基于MIC筛选规则和BP神经网络的变换装置建模及产品预测

doi:10.16085/j.issn.1000-6613.2021-2552

摘要/Abstract

摘要：

在石化企业数字工厂建设中，装置的数字智能化建设十分重要。本文针对某石化企业变换装置数字化建设的需要，结合装置特点，建立了基于最大信息系数方法（MIC）的装置实时数据筛选规则和基于BP神经网络的装置产品质量预测模型。结果表明，利用实时数据筛选规则对采集到的44天共1041组装置实际运行数据进行分析，将161个变量参数删减到23个变量参数，有效降低了数据的维度，数据简化率达到85.63%；进一步采用Levenberg-Marquardt方法，用3层隐含层的网络结构建立装置的产品质量预测模型，模拟得到的装置出口变换气CO摩尔含量值与实际生产偏差很小（平均偏差1.193%），说明本文所建模型可以很好地预测装置产品组成。以上建立的模型可为装置生产优化提供支撑，并可集成到工厂信息物理系统（CPS）中，支撑装置数字化和智能化建设需要。本文所提出的建模方法同样可用于其他类似装置的建模参考。

关键词: 变换装置, 混合建模, BP神经网络, 最大信息系数方法, 信息物理系统

Abstract:

In the construction of digital factories in petrochemical enterprises, the digital intelligent construction of the unit is a key step. According to the need of a shift unit intelligent construction in a petrochemical enterprise, combined with the characteristics of the unit, a set of real-time data filtering rules was created based on maximum information coefficients (MIC), and the product quality prediction model was constructed based on the BP neural network. Results showed that own-developed real-time data filtering rules were used to analyze the actual operating data of a total of 1041 sets collected for 44 days. After the number of variables had been reduced from 161 to 23, it was found that the screening rules could effectively reduce the dimensionality of the data, with a data simplification rate of 85.63%. Furthermore, this model with the best product quality prediction was constructed by the Levenberg-Marquardt method and a network structure with three hidden layers. Comparing the CO molar content value of the shift gas at the outlet of the unit obtained by simulation with the actual production value, it was found that the simulated value deviated very little from the actual value (average deviation 1.193%), which showed that the model built in this paper could predict the product composition of unit very well, and it was proved to be reliable. This model is not only useful for optimizing shift unit production, but also integrated into the factory cyber-physical system (CPS) to facilitate unit digitization and intelligent construction. In addition, the method can be used as modeling reference of other similar units.

Key words: shift unit, hybrid modeling, BP neural network, MIC, CPS

中图分类号:

TQ015

潘艳秋, 李鹏飞, 高石磊, 俞路. 基于MIC筛选规则和BP神经网络的变换装置建模及产品预测[J]. 化工进展, 2022, 41(S1): 36-43.

PAN Yanqiu, LI Pengfei, GAO Shilei, YU Lu. Application of BP neural network based on MIC screening rules in modeling and product prediction of shift unit[J]. Chemical Industry and Engineering Progress, 2022, 41(S1): 36-43.

图/表 13

图1 变换装置工艺物料流程简图

表1 变量数据采集结果

日期	FIC2201	TI2216E	PI2201	PDI2241	…	AI2201
10/01/0：00	83.696	417.92	3.433	25.51	…	0.4237
…	…	…	…	…	…	…
10/02/0：00	81.5675	418.30	3.4201	30.30	…	0.4307
…	…	…	…	…	…	…
10/03/0：00	81.3952	421.16	3.4257	34.16	…	0.4550
…	…	…	…	…	…	…
10/04/0：00	95.8543	420.26	3.5567	42.12	…	0.5284
…	…	…	…	…	…	…
11/13/0：00	100.5424	416.68	3.5665	75.17	…	0.6530

图2 数据筛选流程

图3 数据标准化结果

表2 函数测试结果

方法	sigmoid $y = 1 1 - e - x$	tanh $y = e x - e - x e x + e - x$	lin y=0.2x	co-no $y = s i n (π x 2) + 20 (x - 0.5) 2 + 10 x 600$
Pearson	0.973	0.935	1.000	-0.187
Spearman	1.000	1.000	1.000	-0.140
Kendall	1.000	1.000	1.000	-0.096
MIC	1.000	1.000	1.000	1.000

表2 函数测试结果

方法	sigmoid $y = 1 1 - e - x$	tanh $y = e x - e - x e x + e - x$	lin y=0.2x	co-no $y = s i n (π x 2) + 20 (x - 0.5) 2 + 10 x 600$
Pearson	0.973	0.935	1.000	-0.187
Spearman	1.000	1.000	1.000	-0.140
Kendall	1.000	1.000	1.000	-0.096
MIC	1.000	1.000	1.000	1.000

表3 输入变量与目标变量相关性结果

序号	MIC	序号	MIC
1	0.6117	18	0.4051
2	0.2278	19	0.3978
3	0.2758	20	0.4371
4	0.6272	21	0.3866
5	0.3623	22	0.3996
6	0.3246	23	0.3918
7	0.4250	24	0.5599
8	0.2665	25	0.4182
9	0.3395	26	0.2440
10	0.4806	27	0.5809
11	0.2833	28	0.3735
12	0.5588	29	0.7055
13	0.5655	30	0.2579
14	0.5568	31	0.2424
15	0.5577	…	…
16	0.5568	107	0.3020
17	0.5657	均值	0.4553

