基于KELM-ANFIS的模拟移动床自适应建模

doi:10.11949/0438-1157.20250918

摘要/Abstract

摘要：

模拟移动床色谱分离是一种高效的新型分离技术，但是其分离过程中由于填料的老化、吸附剂活性的下降等会导致吸附特性发生变化。如何使得分离过程长期稳定是一个非常具有挑战性的问题。自适应模型能够适应过程条件的变化，确保模型输出如纯度，产率等始终反映当前的系统状态。为应对吸附剂老化，提出一种基于核极限学习自适应神经模糊推理系统（KELM-ANFIS）的模拟移动床自适应建模方法，模型采用滑动窗口策略进行在线数据采样，前件参数通过梯度下降法，后件参数则使用增量学习法进行参数更新，实现了离线训练与在线更新的结合，有效避免了频繁的离线重新建模，显著节约了时间和人力成本。实验结果与自适应KELM、自适应ELANFIS 等模型对比，验证了所提方法在吸附剂老化下的分离性能预测上的显著优越性，为模拟移动床在实际工况下的操作工艺优化提供了高效的自适应工具。

关键词: 模拟移动床, 色谱, 分离, 吸附, 自适应模型, 核极限学习自适应神经模糊推理系统

Abstract:

Simulated moving bed chromatography (SMB) is a highly efficient new separation technology. However, during the separation process, adsorption characteristics can change due to factors such as packing aging and decreased adsorbent activity. Maintaining long-term stability in the separation process is a significant challenge. A highly accurate adaptive model can adapt to changing process conditions, ensuring that model outputs such as purity and yield consistently reflect the current system state. To address the model mismatch caused by adsorbent aging, a simulated moving bed adaptive modeling approach based on the kernel extreme learning adaptive neural-fuzzy inference system (KELM-ANFIS) is proposed. This method innovatively combines the nonlinear feature capture capabilities of ANFIS with the kernel function mapping advantages of KELM, significantly enhancing the model's nonlinear modeling and generalization capabilities. The model employs a sliding window strategy for online data sampling and a separate parameter update strategy: gradient descent is used for the antecedent parameters, while incremental learning is used for rapid parameter update of the consequent parameters. This approach combines offline training with online updating, effectively avoiding frequent offline re-modeling and significantly saving time and labor costs. The experimental results were compared with models such as adaptive KELM and adaptive ELANFIS models, verifying the significant superiority of the proposed method in predicting separation performance under adsorbent aging, and providing an efficient adaptive tool for process prediction and operation process optimization of simulated moving bed under actual working conditions.

Key words: simulated moving bed, chromatography, separation, adsorption, adaptive model, kernel limit learning adaptive neuro-fuzzy inference system

中图分类号:

TQ 028.8

陈宇翔, 赵忠盖, 刘飞. 基于KELM-ANFIS的模拟移动床自适应建模[J]. 化工学报, DOI: 10.11949/0438-1157.20250918.

Yuxiang CHEN, Zhonggai ZHAO, Fei LIU. Adaptive modeling of simulated moving bed based on KELM-ANFIS[J]. CIESC Journal, DOI: 10.11949/0438-1157.20250918.

图/表 14

图1 模拟移动床分离过程示意图

Fig.1 Schematic diagram of the simulated moving bed separation process

图2 基于平衡理论的纯度区域划分图

Fig.2 Purity zone division based on equilibrium theory

图3 吸附剂老化下的平衡理论完全分离区变化趋势图

Fig.3 Trend diagram of changes in the equilibrium theory complete separation zone under adsorbent aging

图4 KELM-ANFIS的结构图

Fig.4 Structure diagram of KELM-ANFIS

图5 滑动窗口移动过程示意图

Fig.5 Schematic diagram of the sliding window movement process

图6 自适应KELM-ANFIS模型的流程图

Fig.6 Flowchart of the adaptive KELM-ANFIS model

表1 模拟移动床模型参数及工艺参数

Table 1 Simulated moving bed model parameters and process parameters

模型参数		工艺参数
柱分布结构	2-2-2-2	提取液进料浓度	0.5 g/ml
分离组分数	2	提余液进料浓度	0.5 g/ml
柱长	53.6 cm	进料液流量	0.02 ml/s
柱直径	2.6 cm	洗脱液流量	0.0414 ml/s
孔隙率	0.38	提取液流量	0.0348 ml/s
轴向扩散系数	0.0381 cm²/s	提余液流量	0.0266 ml/s
提取液亨利系数	0.54	循环液流量	0.0981 ml/s
提余液亨利系数	0.28	切换时间	1552 s

