Nowcasting systems for irradiance ramp event detection

Abstract

For the operation of a CSP plant, short-term variability of incoming solar irradiance due to passing clouds can be challenging. Scattered cloud conditions, in particular, result in rapidly and continuously changing direct normal irradiance (DNI) distributions across the solar field, which impede optimal plant control. Therefore, the anticipation of such ramp events is crucial for optimized and fully automated control approaches. While recent works based on data-driven methods and hybridization have shown significant improvements in standard forecasting metrics, ramp event detection has often been neglected. In this work we refer to a simple definition for ramp event detection based on a predefined threshold value of change rate in DNI. We present the validation results of a state-of-the-art hybrid nowcasting system, combining physics-based and data-driven nowcasts and a novel generative deep learning approach to detect ramp events. The results confirm that data-driven and hybrid models optimzed on reducing RMSE do not work well to predict ramp events whereas the novel generative model shows good results in anticipating such ramp events.