一种基于卫星的结合云透射率预报与物理晴空辐射模型的短期

Abstract Short-term forecasting of solar radiation is crucial for grid integration of solar photovoltaic (PV) power and for grid scheduling and optimization. Enhancing the interpretability of satellite-based short-term forecasts that rely on artificial intelligence is a research focus. In this study, we presented a novel approach to forecast short-term solar radiation by combining satellite-based cloud transmittance forecast and physical clear-sky radiation forecast. The innovation of this study lies in its foundation on atmospheric physics principles, specifically forecasting cloud transmittance and distinguishing between cloudy and clear skies. The cloud transmittance prediction was conducted based on Himawari-8 observations using widely adopted and well-known convolutional neural network (CNN) and long short-term memory (LSTM) networks, while the clear-sky radiation forecast can be conducted with clear-sky radiation model or prediction based on numerical weather prediction (NWP). Compared to other satellite-based baseline forecasting frameworks, the accuracy of our developed framework for short-term forecasting of solar radiation is improved, with an average root mean square error of about 62 W/m 2 over 116 sites and an average relative root mean square error of about 14.36 % with a forecast horizon of 10 min. When the forecast horizon was increased to ranging from 20 min to 4 h, the corresponding average root mean square error increased from 72.16 W/m 2 to 159.75 W/m 2 , and the relative root mean square error increased from 16.71 % to 37 %. This work can forecast solar radiation maps and assist in the flexible regulation of solar PV generation.