Leveraging machine learning for accurate DNBR prediction using python

Abstract

This study investigates the viability of using Python scikit-learn packages, specifically regression techniques, to forecast the departure from nucleate boiling ratio (DNBR) using previously published datasets as reference. These datasets were generated using the COBRA-IV computer code under specific operational conditions for pressurized water reactors, and are categorized into two groups: steady-state, encompassing cases 1 and 2, and transient-state which includes both the loss of flow accident (LOFA) and rod cluster control assembly (RCCA) withdrawal scenarios. Due to the dataset’s small size, this study also addressed the difficulties encountered with enhancing the accuracy of regression models. Several approaches were investigated in an attempt to improve the regression’s accuracy, these include feature selection and hyperparameters tuning analysis for each regression model. The use of noise perturbation with a realistic standard deviation proved to be the most successful strategy for increasing regression accuracy, especially with small datasets, by interpolating the gaps between observations thus increasing their size. This approach’s algorithm was evaluated using metrics such as R2 score, mean square error (MSE) and mean absolute error (MAE). The evaluation metrics demonstrated that the DTReg model, after incorporating noise, achieved significantly higher accuracy compared to previously referenced artificial neural network (ANN) models, highlighting the substantial impact of noise perturbation on refining the model’s predictive capabilities. Overall, this approach not only improved performance for the steady-state cases but also enhanced accuracy for transient datasets, ultimately demonstrating the effectiveness of noise perturbation in overcoming the limitations of small datasets.