HelioNet-IR: Combining Infrared and Visible Satellite Images for Solar Irradiance Forecasting in the Early-Morning Hours

Abstract

Forecasting of solar irradiance is crucial for integrating large shares of photovoltaics into the electricity grid. On timescales up to a few hours ahead, satellite-based (SAT) forecasts can significantly improve upon numerical weather predictions (NWPs). Conventional SAT methods derive cloud motion vectors (CMV) from consecutive images and extrapolate these to forecast future cloud situations. A semi-empirical version of the Heliosat method is widely used to retrieve global horizontal irradiance from visible-range satellite images via the cloud index (CI) as key parameter. When derived from the visible spectrum, CI computation is restricted to daylight hours, and before sunrise, no SAT forecast is available for the early morning. Here, we present HelioNet_IR, a convolutional neural network with UNet architecture to forecast CI derived from Meteosat second generation images without strictly relying on sun illumination. To do so, input CI is complemented with two additional infrared channels. Forecasts of HelioNet_IR are benchmarked against different CMV and NWP models and its purely CI-based predecessor HelioNet_VIS over one full year (2024) with lead times (LTs) up to 4 hr and 15-min resolution. HelioNet_IR can increase SAT-forecast availability from 22% to 100% for forecasts initiated before 8 AM. It notably outperforms NWP for $LT\le 150\text{min}$, reducing root mean square error by over 40% within the first hour. During daytime, reference SAT models are outperformed for all LTs considered.