Intrahour solar radiation forecasting based on sun visibility for different cloud types

Given certain cloud fractions in the sky, derived from satellite observations or sky images, surface solar radiation (SSR) attenuated by clouds can differ when the sun is visible (based on sunshine duration threshold defined by World Meteorological Organization) or shaded by clouds. Sun visibility is among the most uncertain factors in solar forecasting. With sky images taken by sky cameras and solar radiation measurements at the World Radiation Center (Davos, Switzerland), this work examines the relationship between cloud cover and sun visibility. By integrating sun visibility into a forecasting system, an improvement in accuracy and reliability of the SSR forecast is aimed. A multivariate deep learning model based on a Long- and Short-term Time-series network is used. The model is trained with extensive cloud information including cloud fraction and cloud types extracted from sky images taken at a ground station, and predicts intrahour (15 min- to 1 hr-ahead) sun visibility and cloud modification factor (CMF). With the distinct patterns of visible and occulted sun, initial CMF forecasts are post-processed based on the prediction bias. Next, we derive all-sky global horizontal irradiance (GHI) from both CMF and clear-sky GHI obtained from radiative transfer modeling. The experimental results show a high accuracy of 96% for the 15-min ahead prediction of sun visibility. The SSR forecasts are validated using ground-based measurements. Compared with the persistence reference, our forecast method achieved a significant skill score of 47% for the 15-min ahead prediction. Higher skill scores are found for certain cloud types, e.g., 55% for stratus-altostratus. • Sky image-derived cloud cover and ground-based irradiance measurements are used. • A forecasting method with regime-switching bias correction is developed. • 96% accuracy for sun visibility prediction and 47% forecast skill are achieved. • Considering sun visibility in addition to cloud cover improves solar forecasts.