A wind power prediction method integrating dynamic multi-scale spatio-temporal modelling, adaptive multi-strategy local decomposition, and meta-learning ensemble model

Accurate wind power prediction is critical for advancing the energy transition and ensuring stable power system operation. However, wind power's inherent non-stationarity poses significant challenges to high-precision forecasting. To address this, this study proposes an innovative multi-stage adaptive prediction framework. The framework first uses the hierarchical density-based spatial clustering of applications with noise (HDBSCAN) model to construct homogeneous wind turbine clusters based on their operational characteristics and geographical locations. Subsequently, for each cluster, a multi-scale dynamic spatio-temporal graph convolutional network (MSDGCN-ST) model is designed to construct dynamic graphs, capturing the time-varying interactions and spatial adjacency relationships between turbines to represent deep spatio-temporal features. To handle non-stationarity, an adaptive multi-strategy local decomposition (AMSLD) method is introduced, which adaptively determines the decomposition scale by calculating local signal complexity and selects key predictive components based on mutual information maximization. Finally, a meta-learning-based ensemble system uses the model-agnostic meta-learning ++ algorithm to rapidly optimize base models like Informer, Autoformer, and TFT for different prediction scenarios. Their outputs are dynamically fused via a meta-attention mechanism. Comprehensive validation on a real-world dataset confirms the framework's significant advantages in multi-step prediction tasks, outperforming baseline models.