A TSFLinear model for wind power prediction with feature decomposition-clustering

Wind power forecasting plays a crucial role in ensuring the stable operation of power grids. To enhance the forecasting accuracy, this study introduces a time-step stacking fusion linear model (TSFLinear) with feature decomposition-clustering. Specifically, the time-step stacking operation amplifies the importance of key time steps across the entire time series, while the time-step fusion operation retains short-term dynamic features and integrates long-term trend information, thereby improving the overall quality of the time series. During the model optimization phase, the K-means clustering algorithm is employed to cluster the signals after variational mode decomposition (VMD), generating two new data subsets. This approach effectively eliminates noise and reduces the input of redundant features. Comparative experimental results demonstrate that the TSFLinear model can effectively capture the clustered subsets of high-frequency signals. In addition, through a comparative analysis with seven Transformer variants, the baseline linear model and its variant, TSFLinear, demonstrate a significant challenge to the performance of the Transformer model and its variants in the field of time-series forecasting. The experimental results show that the proposed forecasting strategy achieves optimal R2 values of 0.9848, 0.9914, and 0.9942 across three wind power datasets, indicating excellent prediction accuracy and strong generalization capability.