图4 输入变量相关性分析结果

表4 变量编号及名称

变量编号	变量名称及含义	变量编号	变量名称及含义
1	TI2201/来自粉煤气化的粗合成气温度	13	PI2206/E2201入口低压氮气压力
2	TIC2206/E2201入口HS温度	14	PDI2207/R2201A催化剂床层压差
3	TI2211E/R2201A催化剂床层E点温度	15	PIC2214/E2205压力
4	TI2214A/R2201B催化剂床层A点温度	16	PDI2222/R2204催化剂床层压差
5	TIC2270/R2203入口温度	17	PI2231/P2201A/B出口工艺冷凝液压力
6	TI2225A/R2203催化剂床层A点温度	18	FIC2201/进变换系统MS流量
7	TI2231/E2208出口低压蒸汽温度	19	FIC2208/V2204入口锅炉水流量
8	TI2235/V2202出口变换气温度	20	LI2201/V2201汽液分离器液位
9	TI2237/E2210出口变换气温度	21	LIC2208/V2206液位差
10	TI2246/C2202出口酸性气体温度	22	LIC2210/E2208液位
11	PI2201/V2201粗合成气进口压力	23	AI2201/CO摩尔组成分析
12	PI2203进变换中压蒸汽压力

表5 不同算法结果对比

算法	R	MSE/×10^-4	时间/s
L-M	0.979	1.56	117
B-R	0.983	1.19	1001
SCG	0.943	4.06	151

图5 隐含层数的影响

表6 不同神经元组合计算结果

序号	组合	R	MSE	序号	组合	R	MSE
1	6×6×6	0.9788	1.560×10^-4	10	10×7×6	0.9804	1.447×10^-4
2	7×6×6	0.9794	1.515×10^-4	11	6×8×6	0.9791	1.540×10^-4
3	8×6×6	0.9798	1.489×10^-4	12	7×8×6	0.9800	1.476×10^-4
4	9×6×6	0.9805	1.436×10^-4	13	8×8×6	0.9804	1.444×10^-4
5	10×6×6	0.9807	1.426×10^-4	14	9×8×6	0.9805	1.439×10^-4
6	6×7×6	0.9791	1.537×10^-4	…	…	…	…
7	7×7×6	0.9798	1.492×10^-4	95	10×9×9	0.9813	1.382×10^-4
8	8×7×6	0.9800	1.471×10^-4	…	…	…	…
9	9×7×6	0.9805	1.439×10^-4	125	10×10×10	0.9812	1.383×10^-4

