基于分层特征依赖Transformer的波动性海洋环境下短期海上风电功率预测

Accurate short-term forecasting for offshore wind power is undermined by volatile marine environments, where sudden storms, shifting tides, and intense waves create dynamic spatiotemporal heterogeneity. This volatility threatens grid stability and incurs economic costs due to inefficient reserve allocation. Existing forecasting methods often rely on static correlations and struggle to model the complex, non-linear feature interactions under such conditions. To address this gap, we propose Hieroformer, a novel Transformer-based framework with four key contributions. First, we propose a dynamic feature-dependency hierarchy, built upon directed information flow rather than static correlation, to model the evolving dependencies within volatile marine environments. Second, we propose a hierarchy-aware attention mechanism that instills a physical inductive bias into the Transformer architecture, which overcomes a key limitation of standard, permutation-invariant attention in time series analysis. Third, we introduce a novel frequency-domain filter to refine the input signal, which separates valuable forecasting periodicities from non-informative noise and static components. Finally, we demonstrate through a case study on the IEEE 118-bus system that the model's enhanced forecasting accuracy translates directly into operational cost savings. Experiments on real-world offshore datasets confirm that Hieroformer outperforms state-of-the-art models by an average of 10% in key metrics and maintains high accuracy even during extreme weather events.