A Unified Framework for Uncertainty-Aware Explainable Artificial Intelligence: A Case Study in Power Quality Disturbance Classification

Uncertainty matters for explanations: Traditional AI explanation methods provide a single attribution map, but this paper shows that when models contain uncertainty in their parameters, the explanations themselves vary. By tracking how explanations change across different model versions (an "explanation distribution"), researchers can better understand when and why an AI system is confident about its reasoning.

Ensemble methods improve explanation reliability: Using a deep ensemble approach, the researchers found that averaged explanations aligned better with known problem regions in power quality disturbance classification than single-model explanations, suggesting that combining multiple models produces more trustworthy explanations—though the benefit varies depending on the specific disturbance type being classified.

Different types of input changes affect explanations differently: When the researchers deliberately altered input data in controlled ways, they discovered that adding noise had a much larger impact on explanations than other changes like amplitude scaling or time shifts, indicating that explanation methods may be sensitive to certain types of data variations in ways practitioners should be aware of.

Theoretical rigor in explanation uncertainty: The framework distinguishes between different sources of error in explanations (posterior-approximation error versus finite-sample error), providing a more principled foundation for understanding when and why explanations might be unreliable—important for responsible deployment where stakeholders need to know the limits of AI reasoning.