Long-Term Wind Power Forecasting Using Series Decomposition and Multi-Scale Attention Mechanism

With the growing global concern over carbon emissions, wind energy, as a crucial component of clean energy, plays a vital role in the power system, offering significant environmental and economic benefits. However, the intermittency and volatility of wind energy pose substantial challenges for wind power forecasting, particularly for long-term forecasting, which are essential for grid integration and system stability. Nevertheless, existing forecasting methods face challenges in addressing the complexities and nonlinear relationships inherent in wind power forecasting, such as poor model interpretability, insufficient exploration of implicit relationships between variables, and inadequate multi-scale feature capture. To tackle these issues, this paper proposes a long-term power forecasting model for wind turbines. The proposed model first decomposes the time series into trend and seasonal components. The trend component is forecasted using linear regression, while the seasonal component is forecasted by employing a two-stage attention mechanism combined with multi-scale feature extraction to explore temporal and variable-dimensional features. Finally, the two components are fused to generate a high-precision wind power forecast. Experimental results demonstrate that the proposed model outperforms existing state-of-the-art comparison methods in terms of forecast accuracy. On the SDWPF dataset, the model reduces the mean squared error (MSE) by an average of 8.9% and the mean absolute error (MAE) by $\mathbf{1. 7 \%}$. On the WPF dataset, the MSE decreases by $\mathbf{1 2. 6 \%}$, and the MAE decreases by $\mathbf{4. 5 \%}$. These results indicate that the proposed model offers significant advantages in long-term wind power forecasting, effectively improving prediction accuracy.