Short-Term Solar Radiation Prediction Based on Convolution Neural Network and Fitted Clear-Sky Model

This study proposes an advanced short-term Direct Normal Irradiance (DNI) prediction model for Concentrated Solar Power (CSP) systems, integrating a convolutional neural network (CNN) with a fitted clear-sky DNI model. Leveraging all-sky images and historical DNI data, the model precisely identifies cloud motion patterns through dense optical flow analysis and forecasts DNI using a targeted region-of-interest (ROI) approach. When maximum cloud pixel velocity falls below 5 pixels per minute, the clear-sky DNI model or persistence model directly applies; for higher-velocity conditions, the CNN predicts the clear-sky index to dynamically adjust the forecast. Experimental validation across diverse weather conditions demonstrates superior accuracy, achieving significantly lower normalized Mean Absolute Errors (nMAEs) and normalized Root Mean Squared Errors (nRMSEs) for various forecast horizons under cloudy skies compared to recent state-of-the-art deep learning approaches. This work delivers a robust solution for preventing thermal shock in the receiver and improving the CSP operational stability.