Short-Term Wind Power Forecasting Based on Data Quality Awareness and Global-Local Feature Collaboration

Dian Wang

Abstract: Short-term wind power forecasting is essential for improving power system scheduling and renewable energy integration. However, practical wind farm data are frequently affected by outliers, missing values, redundant meteorological variables, and highly nonlinear temporal characteristics. To address these challenges, this paper proposes a short-term wind power forecasting framework based on data-quality awareness and global local feature collaboration. DBSCAN is employed for outlier detection, KNN imputation is adopted to recover missing values, and correlation analysis is performed to select informative meteorological variables. An improved Transformer incorporating multi-head differential attention is then combined with a temporal convolutional network to jointly capture global temporal dependencies and local fluctuation patterns. Experiments conducted on a real-world SCADA dataset demonstrate that the proposed approach consistently outperforms CNN, TCN, Transformer, and Informer models under multiple forecasting horizons. At the 48-step forecasting horizon, the proposed method reduces MAE and RMSE by 11.7% and 14.5%, respectively. Ablation studies further verify the effectiveness of each component. The proposed framework provides an effective and robust solution for short-term wind power forecasting.

Keywords: short-term wind power forecasting; data quality awareness; global?local feature collaboration; improved Transformer; temporal convolutional network