As an important part of renewable energy installed capacity, offshore photovoltaic(PV) power generation systems are subject to special meteorological environments and limited far-sea meteorological monitoring conditions. Compared with land-based PV prediction, offshore PV prediction needs to accurately grasp the variable cloud conditions over the sea and analyze the characteristics of marine meteorological fluctuations. Therefore, this paper proposes an ultra-short-term power prediction method based on satellite remote sensing data. Aiming at the uncertainty and fluctuation problems of cloud images, segment-weighted Gaussian fusion of remote sensing images in full wavelength band and variational auto-encoder(VAE) based reconstruction technique are used to propose an offshore power model based on the correction of multispectral cloud maps, and a two-layer generative adversarial network (GAN) is used to predict offshore PV power, which significantly reduces the prediction error. Through the data validation of Singapore Johor Bahru Power Station, the results show that the model can realize the ultra-short-term power prediction of 1 hour and above with high accuracy, and the accuracy is improved by 12% compared with the traditional method, which enhances the reliability of real-time grid scheduling and the ability of grid-connected consumption of renewable energy.