Hybrid Wind Power Prediction Model Combining Time–Frequency Analysis and Deep Learning

To address the limitation that traditional single prediction models struggle to capture the multi-scale time–frequency characteristics of wind power sequences and lack targeted modeling capability for the latent features within data, this paper proposes a novel wind power prediction framework that integrates time-frequency analysis with deep learning. Specifically, the Stationary Wavelet Transform (SWT) is first employed to perform multi-scale decomposition of the raw wind power sequence, separating it into low-frequency components that describe the overall trend and mid- and high-frequency components that characterize local fluctuations. For the low-frequency subseries, a pattern recognition module is developed based on structural feature extraction and unsupervised clustering (K-Means), followed by targeted prediction using a Gated Recurrent Unit (GRU) network. Meanwhile, for the mid- and high-frequency components, Short-Time Fourier Transform (STFT) combined with Singular Value Decomposition (SVD) is applied to extract deep representations in the time-frequency domain, which are then fitted by a Multi-Layer Perceptron (MLP). Finally, the predicted results of all subcomponents are linearly superimposed to obtain the final wind power prediction. Experimental results demonstrate that the proposed hybrid prediction model effectively exploits the potential multi-scale patterns embedded in wind power sequences, achieving a more comprehensive characterization of their time-frequency dynamics. Compared with conventional single models, the proposed approach exhibits superior predictive accuracy and robustness across multiple evaluation metrics.