The information in satellite visible images can be used to quantify the cloud movement and thickness

and it has been gradually applied to the field of photovoltaic power prediction. To solve the problem that the selection of formula parameters in satellite cloud image processing is mostly based on artificial experiences

we proposed a general selection method of empirical parameters. On this basis

we established an ultra-short-term photovoltaic power prediction model considering the satellite image feature region positioning

aiming to accurately realize the cloud region localization of blocking the sunlight. Firstly

the satellites visible images need to be standardized and bottom-removed to remove the intraday difference of satellite visible images. Then

based on the feature region positioning algorithm

the occlusion region in the satellite images is located in real time

and the influencing characteristics of cloud occlusion are obtained through the convolutional neural network. Finally

the cloud occlusion characteristics are fused with other influencing factors

and the mapping relationship with photovoltaic power is established to achieve the prediction. The results show that the proposed model can be adopted to effectively solve the problem of intraday difference of the satellite visible images and can realize the accurate positioning of the cloud feature area

and the model shows better prediction performance. The research in this paper can provide a reference for photovoltaic prediction based on cloud images.