基于动态时空注意力图卷积网络与误差修正的光伏功率预测方法

Abstract Superior short-term photovoltaic (PV) power forecasting is essential for formulating PV generation plans and implementing economic dispatch in power systems. However, existing short-term forecasting methods do not delve deeply into the interpretability of input features. They rely on static correlation analysis methods for processing numerical weather prediction (NWP), and often overlook the crucial step of correcting power forecasting errors. Addressing these three research deficiencies, this study develops a short-term PV power forecasting method employing a dynamic spatiotemporal attention graph convolutional network (STAGCN) with decomposition-based error correction. Initially, the spatio-temporal importance model explanation method is employed for feature extraction from multiple NWP sources, identifying key features for model forecasting. Furthermore, the spatiotemporal feature transformation and fusion technique is incorporated to construct a dynamic adjacency matrix and spatiotemporal dynamic graph structure using derived multi-source features. This structure feeds into the STAGCN model, yielding initial forecasting results. Ultimately, the improved complementary ensemble empirical mode decomposition with adaptive noise and variational mode decomposition, long short-term memory network and mixed kernel density estimation are utilized for the decomposition and forecasting of error sequences, subsequently refining the preliminary forecasting results. Simulation results demonstrate that the proposed forecasting approach enhances model interpretability, reduces forecasting errors, and achieves improved forecasting accuracy.