Multi-Modal Generative Video Prediction of All-Sky and MSG/MTG Satellite Imagery for Solar Irradiance Nowcasting

Abstract

Reliable photovoltaic operation necessitates high-resolution solar irradiance forecasting to mitigate the challenges of solar power intermittency. State-of-the-art generative models have demonstrated exceptional performance in forecasting utilizing EUMETSAT’s Meteosat Second Generation (MSG) satellite [1, 2] and All-Sky-Imager (ASI) data [3]. Since these data sources cover disparate scales in time and space, leveraging jointly their distinct advantages in forecasting models is subject of current research [4, 5, 6]. In this PhD project work, we propose a deep learning, diffusion-transformer-based generative video predicting architecture that processes ASI and satellite data, including MSG or in future next-generation MTG, to simultaneously generate future image frames and irradiance target quantities. Preliminary results are presented for a model variant utilizing MSG-only input data to perform both MSG video prediction and irradiance estimation.