Leveraging explainable AI for reliable prediction of nuclear power plant severe accident progression

Abstract

Past severe accidents have highlighted the importance of reducing human error by operators in accident situations. To support operators, machine learning-based accident management support tools have been proposed due to its rapid computation and generalization capabilities. However, the lack of explainability in these models, often perceived as "black-boxes," remains a significant challenge. To address this issue, Explainable AI (XAI) techniques are being integrated across various domains. This study evaluates the applicability of XAI techniques in predicting the state of the OPR1000 reactor during a subset scenario of total-loss-of-component-cooling-water accident with dynamic random failure assumption. Accident scenarios, including various safety component failures and mitigation strategies, were simulated using the Modular Accident Analysis Program (MAAP) code. Two types of XAI techniques-Shapley Additive Explanations (SHAP) and attention-based architectures-are tested alongside conventional black-box models. The results demonstrate that relationships among thermalhydraulic variables can be explained via feature importance, and that the impacts of component failures and mitigation strategies are phenomenologically explainable. Additionally, the study highlights the importance of robust, domain knowledge-based data engineering.