FDM-HANet: A Frequency-Decomposed Hierarchical Attention Network for Non-Stationary Wind Power Prediction
Abstract
Accurate short-term wind power prediction plays a critical role in enhancing wind power utilization for power grid dispatching operations. However, existing methods often fail to fully consider the non-stationary characteristics of wind power sequences, with short-term high-frequency fluctuations that give rise to strong time-variability, while long-term structural mutations occur in the trend terms. To address this, a prediction framework considering non-stationary characteristics is developed for wind power forecasting, in which a frequency decomposition module (FDM) and a hierarchical attention (HA) module based network are proposed. Specifically, the FDM decomposes the raw sequence into low- and high-frequency components and fuses multi-scale features to enhance embedding quality. The HA module captures shortterm temporal dependencies, models long-term trend dynamics, and enables cross-channel feature alignment through three specialized attention mechanisms. The experimental results on wind power dataset demonstrate that, the proposed method achieves significant improvements in prediction accuracy, with mean square error (MSE) reduced by $\mathbf{1 9. 2 \%}$.