6-hour prediction of photovoltaic power using effective cloud fraction from GK2A imagery over South Korean power plants

• A hybrid model using geostationary satellite is developed for short-term PV power prediction. • ECF(day), ECF(night) is applied to enhance correlation with PV power. • CNN-based optical flow model is applied to predict cloud movement up to 6 h. • Model achieves less than 2% nMAE across 31 PV plants in South Korea. Accurate prediction of photovoltaic (PV) power generation is becoming increasingly critical in response to climate change and the growing demand for renewable energy. This study introduces a hybrid model that improves short-term PV power prediction for distributed PV plants across South Korea by integrating satellite-based cloud motion vectors (CMV) with multiple linear regression (MLR) techniques. To quantify the impact of cloud, the effective cloud fraction (ECF) is applied to improve the correlation with PV power. Cloud movements up to six hours ahead are predicted using a convolutional neural network (CNN) designed for optical flow estimation. Validation against actual data from 31 PV plants across 10 regions in South Korea shows that the 5-minute normalized mean absolute error (nMAE) remains below 2% across all lead times, including both daytime and nighttime periods. Although this study focuses on South Korea, the methodology —combining CMV and ECF—provides a generalizable framework that can be applied to diverse geographical and meteorological conditions. This approach demonstrates significant potential to enhance the accuracy of short-term PV power predictions, particularly in regions with limited ground-based observational infrastructure, by leveraging satellite data to assess the optical properties of clouds. This methodology enables power prediction across all locations where PV power plants are situated, thereby contributing to global grid stability and facilitating the integration of renewable energy sources.