Data quality enhancement for wind farms using credibility-guided multi-view representation learning

Abstract

Reliable meteorological measurements are essential for wind power forecasting and renewable energy integration. However, under high-impact weather conditions, environmental disturbances, missing observations, and nonlinear inter-variable coupling often degrade data quality, resulting in uncertainty propagation and unreliable forecasting performance. To address these challenges, a credibility-oriented data quality enhancement framework is proposed for multivariate wind farm meteorological data. First, correlation-aware reconstruction is employed to restore consistency among coupled meteorological variables and alleviate data-quality degradation caused by measurement disturbances. Subsequently, a SHAP-guided multi-view learning strategy is introduced to decouple heterogeneous credibility patterns embedded in multivariate measurements. On this basis, a credibility-enhanced variational autoencoder (C-VAE) is developed to suppress uncertainty propagation during latent representation learning through adaptive credibility weighting. Furthermore, a hierarchical credibility discrimination strategy integrating density-aware clustering, temporal continuity, and information entropy is designed to distinguish pseudo-outliers from true abnormal observations. Experimental results on real-world datasets demonstrate that the proposed framework effectively improves data quality, reduces false alarms, and enhances forecasting robustness.