图6 模型的拟合结果

图7 模拟结果对比

参考文献 21

1	MA Y, XIA L, MENG W. A review of advanced manufacturing strategies and development in typical industrialized countries[J]. Journal of Physics: Conference Series, 2019, 1237: 42007.
2	中国电子技术标准化研究院, 中国信息物理系统发展论坛. 信息物理系统建设指南 (2020)[EB/OL]. [2020-12-11]. .
	China Electronics Standardization Institute, China Cyber Physics System Development Forum. Cyber-physical system construction guide (2020) [EB/OL]. [2020-12-11]. .
3	LI D. Perspective for smart factory in petrochemical industry[J]. Computers & Chemical Engineering, 2016, 91: 136-148.
4	FENG Y, ZHAO Y, ZHENG H, et al. Data-driven product design toward intelligent manufacturing: a review[J]. International Journal of Advanced Robotic Systems, 2020, 17(2): 1729881420911257-1-1729881420911257-18.
5	EESA A, ARABO W. A normalization methods for backpropagation: a comparative study[J]. Science Journal of University of Zakho, 2017, 5: 319.
6	ANTWI P, LI J, BOADI P O, et al. Estimation of biogas and methane yields in an UASB treating potato starch processing wastewater with backpropagation artificial neural network[J]. Bioresource Technology, 2017, 228: 106-115.
7	刘金鹏, 赵兵涛, 钱魏锋, 等. 基于BP神经网络的旋风分离器分割粒径模化与预测[J]. 化工进展, 2021, 40(2): 671-677.
	LIU Jinpeng, ZHAO Bingtao, QIAN Weifeng, et al. Modeling and prediction of particle cutoff size of cyclone separator based on BP neural network[J]. Chemical Industry and Engineering Progress, 2021, 40(2): 671-677.
8	董玉玺, 李乐宁, 田文德. 基于多层优化PCC-SDG方法的化工过程故障诊断[J]. 化工学报, 2018, 69(3): 1173-1181.
	DONG Yuxi, LI Lening, TIAN Wende. A novel fault diagnosis method based on multilayer optimized PCC-SDG[J]. CIESC Journal, 2018, 69(3): 1173-1181.
9	DOWNS J J, VOGEL E F. A plant-wide industrial process control problem[J]. Computers & Chemical Engineering, 1993, 17(3): 245-255.
10	ZHOU Y, LI S. BP neural network modeling with sensitivity analysis on monotonicity based Spearman coefficient[J]. Chemometrics and Intelligent Laboratory Systems, 2020, 200: 103977.
11	任玉佳, 王骥, 田文德. 基于特征工程和KELM的化工过程故障检测与识别[J].高校化学工程学报, 2019, 33(5): 1271-1284.
	RENG Yujia, WANG Ji, TIAN Wende. Fault detection and identification in chemical processes based on feature engineering and kernel extreme learning machine[J]. Journal of Chemical Engineering of Chinese Universities, 2019, 33(5): 1271-1284.
12	RESHEF D N, RESHEF Y A, FINUCANE H K, et al. Detecting novel associations in large data sets[J]. Science, 2011, 334(6062): 1518.
13	马立平. 统计数据标准化-无量纲化方法-现代统计分析方法的学与用(三)[J]. 北京统计, 2000(3): 34-35.
	MA Liping. Statistical data standardization-dimensionless method-learning and application of modern statistical analysis methods (3)[J]. Beijing statistics,2000(3):34-35.
14	顾俊发, 许明阳, 马方圆, 等. 基于MIC的支持向量回归及其在化工过程中的应用[J].化工学报, 2021, 72(3): 1480-1486.
	GU Junfa, XU Mingyang, MA Fangyuan, et al. Support vector regression based on maximal information coefficient and its application in chemical industrial processes[J]. CIESC Journal, 2021, 72(3): 1480-1486.
15	高石磊. 基于数字化建模的变换装置产品预测与优化[D]. 大连: 大连理工大学, 2021.
	GAO Shilei. Product prediction and optimization of shift device based on digital modeling[D]. Dalian: Dalian University of Technology, 2021.
16	孟燕霞. 最大信息系数算法研究[D]. 太原: 太原理工大学, 2019.
	MENG Y X. The maximal information coefficient algorithm research[D]. Taiyuan: Taiyuan University of Technology, 2019.
17	RESHEF D, RESHEF Y, MITZENMACHER M, et al. Equitability analysis of the maximal information coefficient, with comparisons[J]. Arxiv Preprint Arxiv:, 2013.
18	马文礼, 李治平, 高闯, 等. 页岩气井初期产能主控因素“Pearson-MIC”分析方法[J]. 中国科技论文, 2018, 13(15): 1765-1771.
	MA Wenli, LI Zhiping, GAO Chuang, et al. “Pearson-MIC” analysis method for the initial production key controlling factors of shale gas wells[J]. China Sciencepaper, 2018, 13(15): 1765-1771.
19	姜丰, 朱家玲, 胡开永, 等. Pearson相关系数评价ORC系统蒸发器特性的应用研究[J].太阳能学报, 2019, 40(10): 2732-2738.
	JIANG Feng, ZHU Jialing, HU Kaiyong, et al. Pearson correlation coefficient evaluation of orc system evaporator characteristics applied research[J]. Acta Energiae Solaris Sinica, 2019, 40(10): 2732-2738.
20	DAMILOLA E B, AMBROSE A, OMOLEYE J. Artificial neural network and its applications in the energy sector: an overview[J]. International Journal of Energy Economics and Policy, 2020, 10(2): 250-264.
21	ASTHANA S, PANDIT A, BAHRDWAJ A. Analysis of multiple hidden layers vs. accuracy in performance using back propagation neural network[J]. Indian Journal of Science and Technology, 2017, 10(4): doi:10.17485/ijst/2017/v10i4/110899.

[1]	王子宗, 高立兵, 索寒生. 未来石化智能工厂顶层设计：现状、对比及展望[J]. 化工进展, 2022, 41(7): 3387-3401.
[2]	张梦轩, 刘洪辰, 王敏, 蓝兴英, 石孝刚, 高金森. 化工过程的智能混合建模方法及应用[J]. 化工进展, 2021, 40(4): 1765-1776.
[3]	高石磊, 潘艳秋, 李鹏飞, 张春超, 俞路, 王振兴. 基于混合建模的变换反应器产品预测与优化[J]. 化工进展, 2021, 40(12): 6604-6612.
[4]	韩丰磊, 李梦雨, 刘杰夫, 李丹丹, 郭雯雯, 周硕, 赵朝成, 赵东风. 基于BP神经网络的煤直接液化装置VOCs包袋法检测与预测建模[J]. 化工进展, 2020, 39(10): 3949-3955.
[5]	付柯, 谢良才, 闫雨瑗, 李波, 贺改, 徐龙, 马晓迅. 改进BP神经网络预测Ni/Al₂O₃催化CH₄-CO₂重整反应[J]. 化工进展, 2017, 36(07): 2393-2399.
[6]	赵斌1,3，吴桂芳2，郝培文1. 应用红外光谱进行不同性状沥青的判别[J]. 化工进展, 2012, 31(12): 2753-2755.
[7]	宋波，江青茵，曹志凯. 循环流化床锅炉热效率在线计算与影响因素分析 [J]. 化工进展, 2008, 27(4): 616-.