基于多尺度融合动态图卷积网络的多站点太阳辐照度预测OACHSSCD

Aims:High-precision prediction of multi-site solar irradiance is the key to improving the accuracy of photovoltaic power prediction and the efficiency of power grid operation management.However,there is a lack of modeling methods that can simultaneously capture dynamic spatio-temporal correlations across different time scales.Methods:This paper proposes a multi-scale fused dynamic graph convolutional network(MSFDGCN),which employs an attention-based dynamic graph convolutional network to accurately capture the dynamically evolving spatial relationships among sites over time.Meanwhile,a lightweight decomposable multi-scale fusion(DMSF)module was designed to achieve cross-scale feature interaction by decoupling and recombining seasonal and trend components.Finally,structurally identical yet parameter-independent predictors were configured for each time scale to fully exploit the dominant spatiotemporal patterns at different scales.Results:Experimental results on real-world datasets covering three typical climate regions demonstrated that,compared with the advanced adaptive graph convolution and long short-term memory(AGCLSTM),MSFDGCN reduced the mean absolute error by an average of 9.83%while maintaining high prediction accuracy and achieving superior computational efficiency.Conclusions:The proposed method significantly improves the accuracy of multi-site solar irradiance forecasting and provides effective technical support for the optimal operation of photovoltaic power systems.