表2 数据集输入的范围和步长

Table 2 Range and step size of dataset input

输入参数	范围	步长
$m 2$	0.26-0.56	0.004
$m 3$	0.26-0.56	0.004

表2 数据集输入的范围和步长

Table 2 Range and step size of dataset input

输入参数	范围	步长
$m 2$	0.26-0.56	0.004
$m 3$	0.26-0.56	0.004

表3 亨利系数变化的初始值和步长

Table 3 Initial value and step size of Henry coefficient change

亨利系数	初始值	步长
$H A$	0.54	5e-5
$H B$	0.28	2.6e-5

表3 亨利系数变化的初始值和步长

Table 3 Initial value and step size of Henry coefficient change

亨利系数	初始值	步长
$H A$	0.54	5e-5
$H B$	0.28	2.6e-5

表4 静态下的各模型性能指标

Table 4 Static performance metrics of various models

性能指标		RMSE	SSE	MAPE	MAE
提取液纯度	BP神经网络	0.0385	0.2980	78.99%	0.0320
	ELANFIS	0.0245	0.1209	45.83%	0.0173
	KELM	0.0142	0.0407	26.04%	0.0095
	KELM-ANFIS	0.0049	0.0047	5.220%	0.0025
提余液纯度	BP神经网络	0.0202	0.0819	4.543%	0.0119
	ELANFIS	0.0123	0.0306	3.266%	0.0097
	KELM	0.0051	0.0052	2.157%	0.0042
	KELM-ANFIS	0.0008	0.0001	0.118%	0.0003
产率	BP神经网络	0.0330	0.2183	36.15%	0.0308
	ELANFIS	0.0077	0.0119	9.703%	0.0060
	KELM	0.0065	0.0085	5.476%	0.0334
	KELM-ANFIS	0.0020	0.0008	1.543%	0.0048

图7 亨利系数变化下的各模型输出曲线对比图

Fig.7 Comparison of output curves of various models under changes in Henry parameters

图8 亨利系数变化下的各模型预测输出误差曲线对比图

Fig.8 Comparison of the predicted output error curves of each model under the change of Henry coefficient

表5 亨利系数改变下的各模型性能指标

Table 5 Performance indicators of each model under the change of Henry coefficient

性能指标		RMSE	SSE	MAPE	MAE
提取液纯度	KELM-ANFIS	0.0711	1.6232	12.31%	0.0470
	自适应ELANFIS	0.0279	0.6225	6.32%	0.0177
	自适应KELM	0.0040	0.0130	0.70%	0.0027
	自适应KELM-ANIFS	0.0019	0.0029	0.29%	0.0006
提余液纯度	KELM-ANFIS	0.2526	1.9271	17.91%	0.4212
	自适应ELANFIS	0.0186	0.2754	5.29%	0.0133
	自适应KELM	0.0027	0.0059	0.68%	0.0017
	自适应KELM-ANIFS	0.0005	0.0002	0.11%	0.0003
产率	KELM-ANFIS	0.1475	1.6466	14.94%	0.0906
	自适应ELANFIS	0.0213	0.3626	3.53%	0.0160
	自适应KELM	0.0031	0.0077	0.40%	0.0023
	自适应KELM-ANIFS	0.0013	0.0013	0.14%	0.0007

