Vortex-Net: a hybrid short-term wind power prediction model with adaptive decomposition and temporal attention

Abstract Accurate and reliable wind power forecasting models are essential for the optimal dispatch of power systems. However, existing methods still face several challenges, including ineffective signal decomposition, limited feature extraction capability in attention mechanisms, and the diminishing performance gains of RNN-based models. To address these issues, this paper proposes Vortex-Net, a hybrid model for short-term wind power forecasting that integrates the Squirrel Search Algorithm (SSA), Variational Mode Decomposition (VMD), Temporal Pattern Attention (TPA), and a Multi-Layer Stacked Bidirectional LSTM (MBLSTM) network. Specifically, the proposed framework incorporates multiple optimization strategies: (1) raw data are preprocessed through feature correlation analysis and anomaly detection; (2) the non-stationary wind power signal is decomposed into several near-stationary components using SSA-optimized VMD; (3) the TPA mechanism is employed to extract latent temporal features and capture the intrinsic relationships between the input variables and the target output; and (4) the enhanced features are subsequently fed into the MBLSTM network for prediction. The effectiveness of the proposed model is validated through ablation studies and comparisons with baseline models on multiple public datasets. Experimental results show that the proposed model achieves R2 values of 0.99, 0.96, and 0.98 on the three datasets, respectively. Compared with competing models, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are reduced by at least 37% and 36%, respectively. Overall, Vortex-Net demonstrates superior predictive accuracy and generalization capability, thereby improving the reliability and economic efficiency of wind power systems.