Comparative Study of Neural Network-based Source Term Inversion Methods for Nuclear Accidents
-
摘要: 核事故源项是核事故后果评价的重要依据。准确、快速地获取源项对于核事故应急响应决策的制定具有重大意义。本文利用径向基神经网络、误差反向传播神经网络和深度前馈神经网络3种典型的神经网络算法，对核事故源项进行反演估算对比。针对单核素和多核素2种源项反演情况，以放射性核素的释放速率为反演目标，利用课题组开发的放射性大气扩散模拟程序RADC进行模拟实验获取监测数据。误差分析和反演时间的对比结果表明，基于深度前馈神经网络搭建的核素释放速率反演模型表现出优异的性能。对于单核素的释放速率反演，在十折交叉验证的方式下，反演相对误差在1.7%~3.0%以内，平均绝对百分比误差为2.2%；对于多核素情况，深度前馈神经网络的反演平均绝对百分比误差为8.05%，其反演估算准确率和稳定性均优于误差反向传播神经网络和径向基神经网络。Abstract: The nuclear accident source term is an important basis for the evaluation of the consequences of nuclear accidents. The accurate and fast acquisition of the source term is of great significance for the development of nuclear accident emergency response decision-making. This study utilizes three typical neural network algorithms, namely radial basis neural network, back propagation neural network and deep feed-forward neural network, to compare the inverse estimation of nuclear accident source terms. For the inversion of single nuclide and multi-nuclide source term, taking the release rate of radionuclides as the inversion target, the simulation experiment is conducted to obtain the monitoring data by using the radioactive atmospheric dispersion simulation code RADC developed by the group. The results of error analysis and inversion time comparison show that the nuclide release rate inversion model based on deep feedforward neural network exhibits excellent performance. For the release rate inversion of single nuclide, the relative error of the inversion is within 1.7%~3.0% and the average absolute percentage error is 2.2% under the ten-fold cross-validation. For the multi-nuclide case, the average absolute percentage error of the inversion of deep feedforward neural network is 8.05%. Its inversion estimation accuracy and stability are better than that of the BP neural network and radial basis neural network.
-
Key words:
- Nuclear accidents /
- Source term inversion /
- Neural network /
- Radionuclide release rate
-
图 1 基于神经网络的核事故源项反演建模过程
Figure 1. Neural Network-based Modeling Process of Nuclear Accident Source Term Inversion
图 3 BPNN模型不同超参数调整结果
Figure 3. Results of Different Hyperparameter Tuning of BPNN Model
图 4 DFNN不同激活函数模型性能
Figure 4. DFNN Model Performance with Different Activation Functions
图 5 3种模型反演预测散点分布结果
Figure 5. Results of Three Model Inversions for Predicting Scatter Distribution
图 6 不同噪声等级下3个模型预测误差变化对比
Figure 6. Comparison of the Changes in Prediction Errors of the Three Models under Different Noise Levels
图 7 4种核素释放速率的反演APE误差箱型图
Figure 7. Box Plots of APE of Inversion for Four Nuclide Release Rates
表 1 单核素部分数据集示例
Table 1. Examples of Partial Datasets for Single Nuclide
输入参数 输出目标 大气稳定度 释放
高度/m风速/
(m·s−1)降水量/
(mm·h−1)混合层
高度/m监测站点核素空气浓度/(104 Bq·m−3) 131I释放速率/(1011 Bq·s−1) 0.5 km 1 km 2 km 5 km 10 km D 45 7.5 0 560 160.00 173.00 87.00 27.10 11.20 2.99 A 13 1.9 1 750 799.00 201.00 50.10 13.20 6.13 5.45 B 12 2.3 5 810 2020.00 498.00 118.00 15.90 5.43 7.49 
下载: 导出CSV
表 2 RBFNN模型不同参数组合均方误差
Table 2. Mean Squared Error of Different Parameter Combinations in the RBFNN Model
参数组合 （1000，1） （3000，1） （3000，2） （5000，1） （5000，2） MSE 0.019 0.002 0.007 0.001 0.005
下载: 导出CSV
表 3 不同模型在测试集上的反演精度和推理时间
Table 3. Inversion Accuracy and Inference Time for Different Models on the Test Set
模型 MSE RMSE MAE MAPE/% 时间/min RBFNN 0.0011 0.0335 0.0162 4.05 16.480 BPNN 0.0053 0.0729 0.0494 12.99 0.016 DFNN 0.0003 0.0181 0.0096 2.07 0.027
下载: 导出CSV
表 4 不同模型的训练集误差
Table 4. Training Set Errors for Different Models
模型 RBFNN BPNN DFNN MAPE/% 2.81 3.06 1.78
下载: 导出CSV
表 5 不同误差上限测试样本比例
Table 5. Proportion of Test Samples with Different Error Limits
APE/% <0.1 <0.2 <0.5 <1 <2 <3 <4 <5 样本比例/% 8.66 17.04 39.47 63.56 82.42 88.80 91.84 93.74
下载: 导出CSV
表 6 多核素释放速率反演准确性评价
Table 6. Accuracy Evaluation of Multi-nuclide Release Rate Inversion
评价指标模型 MSE MAE MAPE/% RBFNN 0.0143 0.0783 13.21 BPNN 0.1515 0.2363 29.39 DFNN 0.0097 0.0486 8.05
下载: 导出CSV
-
