PV output forecasting based on the combinations of scene-adaptive decomposition and SENet-reweighted informer

Photovoltaic (PV) output forecasting is vital for grid stability. To improve the accuracy of PV output forecasting, many deep learning models have been adopted. However, traditional forecasting models often struggle to adapt to diverse weather conditions and neglect local feature details. To cope with these limitations, this paper proposes a feature-reweighted, scene-adaptive framework. This framework integrates Whale Migration Algorithm (WMA)-optimized Variational Mode Decomposition (VMD) and SENet-reweighted Informer. In the first stage, features are filtered by correlation analysis and clustered into distinct weather scenes. Next, WMA adaptively optimizes the VMD parameters for each specific scene. This optimization produces intrinsic mode functions (IMFs), which effectively separate multi-scale temporal components. Moreover, the Informer-SENet model incorporates the Squeeze-and-Excitation (SE) block after the self-attention distillation procedure. This approach reweights critical local features while preserving long-sequence processing ability. Finally, the framework was evaluated by using real data from a Ningxia PV power plant. Experimental results demonstrate that the proposed framework achieves a maximum error reduction of 96.8% in MSE over the baseline models.