Risk Modeling and Optimization Using Machine Learning Algorithms for PSA Applications

According to the latest reports from the International Atomic Energy Agency, the global nuclear power plant (NPP) fleet is aging, requiring enhanced risk estimation and safety management tools.Additionally, international organizations are promoting the use of artificial intelligence to improve NPP operation and safety, as highlighted by the U.S. Nuclear Regulatory Commission in its Strategic Plan for 2023-2027.This context presents an opportunity to incorporate artificial intelligence and machine learning techniques into risk-informed applications.This work presents a methodology that integrates a metamodel for predicting Core Damage Frequency (CDF) with an optimization method using genetic algorithms to reduce CDF, thereby enhancing plant safety.The proposed metamodel predicts CDF using reliability parameters related to component failure modes and test intervals (TI) of standby equipment as explanatory variables.Traditionally, TIs are set as fixed values grouped by periodicity (e.g., weekly, monthly, yearly).This study introduces additional levels-higher and lower than standard values-to assess their impact on NPP risk, using a Fractional Factorial Design to efficiently generate a representative dataset.Several metamodels were trained and evaluated, with the best-performing one selected to replace conventional Probabilistic Safety Assessment (PSA) models.The genetic algorithm was then implemented to find the optimal combination of TI values, minimizing the mean CDF.The use of the metamodel drastically reduces computational effort compared to resolving large event and fault trees, which is needed to speed up the genetic algorithm's computational time.The results demonstrate that this methodology enhances NPP safety by providing a powerful tool for risk management.Future developments could include analyzing the impact of maintenance on reliability parameters, and consequently, on the CDF.