NUCLEAR TECHNIQUES, Volume. 48, Issue 7, 070020(2025)
Prediction method of reactor transient thermal-hydraulic parameters based on Seq2Seq model
Figures & Tables(17)
Fig. 5. Wavelet analysis decomposition results of the highest envelope temperature of China experimental fast reactor
Fig. 6. Comparison of Mean Relative Error (MRE) among four encoder-decoder combinations (color online)
Fig. 7. Comparison of maximum relative error among four encoder-decoder combinations (color online)
Fig. 8. Comparison of RMSE among four encoder-decoder combinations (color online)
Fig. 9. Comparison of the training time among four encoder-decoder combination (color online)
Fig. 10. Comparison of temperature between CNN-LSTM predicted value and real value (color online)
Fig. 11. Schematics of K-fold cross-validation based on time series
Fig. 12. Loss function of CNN-LSTM neural network with 5 folds cross-validation (color online)
Fig. 13. Loss function of CNN-LSTM neural network verified five times by self help method (color online)
Table 1. CEFR core design parameters
View tableView in ArticleTable 1. CEFR core design parameters
参数Parameters 数值Values 功率Power / MW 65 组件内燃料棒数Number of fuel rods in the module 61 活性区高度Height of active zone / mm 450 燃料棒直径Fuel rod diameter / mm 6.0 包壳厚度Shell thickness / mm 0.3 芯块外径/内径Core block outer/inner diameter / mm 5.2/1.6 堆芯进口/出口温度Inlet/outlet temperature of the core / ℃ 360/530 定位绕丝螺距Positioning winding pitch / mm 100
Table 2. Seq2Seq model prediction error of different encoders-decoders
View tableView in ArticleTable 2. Seq2Seq model prediction error of different encoders-decoders
预测变量
Predictor variables
神经网络模型结构（编码器-解码器）
Neural network model structure
(encoder-decoder)
MRE×100% RMSE 最大相对误差
Maximum relative
error / %
模型训练时间
Model training
time / s
包壳最高温度
Maximum shell temperature
CNN-CNN 0.142 1.112 74 0.878 190.038 CNN-LSTM 0.099 0.772 28 0.552 404.467 LSTM-CNN 0.247 1.692 33 1.123 467.045 LSTM-LSTM 0.137 1.117 49 0.855 289.722
Table 3. Prediction error of CNN-LSTM neural network with 5 folds cross-validation
View tableView in ArticleTable 3. Prediction error of CNN-LSTM neural network with 5 folds cross-validation
预测变量
Predictor variables
次数
Number of times
MRE×100% RMSE 最大相对误差
Maximum relative error / %
模型训练时间
Model training time / s
包壳最高温度
Maximum shell temperature
1 0.426 2.722 32 0.796 82.953 2 0.288 2.080 65 0.926 154.763 3 0.153 1.093 82 0.528 247.583 4 0.200 1.404 01 0.549 339.037 5 0.116 0.953 69 0.721 413.203
Table 4. Prediction error of CNN-LSTM neural network verified by five self help methods
View tableView in ArticleTable 4. Prediction error of CNN-LSTM neural network verified by five self help methods
预测变量
Predictor variables
次数
Number of times
MRE×100% RMSE 最大相对误差
Maximum relative error / %
模型训练时间
Model training time / s
包壳最高温度
Maximum shell temperature
1 0.062 3 0.457 61 0.796 481.099 2 0.060 2 0.442 18 0.926 448.932 3 0.080 7 0.589 93 0.528 437.014 4 0.060 4 0.438 45 0.549 446.559 5 0.044 1 0.327 50 0.721 438.514
Tools
Get Citation
Copy Citation Text
Jingyu CHEN, Xiyang LIU, Tengwei YANG, Pengcheng ZHAO, Zijing LIU. Prediction method of reactor transient thermal-hydraulic parameters based on Seq2Seq model[J]. NUCLEAR TECHNIQUES, 2025, 48(7): 070020
Paper Information
Category: Special Issue on The First Academic Annual Conference of the Research Reactor and Innovative Reactor Association of Chinese Nuclear Society and Advanced Nuclear Power System Reactor Engineering
Received: Apr. 21, 2024
Accepted: --
Published Online: Sep. 15, 2025
The Author Email: Zijing LIU (LIUZijing)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20