Abstract: It is vital to achieve carbon emission peak and carbon neutrality in China by building new power systems with high renewable energy penetration. Photovoltaic (PV) power will occupy a high proportion due to its broadly distributed nature. Because of the PV randomness, highly integrated renewable energy power grids will require strong abilities of area coordination and interaction, depending on accurate regional predictions. Compared to single forecast, regional forecast should track cloud motion within large areas and study weather variations among PV power sites. It should also avoid repeated modeling for each power plant. Thus, the ultra-short-term forecast method for regional up-scaling PV power was proposed based on satellite remote sensing data. It contains multispectral image fusion, image prediction and two-level generative PV forecasting. The method can take full merits of multispectral satellite images, and can reduce the impacts on PV power predictions caused by image forecast errors. Based on open cases from European weather satellite and Belgium provincial PV regions, it can be proved that the method increases the accuracy under horizons of 1.5 hours and above, meeting requirements of real-time dispatch between regional power grids.