Combining deep learning and physical models for solar nowcasting

Kurzfassung

Sudden changes in solar irradiance on a local scale can significantly influence solar power generation. This intermittent characteristic of the solar resource is mainly caused by passing clouds and represents a challenge when solar energy is integrated into the power system. By making use of intra hour nowcasts (very short-term forecasts), changing conditions on solar irradiance can be anticipated, allowing optimized power plant operation and grid integration. All-sky imagers, capturing sky conditions at high spatial and temporal resolution, can be the basis of such nowcasting systems. However, the benefit of these nowcasting systems heavily depends on the accuracy of the predictions. In a previous work, a hybrid model combining physics-based and persistence nowcasts has proven to be advantageous. In this work, we present a novel deep learning (DL) model based on the transformer architecture for solar irradiance nowcasts and show that integrating this model into the hybrid model further improves the nowcast quality significantly. While the physics-based nowcasts are derived from a pipeline of processing steps to model clouds and anticipating their impact on solar irradiance, the DL model is completely data-driven and trained end-to-end using sequences of sky images and groundbased irradiance measurements as input. For comparison to the literature, evaluation is carried out on a benchmark dataset of 2019 from the same site. First, the nowcast quality of the DL model is analyzed independently on standard forecasting error metrics like root mean square error (RMSE), mean absolute error (MAE), mean bias error (MBE) and forecast skill. For computing the forecast skill, we used the so-called smart persistence (SP) as reference model. Reaching scores of over 28%, the DL model itself already outperforms the previous hybrid model in terms of RMSE. Next, the hybrid model, combining physics-based, DL and SP nowcasts, is evaluated on the same dataset using the same metrics. Compared to the previous hybrid model, the new hybrid model shows significant improvement over all metrics.