Accurate and real-time wind power forecasting can effectively mitigate the impact of wind power grid connection fluctuations on the power system, enhancing the safety, stability, economy, and controllability of wind power generation. However, ultra-short-term wind power forecasting currently faces challenges such as incomplete time-series data, difficult-to-eliminate outliers, low interpretability of prediction features lacking physical meaning, and unstable multi-step predictions. To address these issues, this paper proposes an ultra-short-term wind power forecasting framework based on data-feature-model hybrid optimization. Firstly, the unsupervised clustering silhouette coefficient is used to adaptively determine the optimal parameters of the DBSCAN algorithm within each wind speed interval, enabling the identification and correction of abnormal data. Secondly, three correlation indicators—Pearson correlation coefficient, Grey Relational Analysis (GRA), and Maximum Information Coefficient (MIC)—are weighted and integrated with equal weights to select factors strongly correlated with wind power as model inputs. Finally, the window attention mechanism is combined with Fourier transform to replace the sparse attention module of the original Informer model, constructing an improved prediction model. Experimental results on three real-world datasets (SDWPF, Kaggle, and Penmanshiel) show that compared with state-of-the-art models, the proposed method achieves higher prediction accuracy, better stability, and faster execution efficiency. This method provides reliable data support for the friendly grid connection of large-scale wind power, verifying the effectiveness of synergistic optimization of data quality, feature relevance, and model architecture. • Adaptive DBSCAN-based wind power data preprocessing via wind speed intervals. • Comprehensive correlation-driven feature selection method for wind forecasting. • Enhanced Informer model with window attention and Fourier transformer.
Xu et al. (Sun,) studied this question.