To address the issue that the non-stationarity of wind power makes it difficult for traditional forecasting models to accurately capture variation patterns

thus leading to insufficient prediction accuracy

this paper proposes a novel forecasting framework based on "secondary decomposition-prediction" collaborative optimization. This framework achieves adaptive parameter matching for variational mode decomposition (VMD) by incorporating a chaotic mapping and an elite-guided weighting mechanism within the multi-parameter whale optimization algorithm (MP-WOA). Combined with complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)

the method enables adaptive multi-scale analysis of power series

forming multi-scale feature representations with physical interpretability. A spatiotemporal attention fusion network is further designed to capture local dynamic patterns of decomposed features via bidirectional long short-term memory networks. An efficient temporal prediction model is employed to establish global dependencies between power sequences and environmental variables. Experiments conducted on real wind farm datasets demonstrate that

compared to traditional wind power forecasting methods

the proposed framework exhibits superior performance and robustness

particularly under highly volatile conditions

offering new insights for power forecasting in complex environments.