Machine learning-enhanced dynamic probabilistic safety assessment for station blackout in Bushehr nuclear power plant

Abstract

Dynamic Probabilistic Safety Assessment (D-PSA) faces significant computational challenges when simulating complex accident scenarios with high-fidelity thermal-hydraulic codes, necessitating innovative approaches to balance accuracy and efficiency. This study introduces a machine learning-enhanced Dynamic Event Tree (DET) module that improves risk analysis quantification by integrating deep learning and Support Vector Machine (SVM). This module utilizes a FORTRAN-based submodule for automated generation and parallel execution of thermal hydraulic model input files, enabling concurrent simulation of several of dynamic scenarios while a random sampling strategy ensures comprehensive coverage of failure sequences with minimized training data requirements. In this research a Station blackout (SBO), as initiating event at Bushehr nuclear power plant (BNPP) is considered for thermal-hydraulic model benchmark. Consequently, dynamic effect of diesel generators recovery in different time steps are evaluated. As a result, SBO turns into the loss of offsite power (LOOP) accident. The proposed module investigates LOOP accident for comprehensive risk analysis at BNPP. The results show that the machine learning architecture developed achieves good predictive performance, surpassing at least 97 % accuracy in classifying core damage states and reducing scenario evaluation time. High-resolution dynamic modeling combined with computational feasibility in this module represents fast and effective method in nuclear safety analysis.