基于LSTM网络预测的水轮机机组运行状态检测

doi:10.6040/j.issn.1672-3961.0.2018.240

摘要/Abstract

摘要：

利用长短期记忆(long short-term memory, LSTM)网络对水轮机机组的运行状态进行预测。对水轮机机组的流式监测数据进行标准化处理,并利用滑动窗口技术将数据转换为LSTM网络训练所需的训练数据集与测试数据集;给出LSTM预测模型结构,并通过调节网络层数、隐层神经元数目等参数对模型进行优化,建立水轮机机组的时间序列数据预测模型。经试验分析验证,与其它模型相比,基于多测点的多元长短期记忆网络预测模型具备更高的预测精度,并基于改进的雷达图分析法计算健康偏离度,成功地检测出某水电厂5号水轮机机组5月末的数据出现异常,验证了模型的有效性。

关键词: 水轮机, 数据预测, 长短期记忆网络

Abstract:

Long short-term memory (LSTM) networks was adapted to make accurate prediction of the unit's operation status. The streaming monitoring data of the turbine unit was standardized, and the sliding window technology was used to convert the data into the training data set and test data set for LSTM network training. The LSTM prediction model structure was given, and the structure of LSTM prediction model was fine-tuned, such as the number of network layers and the number of hidden layer neurons. The time series data prediction model of the hydro turbine unit was established. The experimental analysis proved that the multi-measurement-based LSTM network prediction model had higher prediction accuracy than other models, which calculated the health deviation based on the improved radar image analysis method and successfully detected the abnormality of the No. 5 hydraulic turbine unit of a hydropower plant at the end of May, and verified the validity of the model.

Key words: hydraulic turbine, data forecasting, long short-term memory networks

中图分类号:

TP29

陈畅,李晓磊,崔维玉. 基于LSTM网络预测的水轮机机组运行状态检测[J]. 山东大学学报 (工学版), 2019, 49(3): 39-46.

Chang CHEN,Xiaolei LI,Weiyu CUI. Hydraulic turbine operation status detection based on LSTM network prediction[J]. Journal of Shandong University(Engineering Science), 2019, 49(3): 39-46.

