Machine learning-based intelligent models for accurate prediction of key parameters in pressurized water reactors with operational data

Abstract

In this research, a quick and accurate prediction model has been developed for the key parameters of pressurized water reactors (PWRs) by applying the residual module neural network and single-channel convolutional neural network. The improved Gen-II PWRs, including CNP1000 and CPR1000, constitute the principal operating units in China and have accumulated a great many measurement values for critical boron concentration (CBC) and power distributions. To improve the prediction accuracy of these key parameters, this study utilizes approximately 19,971 sets of measurement values for CBC and 618 sets of data for 3-D power distribution. These datasets are combined with core state parameters provided by the PWR-core analysis code LOCUST/SPARK, which include factors such as fuel-assembly loading, assembly-averaged burnup, and poison-isotope distribution, among others, to train the intelligent models. As a result the intelligent model linking simulated core state parameters to measured key parameters has been established, making it possible to accurately and rapidly predict the measured CBC and power distribution. The verification tests indicate that the predicted CBCs have an average error of about 3.32 ppm with a 95 % confidence interval of [-6.0 ppm, 7.5 ppm], taking only 1.8 ms for each prediction. The predicted axial-power distributions and axial offset (AO) have an average relative error of about 0.46 % and 0.069 %, with 95 % confidence intervals of [-1.4 %, 1.4 %] and [-0.124 %, 0.071 %], respectively, and require only 8.3 ms for each prediction. Additionally, the predicted radial core power distributions have an average error of about 0.31 % with a 95 % confidence interval of [-0.92 %, 0.92 %] and require only 3.2 ms for each prediction. These results demonstrate that the intelligent models established in this research are capable of predicting key parameters with high accuracy and efficiency, which provides a novel technical approach for further enhancing the operational economy and safety of PWR nuclear power plants.