Design of the MDFF-EPA photovoltaic ultra-short-term power prediction algorithm based on FY-4A

The variability of solar radiation significantly impacts the energy output of photovoltaic (PV) power plants, with cloud cover being the primary cause of such fluctuations. To address the issues of power output variability caused by cloud cover and the complexity of ultra-short-term PV power forecasting models, this paper proposes an ultra-short-term PV power forecasting framework and model based on Fengyun-4A (FY-4A) data, named Convolutional-Stacked-LSTM (CS-LSTM). The core task of this framework is first to extract radiation data for specific coordinates, then analyze the corresponding dynamic changes using the Farneb & auml;ck optical flow method to capture the movement trends of radiation, forming optical flow input features. Subsequently, this study integrates FY-4A data with optical flow input features and other information to form a comprehensive multimodal input for the CS-LSTM algorithm to predict. The innovation of this research lies in combining multimodal inputs and optical flow features, effectively enhancing prediction accuracy. Additionally, this paper proposes an ultra-short-term forecasting algorithm that combines multidimensional feature fusion with error gradual approximation (MDFF-EPA). By incorporating historical prediction errors into the feature set, the model can progressively correct and optimize prediction results, improving prediction accuracy and stability. Experimental validation using real operation data from the Huadian Hubei Suixian Yindian PV power plant demonstrates that, compared to the control group, the experimental group with optical flow input features achieved over a 1% improvement in prediction accuracy for certain months. Moreover, in the comparative analysis of prediction accuracy across four weather types - clear, cloudy, rainy, and heavy rain - the experimental group showed a reduction in RMSE (Root Mean Square Error) percentage by 15.04%, 11.97%, 14.84%, and 18.86% respectively. Future research could further consider capturing cloud movement information and improving the optical flow model to enhance the accuracy of ultra-short-term PV power forecasting.