图/表 12

图1

图2

图3

图4

表1

表2

表3

表4

图5

表5

图6

图7

参考文献 22

1	DAI H, MAO J, CAO G. High intensity concrete construction technology and management mode of large-scale hydropower project[C]//International Conference on Management and Service Science. Wuhan, China: IEEE, 2011: 1-4.
2	林道远, 林建兴. 水轮机运行稳定性的研究与展望[J]. 山东工业技术, 2018, (1): 35- 35.
	LIN Daoyuan , LIN Jianxing . Research and prospect of turbine operation stability[J]. Shandong Industry Science and Technology, 2018, (1): 35- 35.
3	LI C , CHANG L , HUANG Z , et al. Parameter identification of a nonlinear model of hydraulic turbine governing system with an elastic water hammer based on a modified gravitational search algorithm[J]. Engineering Applications of Artificial Intelligence, 2016, 50 (C): 177- 191.
4	曹剑绵, 陈昌林. 大型水轮机转轮异常振动及叶片裂纹分析[J]. 西南交通大学学报, 2002, 37 (增刊1): 68- 72.
	CAO Jianmian , CHEN Changlin . Analysis of abnormal vibration and blade cracks of a large hydraulic turbine Runner[J]. Journal of Southwest Jiao Tong University, 2002, 37 (Suppl.1): 68- 72.
5	郭峰, 王斌, 刘敏. 基于BP神经网络的时间序列预测研究[J]. 价值工程, 2010, 29 (35): 128- 129. doi: 10.3969/j.issn.1006-4311.2010.35.094
	GUO Feng , WANG Bin , LIU Min . Research on time series prediction based on BP neural network[J]. Value Engineering, 2010, 29 (35): 128- 129. doi: 10.3969/j.issn.1006-4311.2010.35.094
6	文新辉. 时间序列神经网络预测方法[J]. 电子与信息学报, 1994, 16 (5): 456- 462.
	WEN Xinhui . Time series neural network prediction method[J]. Journal of Electronics and Information Technology, 1994, 16 (5): 456- 462.
7	张拥军, 叶怀珍, 任民. 神经网络模型预测运输货运量[J]. 西南交通大学学报, 1999, 34 (5): 602- 605.
	ZHANG Yongjun , YE Huaizhen , REN Min . Neural network model for forecasting freight transport volume[J]. Journal of Southwest Jiao Tong University, 1999, 34 (5): 602- 605.
8	林杰, 郭耀煌. 用神经网络方法预测股票短期走势[J]. 西南交通大学学报, 1998, 33 (3): 299- 304.
	LIN Jie , GUO Yaohuang . Prediction of stock short-term trend using neural network[J]. Journal of Southwest Jiaotong University, 1998, 33 (3): 299- 304.
9	钟颖, 汪秉文. 基于遗传算法的BP神经网络时间序列预测模型[J]. 系统工程与电子技术, 2002, 24 (4): 9- 11. doi: 10.3321/j.issn:1001-506X.2002.04.003
	ZHONG Ying , WANG Bingwen . BP neural network time series prediction model based on genetic algorithm[J]. Systems Engineering and Electronics, 2002, 24 (4): 9- 11. doi: 10.3321/j.issn:1001-506X.2002.04.003
10	辛治运, 顾明. 基于最小二乘支持向量机的复杂金融时间序列预测[J]. 清华大学学报(自然科学版), 2008, 48 (7): 1147- 1149. doi: 10.3321/j.issn:1000-0054.2008.07.021
	XIN Zhiyun , GU Ming . Complex financial time series prediction based on least squares support vector machine[J]. Journal of Tsinghua University(Natural Science), 2008, 48 (7): 1147- 1149. doi: 10.3321/j.issn:1000-0054.2008.07.021
11	杨平, 彭道刚, 韩璞, 等. 神经网络预测控制算法及其应用[J]. 控制工程, 2003, 10 (4): 349- 351. doi: 10.3969/j.issn.1671-7848.2003.04.019
	YANG Ping , PENG Daogang , HAN Pu , et al. Neural network predictive control algorithm and its application[J]. Control Engineering, 2003, 10 (4): 349- 351. doi: 10.3969/j.issn.1671-7848.2003.04.019
12	吴微, 陈维强, 刘波. 用BP神经网络预测股票市场涨跌[J]. 大连理工大学学报, 2001, 41 (1): 9- 15. doi: 10.3321/j.issn:1000-8608.2001.01.002
	WU Wei , CHEN Weiqiang , LIU Bo . Prediction of stock market ups and downs with BP neural network[J]. Journal of Dalian University of Technology, 2001, 41 (1): 9- 15. doi: 10.3321/j.issn:1000-8608.2001.01.002
13	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, 27 (3): 35- 38.
	YANG Yiyue , FU Qian , WAN Dingsheng . Time series prediction model based on deep circulation neural network[J]. Journal of Computer Technology and Development, 2017, 27 (3): 35- 38.
14	李洋, 董红斌. 基于CNN和BiLSTM网络特征融合的文本情感分析[J]. 计算机应用, 2018, 38 (11): 3075- 3080. doi: 10.11772/j.issn.1001-9081.2018041289
	LI Yang , DONG Hongbin . Text sentiment analysis based on feature fusion of CNN and BiLSTM networks[J]. Journal of Computer Applications, 2018, 38 (11): 3075- 3080. doi: 10.11772/j.issn.1001-9081.2018041289
15	LE T , BUI G , DUAN Y . A multi-view recurrent neural network for 3D mesh segmentation[J]. Computers & Graphics, 2017, 66 (Secial): 103- 112.
16	PASCUAL S, BONAFONTE A. Multi-output RNN-LSTM for multiple speaker speech synthesis with α-interpolation model[C]//ISCA Speech Synthesis Workshop.[S.l.]: [s.n], 2016: 112-117.
17	陈亮, 王震, 王刚. 深度学习框架下LSTM网络在短期电力负荷预测中的应用[J]. 电力信息与通信技术, 2017, 15 (5): 8- 11.
	CHEN Liang , WANG Zhen , WANG Gang . Application of LSTM network in short-term load forecasting under deep learning framework[J]. Electric Power Information and Communication Technology, 2017, 15 (5): 8- 11.
18	张清良, 李先明. 一种确定神经网络隐层节点数的新方法[J]. 吉首大学学报(自然科学版), 2002, 23 (1): 89- 91. doi: 10.3969/j.issn.1007-2985.2002.01.024
	ZHANG Qingliang , LI Xianming . A new method for determining the number of hidden layer nodes in neural networks[J]. Journal of Jishou University(Natural Science), 2002, 23 (1): 89- 91. doi: 10.3969/j.issn.1007-2985.2002.01.024
19	蔡荣辉, 崔雨轩, 薛培静. 三层BP神经网络隐层节点数确定方法探究[J]. 电脑与信息技术, 2017, 25 (5): 29- 33. doi: 10.3969/j.issn.1005-1228.2017.05.009
	CAI Ronghui , CUI Yuxuan , XUE Peijing . Research on determining the number of hidden layer nodes of three-layer BP neural network[J]. Computer and Information Technology, 2017, 25 (5): 29- 33. doi: 10.3969/j.issn.1005-1228.2017.05.009
20	金钊, 鲁淑霞. 非线性SGD算法改进预测超平面表示方法的研究[J]. 信息与电脑, 2016, (13): 99. doi: 10.3969/j.issn.1003-9767.2016.13.048
	JIN Zhao , LU Shuxia . Research on improved prediction of hyperplane representation by nonlinear SGD algorithm[J]. Information & Computer, 2016, (13): 99. doi: 10.3969/j.issn.1003-9767.2016.13.048
21	贺昱曜, 李宝奇. 一种组合型的深度学习模型学习率策略[J]. 自动化学报, 2016, 42 (6): 953- 958.
	HE Yuyao , LI Baoqi . A combined learning strategy for deep learning models[J]. IEEE/CAA Journal of Automatica Sinica (JAS), 2016, 42 (6): 953- 958.
22	崔广新, 李殿奎. 基于AdamOptimizer自编码的手写数字识别研究[J]. 佳木斯大学学报(自然科学版), 2018, 36 (3): 35- 38.
	CUI Guangxin , LI Diankui . Research on handwritten digit recognition rased on AdamOptimizer self-encoding[J]. Journal of Jiamusi University(Natural Science Edition), 2018, 36 (3): 35- 38.

