基于LSTM的核电传感器多特征融合多步状态预测

张思原; 卢忝余; 曾辉; 徐春; 张倬; 黄擎宇; 张尧毅; 王媛美

doi:10.13832/j.jnpe.2021.04.0208

基于LSTM的核电传感器多特征融合多步状态预测

doi: 10.13832/j.jnpe.2021.04.0208

张思原^{1, 2,},
卢忝余²,
曾辉^{1, 2},
徐春^{1, 2},
张倬^{1, 2},
黄擎宇^{1, 2},
张尧毅^{1, 2},
王媛美^{1, 2}

1.
中国核动力研究设计院反应堆系统设计技术国家重点实验室，成都， 610213
2.
中国核动力研究设计院，成都， 610213

详细信息

作者简介:
张思原（1992—），女，工程师，现主要从事反应堆系统软件研发方面工作，E-mail: siyuan_Tooru@163.com

中图分类号: TL363
计量
- 文章访问数: 340
- HTML全文浏览量: 183
- PDF下载量: 67
- 被引次数: 0
出版历程
- 收稿日期: 2020-06-04
- 修回日期: 2021-01-07
- 刊出日期: 2021-08-15

Multi-Feature Fusion Multi-Step State Prediction of Nuclear Power Sensor Based on LSTM

Zhang Siyuan^{1, 2
,},
Lu Tianyu²,
Zeng Hui^{1, 2},
Xu Chun^{1, 2},
Zhang Zhuo^{1, 2},
Huang Qingyu^{1, 2},
Zhang Yaoyi^{1, 2},
Wang Yuanmei^{1, 2}

1.
Science and Technology on Reactor System Design Technology Laboratory, Nuclear Power Institute of China, Chengdu, 610213, China
2.
Nuclear Power Institute of China, Chengdu, 610213, China

摘要

摘要: 针对核电工况参数预测的问题，利用核电站传感检测系统采集的大量时间序列，提出了基于长短时记忆网络（LSTM）的多特征融合多步状态预测模型。以某核电厂实时参数系统采集到的SG1蒸汽压力传感数据为研究对象，首先针对数据缺失、采样时标不一致问题进行数据预处理，然后完成基于LSTM的多特征融合多步状态预测模型的结构设计与建模，最后将本文提出的预测模型与循环神经网络（RNN）、门控循环单元（GRU）、本文模型-全连接层1以及单变量LSTM等多步预测模型进行比较。实验结果表明，本文提出的预测模型的拟合性能和预测性能整体最优，同时也验证了基于LSTM模型的深度学习方法在核电站运行安全保障领域的适用性。
- 核电安全 /
- 时间序列数据 /
- 状态预测 /
- 深度学习
Abstract: Aiming at the problem in the prediction of nuclear power working condition parameter, this paper uses a large number of time series collected by the nuclear power plant sensor detection system to propose a multi-feature fusion multi-step state prediction model based on long short-term memory network (LSTM). This paper takes the SG1 steam pressure sensor data collected by the real-time parameter system of nuclear power plants as the research object. Firstly, the data is preprocessed for the problems of missing data and inconsistent sampling time scales, and then the structural design and modeling of the multi-feature fusion multi-step state prediction model is completed based on LSTM. Finally, the prediction model proposed in this paper is compared with the multi-step prediction models such as Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Model-S1 layer and univariate LSTM. Experimental results show that the fitting performance and prediction performance of the prediction model proposed in this paper are with overall optimization, and it also verifies the applicability of the deep learning method based on the LSTM model in the field of nuclear power plant operation safety assurance.
- Nuclear power safety /
- Time series data /
- State prediction /
- Deep learning

HTML全文

图 1 基于LSTM的多特征融合多步状态预测模型示意图　　　

S₁—全连接层1的激活函数输出向量；S₂—全连接层2的激活函数输出向量，A—LSTM层的输出向量

Figure 1. Schematic Diagram of Multi-Feature Fusion Multi-Step State Prediction Model Based on LSTM

下载: 全尺寸图片幻灯片

图 2 本文模型测试集预测结果

Figure 2. Prediction Results of Model Test Set in this Paper

下载: 全尺寸图片幻灯片

图 3 本文模型预测效果图（40个采样点）

Figure 3. Model Prediction Effect Chart of this Paper (40 Sampling Points)

下载: 全尺寸图片幻灯片

表 1 SG1蒸汽压力关联特征

Table 1. SG1 Steam Pressure Sensor Related Features

测点序号	参数名称	测点序号	参数名称
1	SG1蒸汽压力	10	上充流量
2	SG1蒸汽压力	11	SG2给水流量
3	SG1蒸汽压力	12	SG3给水流量
4	SG1给水流量	13	SG2蒸汽压力
5	SG1蒸汽流量	14	SG3蒸汽压力
6	SG1窄量程水位	15	SG2蒸汽流量
7	SG1宽量程水位	16	SG3蒸汽流量
8	最大平均温度	17	2环路流量
9	1环路流量	18	3环路流量