图9 各自适应模型参数更新耗时对比图

Fig.9 Comparison chart of parameter update time for various adaptive models

参考文献 32

[1]	Lim Y I, Jorgensen S B. A fast and accurate numerical method for solving simulated moving bed (SMB) chromatographic separation problems[J]. Chemical Engineering Science, 2004, 59(10): 1931-1947.
[2]	Ruthven D M, Ching C B. Counter-current and simulated counter-current adsorption separation processes[J]. Chemical Engineering Science, 1989, 44(5): 1011-1038.
[3]	Lee J, Shin N C, Lim Y, et al. Modeling and simulation of a simulated moving bed for adsorptive para-xylene separation[J]. Korean Journal of Chemical Engineering, 2010, 27(2): 609-618.
[4]	Rodrigues A. Simulated moving bed technology: principles, design and process applications[M]. Butterworth-Heinemann, 2015.
[5]	Aniceto J P S, Silva C M. Simulated moving bed strategies and designs: from established systems to the latest developments[J]. Separation & Purification Reviews, 2015, 44(1): 41-73.
[6]	孙磊.模拟移动床分离过程的建模及预测控制策略研究[D].浙江大学,2007.
	Sun L. Research on modeling and predictive control approach of simulated moving bed separation process[D]. Zhejiang University, 2007.
[7]	Van Duc Long N, Le T H, Kim J I, et al. Separation of D-psicose and D-fructose using simulated moving bed chromatography[J]. Journal of Separation Science, 2009, 32(11): 1987-1995.
[8]	Kim K M, Han K W, Kim S I, et al. Simulated moving bed with a product column for improving the separation performance[J]. Journal of Industrial and Engineering Chemistry, 2020, 88: 328-338.
[9]	Ribeiro A E, Gomes P S, Pais L S, et al. Chiral separation of ketoprofen enantiomers by preparative and simulated moving bed chromatography[J]. Separation Science and Technology, 2011, 46(11): 1726-1739.
[10]	Azevedo D C S, Rodrigues A E. Design of a simulated moving bed in the presence of mass‐transfer resistances[J]. AIChE Journal, 1999, 45(5): 956-966.
[11]	Migliorini C, Mazzotti M, Morbidelli M. Simulated moving-bed units with extra-column dead volume[J]. AIChE Journal, 1999, 45(7): 1411-1421.
[12]	顾金生, 蒋慰孙, 顾幸生. 模拟移动床 Parex 过程动态模型[J]. 华东理工大学学报, 1997, 23(6): 725-730.
	Gu J S, Jiang W S, Gu X S. Dynamic model for simulating the Parex process of a moving bed[J]. Journal of East China University of Science and Technology, 1997, 23(6): 725-730.
[13]	Klatt K U, Hanisch F, Dünnebier G. Model-based control of a simulated moving bed chromatographic process for the separation of fructose and glucose[J]. Journal of Process Control, 2002, 12(2): 203-219.
[14]	盖旭东,佟泽民,杨春育. C_8芳烃吸附分离系统动态混合池模型[J]. 石油炼制与化工, 1994, 25(12): 45-48.
	Gai X D, Tong Z M, Yang C Y. A dynamic approach to the adsorption separation process of c, aromatics[J]. Petroleum Processing and Petrochemicals, 1994, 25(12): 45-48.
[15]	Lim Y I, Lee J, Bhatia S K, et al. Improvement of para-xylene SMB process performance on an industrial scale[J]. Industrial & Engineering Chemistry Research, 2010, 49(7): 3316-3327.
[16]	Suzaki P Y R, Munaro M T, Triques C C, et al. Phenomenological mathematical modeling of heavy metal biosorption in fixed-bed columns[J]. Chemical Engineering Journal, 2017, 326: 389-400.
[17]	Kim Y, Cho S, Jang K, et al. Effect of radial distribution of injected flow on simulated moving bed performance[J]. Journal of Chromatography A, 2022, 1662: 462703.
[18]	Aniceto J P S, Cardoso S P, Silva C M. General optimization strategy of simulated moving bed units through design of experiments and response surface methodologies[J]. Computers & Chemical Engineering, 2016, 90: 161-170.
[19]	郑震玮.基于机器学习的模拟移动床分离过程优化[D].厦门大学,2021.
	Zheng Z W. Optimization of simulated moving bed separation process based on machine learning[D]. Xiamen University, 2021.
[20]	Yan Z, Wang J S, Wang S Y, et al. Model predictive control method of simulated moving bed chromatographic separation process based on subspace system identification[J]. Mathematical Problems in Engineering, 2019, 2019(1): 2391891.
[21]	李凌, 袁德成, 井元伟. 模拟移动床过程的软测量建模仿真研究[J]. 计算机与应用化学, 2014, 31(11): 1298-1302.
	Li L, Yuan D C, Jing Y W. Simulation research on soft-sensor modeling of Simulated Moving Bed process[J]. Computers and Applied Chemistry, 2014, 31(11): 1298-1302.
[22]	Yao C, Zheng Z, Chen K, et al. Using a machine learning model for the optimal design of simulated moving bed processes and its application to separate rebaudioside A and stevioside[J]. Journal of Chemical Technology & Biotechnology, 2021, 96(9): 2558-2568.
[23]	Wang D, Wang J S, Wang S Y, et al. Soft sensing modeling of the SMB chromatographic separation process based on the adaptive neural fuzzy inference system[J]. Journal of Sensors, 2019, 2019(1): 1312709.
[24]	Huang G B, Zhu Q Y, Siew C K. Extreme learning machine: theory and applications[J]. Neurocomputing, 2006, 70(1-3): 489-501.
[25]	Huang G B, Zhou H, Ding X, et al. Extreme learning machine for regression and multiclass classification[J]. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 2011, 42(2): 513-529.
[26]	Wu K, Xu C, Yan J, et al. Error-distribution-free kernel extreme learning machine for traffic flow forecasting[J]. Engineering Applications of Artificial Intelligence, 2023, 123: 106411.
[27]	Mueller I, Seidel-Morgenstern A, Hamel C. Simulated-moving-bed technology for purification of the prebiotics galacto-oligosaccharides[J]. Separation and Purification Technology, 2021, 271: 118829.
[28]	Sun Y C, Wang J S, Xing C, et al. Bidirectional online sequence extreme learning machine and switching strategy for soft-sensor model of SMB chromatography separation process[J]. Soft Computing, 2025, 29(2): 485-507.
[29]	Mazzotti M, Storti G, Morbidelli M. Optimal operation of simulated moving bed units for nonlinear chromatographic separations[J]. Journal of Chromatography A, 1997, 769(1): 3-24.
[30]	Gomes P S, Minceva M, Rodrigues A E. Operation of an industrial SMB unit for p-xylene separation accounting for adsorbent ageing problems[J]. Separation Science and Technology, 2008, 43(8): 1974-2002.
[31]	魏朋, 陈珺, 王志国, 等. 基于平衡理论的模拟移动床工艺参数鲁棒寻优[J]. 化工学报, 2022, 73(2): 792-800.
	Wei P, Chen J, Wang Z G, et al. Robust optimization of process parameters of simulated moving bed based on equilibrium theory[J]. CIESC Journal, 2022, 73(2): 792-800.
[32]	Sá Gomes P, Minceva M, Rodrigues A E. Operation strategies for simulated moving bed in the presence of adsorbent ageing[J]. Separation Science and Technology, 2007, 42(16): 3555-3591.