In the study of photovoltaic power prediction based on satellite cloud images

capturing the cloud features that interfere with photovoltaic power generation can effectively improve the accuracy of power prediction. However

the current method has the following shortcomings: 1) The limited spatial information of photovoltaic power station leads to the lack of fine spatial modeling. 2) There is a significant deviation in the position of the occluded clouds calculated by experience. This problem will make the model obtain invalid interference cloud features. Aiming at these problems

this paper introduces the spatial distribution information of photovoltaic power stations

and proposes an end-to-end photovoltaic power prediction model to replace the empirical calculation of the interference cloud position. In this model

a two-stage cloud feature dynamic coding module is proposed. The module uses the cross-attention mechanism to construct the coupling relationship between multi-modal data. This operation can guide the shadowing effect of clouds in the model learning space. Then

the extracted cloud features are sent to the feedforward neural network to generate an interference factor that describes the degree of interference to photovoltaic power generation. The interference factor and other factors constitute a multi-variate time series. The hybrid neural network is used to capture the dependencies in the time series to achieve accurate power prediction. Simulation results show that compared with other advanced methods

the mean absolute error (MAE) and root mean square error (RMSE) are reduced by 15.22% and 15.28 % respectively

and the R2 is increased by 3.17% on average in the prediction results of 15 minutes in advance.