AI-based Assimilation of Clouds into WRF and ICON Using Sky Cameras and Webcams

Abstract:

Cloud-resolving numerical weather prediction (NWP) models allow to forecast cloud processes at high spatio-temporal resolutions of a few kilometres and minutes. Nevertheless, they often fail to accurately predict cloudiness evolution which affects the prediction of severe weather events as well as solar irradiance. Refining the initial conditions of regional-scale models in terms of clouds using data assimilation is an efficient means for improving short-term cloud cover and irradiance forecasts. The growing amount of sky cameras offers frequently updated cloud observations that are not used in data assimilation so far. Both visible- and infrared-wavelength range images of the sky are considered in this work. The infrared images are obtained from Reuniwatt’s thermal infrared all-sky imager Sky InSight. This all-sky camera is already used for minute-scale PV production forecasting around the globe. It allows the retrieval of cloud parameters such as cloud fraction, cloud-base height, or cloud type at day and night time with a constant accuracy. Moreover, the German Weather Service operates multiple high-resolution webcams all over Germany, that offer visible-range cloud observations. The aim of this work is to detect cloud features from the visible and infrared camera images of the sky and to explore how that would translate into forward operators based on artificial intelligence. Different approaches for cloud data assimilation are evaluated within the framework of the models WRF and ICON over Germany.