Artificial Intelligence-Enabled Risk Forecasting in Air Traffic Control: An Interpretable Machine-Learning Framework for Safety Management

This study aimed to develop and evaluate a human-interpretable artificial intelligence (AI) model for short-term operational risk forecasting in air traffic control (ATC). To do so, it compared logistic and Poisson regression models with random forest and gradient boosting classifiers using a transparently generated synthetic dataset designed to reflect realistic workload, weather, staffing, and sector complexity conditions. Model performance was assessed through cross-validation, calibration analysis, sensitivity to class imbalance, and decision-curve analysis. Among the tested approaches, gradient boosting achieved the best predictive performance, with an area under the curve of 0.93, and provided the most reliable probability estimates, outperforming the regression-based baseline models. Explainability analysis using Shapley additive explanations showed that the most influential predictors were controller workload, weather severity, sector complexity, and staffing ratio, which is consistent with established human factors theory. Decision-curve analysis also indicated measurable operational benefit at realistic alert thresholds, supporting potential applications in dynamic staffing and flow management. These findings suggest that responsible AI can strengthen safety management systems by providing accurate, transparent, and reproducible risk forecasts while supporting regulatory expectations for documentation, calibration, and interpretability.