多维度评价

Viewed

Full text

Abstract

Cited

Shared

Discussed

网络层数	MSE	编译时间/s	训练时间/s
5	0.026 8	0.046 0	2 007.956
3	0.072 2	0.078 0	1 131.730

节点数	MSE	计算时间/s
25	0.135	973.77
30	0.174	1 039.30
35	0.034	1 201.15
40	0.089	1 231.96
50	0.177	1 363.47

算法	残差均值	残差方差
SGD	0.136 830	0.026 786
Adagrad	0.112 304	0.026 312
RMSprop	0.119 042	0.027 368

训练(测试集)	普通神经网络	RNN	一元LSTM	多元LSTM
Train_X	(72220, 20)	(72220, 20, 1)	(72220, 20, 1)	(72220, 20, 6)
Train_Y	(72220, 1)	(72220, 1)	(72220, 1)	(72220, 6)
Test_X	(3800, 20)	(3800, 20, 1)	(3800, 20, 1)	(3800, 20, 6)
Test_Y	(3800, 1)	(3800, 1)	(3800, 1)	(3800, 1)

时间序列长度	一元LSTM		多元LSTM
时间序列长度	MSE	MAD	MSE	MAD
5	0.057 3	0.028 5	0.022 0	0.017 0
10	0.037 7	0.019 9	0.017 1	0.008 7
15	0.038 9	0.017 6	0.005 3	0.005 6
20	0.039 8	0.012 0	0.006 8	0.008 1
25	0.044 3	0.037 4	0.011 9	0.019 2