Combining deep learning and physical models: a benchmark study on all-sky imagerbased solar nowcasting systems

Abstract

Intermittent solar irradiance due to passing clouds poses challenges for integrating solar energy into existing infrastructure. By making use of intrahour nowcasts (very short-term forecasts), changing conditions of solar irradiance can be anticipated. All-sky imagers, capturing sky conditions at high spatial and temporal resolution, can be the basis of such nowcasting systems. Herein, a deep learning (DL) model for solar irradiance nowcasts based on the transformer architecture is presented. The model is trained end-to-end using sequences of sky images and irradiance measurements as input to generate point-forecasts up to 20 min ahead. Further, the effect of integrating this model into a hybrid system, consisting of a physics-based model and smart persistence, is examined. A comparison between the DL and two hybrid models (with and without the DL model) is conducted on a benchmark dataset. Forecast accuracy for deterministic point-forecasts is analyzed under different conditions using standard error metrics like root-mean-square error and forecast skill. Furthermore, spatial and temporal aggregation effects are investigated. In addition, probabilistic nowcasts for each model are computed via a quantile approach. Overall, the DL model outperforms both hybrid models under the majority of conditions and aggregation effects.