A study presents a novel method for accurately predicting cloud cover at solar photovoltaic (PV) plants using geostationary satellite images and an advanced recurrent neural network (RNN). The method, called NCP_CF, uses sequential Himawari-8/9 satellite images with high spatio-temporal resolution to predict cloud fraction (CF) at leading times of 0–4 hours. The system is tested at five PV plants and several stations in China, demonstrating high accuracy and adaptability. The results show that the predicted CF aligns closely with actual power generation, with an average correlation coefficient of nearly 0.8 for the first 2-hour leading time. This technique improves the reliability of solar PV energy generation and enhances its competitiveness in electricity markets. The study highlights the potential of using geostationary satellite data and advanced machine learning models to forecast solar radiation and improve the efficiency of PV systems. The developed system is cyclically updated and operates in near real-time, providing reliable CF information for better power generation forecasting. The method is particularly effective for short-term forecasts, with performance degrading beyond 2 hours due to the vanishing gradient problem. The system is applicable to various PV plants and can be used for power smoothing and load-following applications. The study also emphasizes the importance of accurate cloud cover nowcasting for improving the stability and efficiency of solar PV energy generation.A study presents a novel method for accurately predicting cloud cover at solar photovoltaic (PV) plants using geostationary satellite images and an advanced recurrent neural network (RNN). The method, called NCP_CF, uses sequential Himawari-8/9 satellite images with high spatio-temporal resolution to predict cloud fraction (CF) at leading times of 0–4 hours. The system is tested at five PV plants and several stations in China, demonstrating high accuracy and adaptability. The results show that the predicted CF aligns closely with actual power generation, with an average correlation coefficient of nearly 0.8 for the first 2-hour leading time. This technique improves the reliability of solar PV energy generation and enhances its competitiveness in electricity markets. The study highlights the potential of using geostationary satellite data and advanced machine learning models to forecast solar radiation and improve the efficiency of PV systems. The developed system is cyclically updated and operates in near real-time, providing reliable CF information for better power generation forecasting. The method is particularly effective for short-term forecasts, with performance degrading beyond 2 hours due to the vanishing gradient problem. The system is applicable to various PV plants and can be used for power smoothing and load-following applications. The study also emphasizes the importance of accurate cloud cover nowcasting for improving the stability and efficiency of solar PV energy generation.