Uncertainty Quantification of Wind Gust Predictions in the Northeast United States: An Evidential Neural Network and Explainable Artificial Intelligence Approach

Machine learning can improve weather predictions while being honest about limitations: The evidential neural network approach reduced prediction errors by 47% compared to traditional methods, but crucially, it also quantified when predictions were unreliable—addressing a key responsible AI principle of transparency about model confidence.

Uncertainty quantification helps prevent overconfidence in high-stakes decisions: By identifying which predictions needed "cautious interpretation," the model enabled emergency responders and weather forecasters to distinguish between reliable forecasts and those requiring extra scrutiny, potentially preventing poor decisions based on false certainty.

Explainable AI revealed that uncertainty itself contains valuable information: The study found that features contributing to higher uncertainty correlated with storm intensity and spatial gust gradients—meaning the model's "doubt" was actually a signal about genuinely difficult-to-predict conditions, not a flaw to be hidden.

Practical accountability through prediction intervals: The model successfully captured at least 95% of observed gusts at 179 out of 266 stations by providing ranges rather than single predictions, giving stakeholders concrete information about risk levels for emergency planning rather than false precision.

Responsible AI requires both accuracy and interpretability: The combination of improved performance (47% error reduction) with explainability techniques and uncertainty quantification demonstrates that responsible AI isn't just about making better predictions—it's about making predictions stakeholders can actually trust and act upon safely.