[1] SAUNIER O, MATHIEU A, DIDIER D, et al. An inverse modeling method to assess the source term of the Fukushima Nuclear Power Plant accident using gamma dose rate observations[J]. Atmospheric Chemistry and Physics, 2013, 13(22): 11403-11421. doi: 10.5194/acp-13-11403-2013 [2] CHAI T F, DRAXLER R, STEIN A. Source term estimation using air concentration measurements and a Lagrangian dispersion model—experiments with pseudo and real cesium-137 observations from the Fukushima nuclear accident[J]. Atmospheric Environment, 2015, 106: 241-251. doi: 10.1016/j.atmosenv.2015.01.070 [3] DUMONT LE BRAZIDEC J, BOCQUET M, SAUNIER O, et al. Bayesian transdimensional inverse reconstruction of the Fukushima Daiichi caesium 137 release[J]. Geoscientific Model Development, 2023, 16(3): 1039-1052. doi: 10.5194/gmd-16-1039-2023 [4] TICHÝ O, ŠMÍDL V, HOFMAN R, et al. Bayesian inverse modeling and source location of an unintended 131I release in Europe in the fall of 2011[J]. Atmospheric Chemistry and Physics, 2017, 17(20): 12677-12696. doi: 10.5194/acp-17-12677-2017 [5] 唐秀欢,李华,包利红. 核事故实时释放量集合卡尔曼滤波反演算法研究[J]. 原子能科学技术,2014, 48(S1): 415-420. [6] KOVALETS I, ANDRONOPOULOS S, HOFMAN R, et al. Advanced method for source term estimation and status of its integration in JRODOS[J]. Radioprotection, 2016, 51(HS2): S121-S123. [7] SUN S D, LI X P, LI H, et al. Site-specific (Multi-scenario) validation of ensemble Kalman filter-based source inversion through multi-direction wind tunnel experiments[J]. Journal of Environmental Radioactivity, 2019, 197: 90-100. doi: 10.1016/j.jenvrad.2018.12.005 [8] DI RONCO A, GIACOBBO F, CAMMI A. A Kalman filter-based approach for online source-term estimation in accidental radioactive dispersion events[J]. Sustainability, 2020, 12(23): 10003. doi: 10.3390/su122310003 [9] GENG X B, XIE Z H, ZHANG L J, et al. An inverse method to estimate emission rates based on nonlinear least-squares-based ensemble four-dimensional variational data assimilation with local air concentration measurements[J]. Journal of Environmental Radioactivity, 2018, 183: 17-26. doi: 10.1016/j.jenvrad.2017.12.004 [10] 刘蕴,刘新建,李红,等. 截断总体最小二乘变分核事故源项反演数值研究[J]. 核动力工程,2019, 40(1): 120-125. [11] 宁莎莎,蒯琳萍. 混合遗传算法在核事故源项反演中的应用[J]. 原子能科学技术,2012, 46(S1): 469-472. [12] MA D L, ZHANG Z X. Contaminant dispersion prediction and source estimation with integrated Gaussian-machine learning network model for point source emission in atmosphere[J]. Journal of Hazardous Materials, 2016, 311: 237-245. doi: 10.1016/j.jhazmat.2016.03.022 [13] 沈泽亚,郎建垒,程水源,等. 典型耦合优化算法在源项反演中的对比研究[J]. 中国环境科学,2019, 39(8): 3207-3214. doi: 10.3969/j.issn.1000-6923.2019.08.010 [14] LI X P, SUN S D, HU X F, et al. Source inversion of both long- and short-lived radionuclide releases from the Fukushima Daiichi nuclear accident using on-site gamma dose rates[J]. Journal of Hazardous Materials, 2019, 379: 120770. doi: 10.1016/j.jhazmat.2019.120770 [15] JANG S, PARK J, LEE H H, et al. Comparative study on gradient-free optimization methods for inverse source-term estimation of radioactive dispersion from nuclear accidents[J]. Journal of Hazardous Materials, 2024, 461: 132519. doi: 10.1016/j.jhazmat.2023.132519 [16] 凌永生,侯闻宇,贾文宝,等. 基于BP神经网络的核事故源项反演方法研究[J]. 中国安全科学学报,2014, 24(8): 21-25. [17] 侯闻宇,凌永生,赵丹,等. BP神经网络反演核事故源项中重要参数的研究[J]. 南京航空航天大学学报,2015, 47(5): 778-784. [18] 侯闻宇,凌永生,赵丹,等. GA-BP算法应用于核事故源项反演的研究[J]. 安全与环境学报,2016, 16(6): 24-28. [19] 赵丹,凌永生,侯闻宇,等. 基于BP神经网络的核事故多核素源项反演方法[J]. 南京航空航天大学学报,2016, 48(1): 130-135. [20] LING Y S, YUE Q, CHAI C J, et al. Nuclear accident source term estimation using kernel principal component analysis, particle swarm optimization, and backpropagation neural networks[J]. Annals of Nuclear Energy, 2020, 136: 107031. doi: 10.1016/j.anucene.2019.107031 [21] KATATA G, OTA M, TERADA H, et al. Atmospheric discharge and dispersion of radionuclides during the Fukushima Dai-ichi nuclear power plant accident. Part I: Source term estimation and local-scale atmospheric dispersion in early phase of the accident[J]. Journal of Environmental Radioactivity, 2012, 109: 103-113. doi: 10.1016/j.jenvrad.2012.02.006 [22] 淳思琦,冯欢,张安妮,等. 基于软注意力GRU模型的堆芯瞬态热工水力参数预测方法研究[J]. 核技术,2024, 47(1): 124-132. doi: 10.11889/j.0253-3219.2024.hjs.47.010603 -
计量
- 文章访问数: 24
- HTML全文浏览量: 8
- PDF下载量: 7
- 被引次数: 0
