🤖 AI Summary
This study addresses the common conflation of epistemic uncertainty (EU) and aleatoric uncertainty (AU) in existing wind power forecasting methods, which undermines decision reliability. Building upon the law of total variance, the authors propose a unified framework combining heteroscedastic neural networks with Bayesian posterior approximation to explicitly disentangle total predictive uncertainty into its AU and EU components. A β-negative log-likelihood (β-NLL) loss is introduced to balance mean and variance learning. The work presents the first evaluation framework for uncertainty disentanglement in wind power prediction that operates without ground-truth uncertainty labels. Through synthetic experiments, data attribute analysis, and scaling studies, the approach demonstrates theoretically consistent responses of AU and EU to noise structure, distribution shifts, and training set size on both synthetic and real-world SCADA data, validating its theoretical soundness and practical utility.
📝 Abstract
Accurate wind power forecasting requires reliable uncertainty quantification, yet most existing methods report a single predictive uncertainty that conflates epistemic and aleatoric sources. This paper applies the law of total variance to the joint setting of heteroscedastic neural network regression and Bayesian posterior approximation, deriving an explicit decomposition of total uncertainty (TU) into aleatoric (AU) and epistemic (EU) components. The resulting estimators are compatible with standard posterior-approximation methods and with $β$-NLL training to regulate the mean--variance learning trade-off. A wind power--specific evaluation framework is proposed to validate disentanglement without access to ground-truth uncertainty labels, comprising three modules: controlled synthetic experiments to verify responses to heteroscedastic noise and distribution shift; data-property--driven validation on a real-world wind turbine SCADA dataset; and dataset-size scaling experiments to examine the predicted asymptotic behavior of EU. Across synthetic and real-world experiments, the decomposed AU and EU components respond in theoretically consistent directions to noise structure, distributional shift, and training-scale variation, supporting the theoretical consistency and operational utility of the proposed decomposition and evaluation protocol.