Data-driven excess Xe reactivity modeling in CANDU nuclear reactor using machine learning methods

The prediction and control of excess Xenon reactivity are critical challenges in ensuring the safe and stable operation of nuclear reactors, particularly in CANDU-type reactors, where xenon-induced oscillations can lead to significant operational risks. In this study, machine learning models are developed to predict excess Xenon reactivity and determine whether it remains within prescribed safety limits. Five regression models – Artificial Neural Network (ANN), Gaussian Process Regression (GPR), Linear Regression (LR), Support Vector Machine (SVM), and Decision Tree (DT), are implemented to predict the magnitude of excess Xenon reactivity, while three classifiers – Logistic Regression (LogR), SVM classifier, and Decision Tree classifier, are employed to classify whether the reactivity levels remain within safe operational thresholds. A sensitivity analysis is performed on the best-performing model by removing individual input parameters to assess their influence on predictions. The results demonstrate that the ML models can provide accurate predictions and real-time classification of reactivity levels, offering potential integration into nuclear reactor control systems. Specifically, the models can enhance existing control strategies, such as those used in Advanced CANDU reactors, by predicting xenon-induced oscillations and enabling automatic adjustments to control mechanisms like zone controllers and mechanical absorbers, thus improving operational stability and safety.