Transient level determination with machine learning for pressurized water reactor VVER-1000

Nuclear reactors carry significant risks of severe consequences in the event of accidents. Therefore, the imperative of ensuring their safe operation is paramount. During transient conditions, key parameter time series may fluctuate in ways that are not immediately visible. Therefore, the detection of these transients is essential for preventive measures and protective actions. This work focuses on identification of 53 transient sub-scenarios derived from reactivity insertion via rod withdrawal, steam leak from pressurizer, loss of flow and loss of coolant accidents affecting both hot and cold legs main transient for VVER type reactor. 91 features have been handled for model assessment with 465,465 data points. K-Nearest Neighbor, Decision Tree Classifier, Random Forest Classifier, Gradient Boosting, Logistic Regression, Support Vector Machine, Naïve Bayes and Multilayer Perceptron methods were applied for three different approaches. A one-step, two-steps, and grouped one-step approaches were considered for identification sub-scenarios. In the one-step approach, the exact sub-scenario (e.g. 25% rod withdrawal) was identified. In the two steps approach, the main transients (e.g. rod withdrawal) were identified first and then the sub-scenario (e.g. 25% rod withdrawal) were identified. In the grouped one-step approach, sub-scenarios were grouped (e.g. 20 to 30% rod withdrawal) to increase the accuracy of predictions. While the highest accuracy in one-step approach was 74.66%, two-steps approach had 99.51% main transient identification but 86.44% total accuracy. The grouping of sub-scenarios achieved a less precise but more accurate result with 92.33% accuracy. In conclusion, fast transient identification for VVER type reactors was achieved with two-steps approach.