下载: 导出CSV

表 2 数据集统计结果

Table 2. Data Set Statistics

测点序号	参数	2018年1月总测量数	缺失统计数量
1	SG1蒸汽压力	676731	2632
2	SG1蒸汽压力	1071559	206
3	SG1蒸汽压力	956924	561
4	SG1给水流量	4696573	0
5	SG1蒸汽流量	1175725	0
6	SG1窄量程水位	3335024	0
7	SG1宽量程水位	2401034	0
8	最大平均温度	4274564	171
9	1环路流量	4233016	0
10	上充流量	552510	5903
11	SG2给水流量	4620442	0
12	SG3给水流量	4557953	0
13	SG2蒸汽压力	887091	835
14	SG3蒸汽压力	425889	7954
15	SG2蒸汽流量	1050003	0
16	SG3蒸汽流量	859573	19
17	2环路流量	4324677	0
18	3环路流量	4242478	0

下载: 导出CSV

表 3 重要超参数设置

Table 3. Important Hyperparameter Settings

参数名	参数描述	设置值
batch_size	批处理大小	256
learning_rate	学习率	0.001
epoch	迭代次数	50
s1_num	特征融合层神经元个数	16
lstm_num	lstm层隐藏单元个数	128
s2_num	输出变换层隐藏单元个数	取值与k相同（k=10）
m	输入数据的时间步数	25

下载: 导出CSV

表 4 不同模型预测结果比较

Table 4. Comparison of Prediction Results of Different Models

模型	SG1蒸汽压力RMSE×10²值
模型	t+1	t+2	t+3	t+4	t+5	t+6	t+7	t+8	t+9	t+10	平均值
本文模型	0.36101	0.36068	0.36851	0.37483	0.38161	0.39794	0.39587	0.39634	0.40058	0.41057	0.38479
RNN	0.35700	0.36627	0.37435	0.37549	0.37685	0.38800	0.40119	0.40595	0.41155	0.41172	0.38684
GRU	0.37124	0.37457	0.38114	0.38288	0.38405	0.40277	0.41428	0.40911	0.40475	0.41421	0.39390
本文模型-全连接层1	0.39449	0.40396	0.40544	0.39090	0.39658	0.42582	0.43874	0.42343	0.41843	0.42890	0.41267
单变量LSTM	0.40054	0.40575	0.40646	0.40590	0.41001	0.42589	0.42917	0.43273	0.43023	0.44318	0.41899
t+1—从任意时刻t开始后的第1×30 s时刻；其余类同

下载: 导出CSV

参考文献(15)

[1]	BOX G E P, JENKINS G M, REINSEL G C, et al. Time series analysis: forecasting and control, 5th edition[J]. Journal of the Operational Research Society, 2015, 22(2): 199-201.
[2]	TONG W M, LI Y J, SHANY Z. Data mining of time series based on wavelet analysis[J]. Computer Engineering, 2008, 34(1): 26-29.
[3]	潘迪夫,刘辉,李燕飞. 基于时间序列分析和卡尔曼滤波算法的风电场风速预测优化模型[J]. 电网技术,2008, 32(7): 82-86.
[4]	NOTTINGHAM Q J, COOK D F. Local linear regression for estimating time series data[J]. Computational Statistics & Data Analysis, 2001, 37(2): 209-217.
[5]	SAPANKEVYCH N I, SANKAR R. Time series prediction using support vector machines[J]. IEEE Computational Intelligence Magazine, 2009, 4(2): 24-38. doi: 10.1109/MCI.2009.932254
[6]	LIN K, LIN Q, ZHOU C, et al. Time series prediction based on linear regression and SVR[C]//Third International Conference on Natural Computation (ICNC 2007). Haikou: IEEE Computer Society, 2007: 688-691.
[7]	HU T S, LAM K C, NG S T. River flow time series prediction with a range-dependent neural network[J]. International Association of Scientific Hydrology Bulletin, 2001, 46(5): 729-745. doi: 10.1080/02626660109492867
[8]	ZHANG G P. A neural network ensemble method with jittered training data for time series forecasting[J]. Information Sciences, 2007, 177(23): 5329-5346. doi: 10.1016/j.ins.2007.06.015
[9]	杨祎玥,伏潜,万定生. 基于深度循环神经网络的时间序列预测模型[J]. 计算机技术与发展,2017(3): 35-38, 43.
[10]	PARK D C. Prediction of time series data using multiresolution-based bilinear recurrent neural network[C]//International Conference on Electronic Computer Technology. Macau: IEEE Computer Society, 2009.
[11]	JIE W, JUN W, WEN F, et al. Financial time series prediction using elman recurrent random neural networks[J]. Computational Intelligence and Neuroscience, 2016(2016): 1-14.
[12]	NGAI E, DRESSLER F, LEUNG V, et al. Guest editorial special section on internet-of-things for smart cities and urban informatics[J]. IEEE Transactions on Industrial Informatics, 2017, 13(2): 748-750. doi: 10.1109/TII.2017.2675379
[13]	龚安, 马光明, 郭文婷, 等. 基于LSTM循环神经网络的核电设备状态预测[J]. 计算机技术与发展,2019(10): 41-45. doi: 10.3969/j.issn.1673-629X.2019.10.009
[14]	ZAGHLOUL M A S, HASSAN A M, CARPENTER D, et al. Optical sensor behavior prediction using LSTM neural network[C]//2019 IEEE Photonics Conference (IPC). US: IEEE, 2019.
[15]	ZHANG A, TENG J, JU Y, et al. Thermal power prediction of nuclear reactor core based on LSTM[C]// 2019 Chinese Automation Congress (CAC). Hangzhou, China: IEEE, 2020.