FACTF-WMTimesNet: Short-term wind power forecasting based on signal reconstruction and 2D multi-scale feature extraction

Accurate short-term wind power forecasting is crucial for the secure integration of wind energy into power systems. However, the volatility, non-stationarity, and complex multi-periodicity of wind power time series present significant challenges for forecasting. To address these challenges, we propose a hybrid deep learning framework: Frequency-Adaptive Cyclical-Trend Forecasting Fusion with Wavelet-based Multi-scale and Mixture of Linear Experts TimesNet (FACTF-WMTimesNet), which integrates three core modules: frequency-adaptive decomposition, multi-scale feature learning, and dynamic weight adjustment. First, to address the issues of volatility and non-stationarity, we propose the Frequency-Adaptive Cyclical-Trend Forecasting Fusion (FACTF) method, which enhances the model's ability to capture periodic trends and components. Second, we design the wavelet-based Multi-scale TimesBlock (WTimesBlock). This architecture employs multi-level wavelet transforms for localized time-frequency analysis and dilated convolutions for multi-scale receptive field expansion, enabling cross-period dependency modeling in a two-dimensional temporal-spectral space. This approach enables multi-scale feature extraction both within and across periods in a two-dimensional space. Finally, we introduce the Mixture of Linear Experts mechanism to dynamically adjust the weights of the predicted results, further improving forecasting performance and generalization. Experimental results demonstrate that the FACTF-WMTimesNet outperforms existing wind power forecasting methods and the TimesNet model on two real-world datasets at three different time resolutions, demonstrating superior accuracy and stability.