WD-SGformer: high-precision wind power forecasting via dual-attention dynamic spatio-temporal learning

Accurate wind power forecasting is essential for grid stability and successful renewable energy integration, yet conventional models exhibit significant limitations in complex terrains due to decoupled spatio-temporal processing, inflexible modeling of dynamic meteorological influences and their associated temporal delays, and static representations of spatial interactions. This paper introduces the Weather-Differentiated Spectro-Geographic Transformer (WD-SGformer) to address these fundamental challenges through a deeply coupled spatio-temporal transformer architecture that enables comprehensive feature interaction throughout the modeling process. The model incorporates two key innovations: weather-differentiated attention that dynamically captures heterogeneous meteorological impacts and their temporal delays on individual turbines, and spectro-geographic attention that adaptively models multi-scale spatial dependencies by integrating geospatial spectral priors with real-time feature similarity measures. Comprehensive evaluation on a field-collected dataset for short-term forecasting demonstrates that WD-SGformer significantly outperforms state-of-the-art methods, achieving a Mean Absolute Error (MAE) of 1.58 MW, with exceptional capability in capturing high-volatility power generation patterns during extreme weather events where conventional approaches demonstrate substantial performance degradation. Although this superior accuracy requires increased computational resources, the results validate the effectiveness of our advanced architecture for high-stakes forecasting applications in complex meteorological environments, representing a significant advancement in wind power prediction technology for renewable energy systems.