Feature-driven dynamic non-crossing quantile ensemble learning for reliable probabilistic wind power forecasting

Wind power forecasting is crucial for grid stability and integrating renewables into power systems. While deterministic methods have been widely adopted, they fail to characterize wind power’s inherent variability and quantify uncertainties. Existing probabilistic approaches face significant limitations, with parametric methods constrained by specific distributional assumptions and non-parametric methods like multi-quantile regression suffering from quantile crossing issues. This study proposes a novel two-stage non-crossing quantile ensemble learning framework for generating reliable probabilistic wind power forecasts. In the first stage, an iTransformer-based non-crossing quantile regression model with parameter-sharing mechanism generates diverse base quantile forecasts, enhancing computational efficiency while ensuring monotonic quantile relationships. In the second stage, an attention-enhanced U-Net-based ensemble model refines and calibrates these quantiles through multi-scale feature fusion via attention mechanisms in skip connections, effectively integrating temporal and quantile-specific features through cross-attention. Experimental validation on three datasets demonstrates the framework’s superiority, achieving pinball loss reductions of 7%–14% compared to benchmarks while maintaining coverage probability within 0.5% of nominal confidence levels. The proposed model advances probabilistic forecasting by dynamically integrating feature-driven patterns with distributional uncertainty, offering enhanced reliability for grid operators in managing wind energy variability and supporting robust decision-making under uncertainty.