An interpretable machine learning approach for classifying nuclear accident release categories

Rapid and reliable identification of accident types is essential for guiding emergency response to nuclear incidents, especially when on-site instrumentation may be compromised. This study proposes an accident classification approach based on off-site monitoring data for a 2905 MW pressurized water reactor at end-of-life, employing the RADC radioactive atmospheric dispersion model to generate simulated off-site data for categorizing nine nuclear power plant accident release categories (PWR1-9) specified in the WASH-1400. The research evaluated classification performance of KNN, RF, and XGBoost, with XGBoost achieving the highest accuracy of 99.58 %. By integrating Harris Hawks Optimization with XGBoost, prediction accuracy was enhanced to 99.97 %. To ensure transparency and trust in real-time decision support, we prioritized interpretability analysis: Partial Dependence Plot and SHAP value reveal that gamma dose rate monitoring data, wind speed and release height drive model predictions. These insights highlight the importance of maintaining data quality for these key parameters and support the integration of high-performance classification with rigorous interpretability in nuclear emergency management.