Network integrating multiscale analysis and nonlinear representation for short-term wind power forecasting

In the field of wind power forecasting, two persistent challenges are: (1) the lack of an integrated approach that combines feature extraction with nonlinear modeling, and (2) limited utilization of multiscale temporal-frequency information. To address these issues, this study proposes a forecasting model termed the Wavelet-Frequency-Time Transformer. The model integrates multiscale signal analysis with nonlinear feature representation to improve forecasting performance. First, an adaptive wavelet packet decomposition layer performs frequency-domain decomposition on the original multivariate signals, extracting sub-band information at multiple scales. A frequency-band selection attention mechanism is then applied to dynamically identify the most representative sub-band features for each variable, which are fused through a gating mechanism. The selected features are subsequently fed into an enhanced temporal fusion transformer, achieving a unified framework for multiscale feature integration and temporal modeling. Experiments conducted on two real-world supervisory control and data acquisition datasets collected during different seasonal periods demonstrate that the proposed method achieves strong forecasting performance, with mean absolute error/root mean square error values of 0.0155/0.0226 on dataset D1 and 0.0291/0.0448 on dataset D2. In addition, the model outputs variable importance and temporal attention scores, providing auxiliary insights for wind farm operation and scheduling.