Pressurized Water Reactor Transient Detection With Artificial Intelligence to Support Reactor Operators

Abstract

Nuclear reactors are subject to strict safety standards due to the critical nature of operational safety. Detecting transients as fast and accurately as possible is essential to reactor safety especially to reduce the human error of operators. In order to enhance this process, artificial intelligence (AI) offers strong opportunities. The robust AI systems that are validated and verified with current reactors’ data will increase the trust of AI systems for next‐generation small modular reactors. In this study, five transients of light water reactors (reactivity insertion with rod withdrawal, steam leak from pressurizer, pump trip, loss of coolant (LOCA) in the hot leg, and LOCA in the cold leg) were selected to generate 53 subscenarios by using the description of VVER‐1000 type reactor. For the 93 features of the system selected based on reactor operation, simulations produced 475.695 data points stored in 5115 rows, 4085 of which were committed to training and 1030 to the validation process. The three methods for reinforced machine learning to detect transients, K‐nearest neighbors, decision tree classifier, and random forest classifier, have been implemented. The success and failure rates of the models have also been analyzed and presented. When accuracy, precision, recall, and F1‐score are compared together, the random forest method showed the best performance.