GateFormer: forecast solar radiation from sky images at half an hour to better control networks

A Transformer Vision that combines sky images and sun measurements reduces the forecast error by 9.6% to 30 minutes, a key horizon to balance solar production.

Balancing power grids and managing storage are increasingly relying on forecasts that can anticipate sudden changes in solar production when a cloud passes in front of the sun. A team of four researchers presents GateFormer, a Vision Transformer that predicts horizontal global radiation in the coming hour based on images of the sky and the latest sunshine measurements, injected directly into the model's visual information stream.

Two weighted forecasts on the fly

The model produces two estimates: a complete trajectory of the future radiation, and the expected change from the last measurement. It then combines them with a weighting that adapts to the situation. On the NREL-TSI reference set, it achieves a root mean square error of 125 W/m² at 30 minutes, which is 9.6% better than the best reference Transformer, and it obtains the lowest error at the end horizon at the three sites tested (NREL-TSI, Chilbolton and SIRTA), with a gain of 23 to 29% compared to a persistence prediction.

Progress is most pronounced during rapid increases in radiation, precisely the episodes that matter most to grid operators. The system only needs a sky-facing camera and a sunlight sensor.