All-sky imager based irradiance nowcasts: combining a physical and a deep learning model

Abstract

Improved solar irradiance nowcasts based on all-sky imagers. Hybrid physical and end-to-end machine learning (ML) model. The ML model is based on an multi-modal deep learning model combining an vision transformer (for images) with an time series transformer (for time series data). Skill score improvements >12% points are achieved. Correct detection of cloud ramp rates improved by >8% points.