Frequency-enhanced hierarchical attention network for short-term wind power forecasting

Wind power has achieved widespread application in electricity generation, and its installed capacity has demonstrated consistent annual growth. However, the inherent unpredictability and stochastic nature of wind resources result in unstable power output, which compromises grid reliability and operational security. To address these challenges, a lightweight network, Frequency-Enhanced Hierarchical Attention Network, has been developed. Specifically, the network captures frequency-domain features and spatial correlations in wind power data through its hierarchical attention module, and the Bidirectional Long Short-term Memory module enables robust modeling of long-term temporal dependencies. First, hierarchical attention incorporates frequency-domain information into both channel and spatial attention mechanisms through the Discrete Cosine Transform, effectively integrating spectral features and suppressing high-frequency noise. Furthermore, the integration of dual-attention mechanisms enables adaptive focus on critical features and spatiotemporal patterns via dynamic weight allocation. Second, the Bidirectional Long Short-term Memory enhances predictive capability for complex wind power data through bidirectional information flow and gated control mechanisms. Our result has been validated on two public datasets from Kaggle and a real-world dataset from power plants. Multiple comparative experiments demonstrate that the proposed network achieves higher predictive accuracy and has a lower computational training time.