the raw fisheye image and a version where the clouds are highlighted by unsupervised segmentation guided by physics. Trained on a single month of data from the Chilbolton Observatory (United Kingdom), then tested over an entire year, the models remain within 1 to 6 W/m² of error close to a reference Vision Transformer trained on the dataset. This is a cheap way to evaluate the solar resource where measurement stations are lacking.