Interpretative analysis of carbon deposition process of methanol to olefins catalyst based on NMI-GBR-SHAP

doi:10.11949/0438-1157.20250952

Abstract

Abstract:

Catalyst carbon deposition in the methanol to olefins (MTO) process is complex and interdependent, making it difficult to effectively control carbon deposition. This study employed an integrated analytical framework based on normalized mutual information (NMI), gradient boosting regression (GBR), and SHAP to explain the key factors influencing catalyst carbon deposition. The results revealed that when the regeneration swirl inlet linear velocity was maintained at 16-18 m/s, combined with a reactor protection steam rate of 10-13 t/h, the total regenerator char air volume was controlled at 24 000-30 000 Nm³/h, the regeneration slide valve position was maintained at 35-39%, and the steam injection rate was no less than 34%, catalyst carbon deposition could be effectively controlled (approximately 6.5-7.3%). A benchmark limit is constructed based on the above-mentioned process indicators. Once the operating conditions exceed the monitoring limit, the corresponding carbon deposition trend (rate of change) can be explained by performing SHAP analysis on the interaction characteristics, providing the necessary monitoring information for the smooth operation of the process, and providing a technical solution for carbon deposition monitoring of MTO process catalysts with both predictive accuracy and mechanism explanation.

Key words: methanol to olefins, catalyst coke deposition, normalized mutual information, gradient boosting regression, interpretable machine learning, process systems, prediction, chemical processes

摘要：

甲醇制烯烃（Methanol to Olefins， MTO）工艺中催化剂积碳过程关联因素多且存在复杂交互作用，难以有效调控积碳量。本研究构建基于归一化互信息（NMI）、梯度提升回归（GBR）与SHAP的集成分析框架对催化剂积碳的主要影响因素进行解释性研究，发现当再生一旋入口线速维持在16-18m/s，同时配合10-13t/h的反应器保护蒸汽，将再生器烧焦总风量控制在24 000-30 000 Nm³/h且再生滑阀阀位保持在35-39%，并且蒸汽配入率不低于34%时，能够对催化剂的积碳情况进行有效调控（≈6.5-7.3%）。基于上述工艺指标构建基准限，一旦运行工况超出该监测限，通过对交互特征进行SHAP分析，能够对相应的积碳趋势（变化率）作出解释，为工艺过程的平稳运行提供必要的监测信息，为MTO过程催化剂积碳监测提供了兼具预测精度和机理解释性的技术方案。

关键词: 甲醇制烯烃, 催化剂积碳, 归一化互信息, 梯度提升回归, 可解释机器学习, 过程控制, 预测, 化学过程

CLC Number:

TQ 028.8

Wenliang LI, Hao ZHANG, Mingrui CHEN, Chen LIANG, Chi ZHAI. Interpretative analysis of carbon deposition process of methanol to olefins catalyst based on NMI-GBR-SHAP[J]. CIESC Journal, DOI: 10.11949/0438-1157.20250952.

李文亮, 张浩, 陈鸣睿, 梁晨, 翟持. 基于NMI-GBR-SHAP的甲醇制烯烃催化剂积碳过程解释性分析[J]. 化工学报, DOI: 10.11949/0438-1157.20250952.

Figures/Tables 10

Fig.1 Schematic of the Methanol to Olefins process

Fig.2 Framework of the NMI-GBR-SHAP model

Fig.3 Flow chart of NMI-GBR modeling and on-line soft sensing of catalyst coke content

Fig.4 NMI values of various process variables with catalyst coke content

Table 1 Evaluation metrics

模型	MAE	MSE	R²
NMI-Bayesian	0.2289	0.0903	0.6323
NMI-PLSR	0.2269	0.0880	0.6415
NMI-SVR	0.1286	0.0276	0.8874
NMI-GBR	0.1108	0.0201	0.9181

Fig.5 Comparison of coke content predictions from different models

Fig.6 Global feature distribution of the MTO process

Fig.7 Local feature characteristics of the MTO process

Fig.8 SHAP dependence of various variables

Fig.9 SHAP interaction contributions of key process-variable pairs to predicted catalyst coke deposition

References 30

[1]	Lin S F, Li H, Tian P, et al. Methanol to olefins (MTO): understanding and regulating dynamic complex catalysis[J]. Journal of the American Chemical Society, 2025, 147(14): 11585-11607.
[2]	季超,武鲁明,李滨,等. 甲醇制烯烃催化剂SAPO-34分子筛的改性研究进展[J]. 无机盐工业, 2022, 54(7): 1-9, 90.
	Ji C, Wu L M, Li B, et al. Research progress on modification of SAPO-34 as catalyst for methanol to olefins[J]. Inorganic Chemicals Industry,2022, 54(7):1-9, 90.
[3]	Zhou J B, Li X, Liu D P, et al. A hybrid spatial-temporal deep learning prediction model of industrial methanol-to-olefins process[J]. Frontiers of Chemical Science and Engineering, 2024, 18(4): 42.
[4]	Tian P, Wei Y X, Ye M, et al. Methanol to olefins (MTO): from fundamentals to commercialization[J]. ACS Catalysis, 2015, 5(3): 1922-1938.
[5]	Zhou J B, Gao M B, Zhang J L, et al. Directed transforming of coke to active intermediates in methanol-to-olefins catalyst to boost light olefins selectivity[J]. Nature Communications, 2021, 12: 17.
[6]	Yang L, Wang C, Zhang L N, et al. Stabilizing the framework of SAPO-34 zeolite toward long-term methanol-to-olefins conversion[J]. Nature Communications, 2021, 12: 4661.
[7]	Qi G Z, Xie Z K, Yang W M, et al. Behaviors of coke deposition on SAPO-34 catalyst during methanol conversion to light olefins[J]. Fuel Processing Technology, 2007, 88(5): 437-441.
[8]	Chen D, Grønvold A, Moljord K, et al. Methanol conversion to light olefins over SAPO-34: reaction network and deactivation kinetics[J]. Industrial & Engineering Chemistry Research, 2007, 46(12): 4116-4123.
[9]	Suvarna M, Pérez-Ramírez J. Embracing data science in catalysis research[J]. Nature Catalysis, 2024, 7(6): 624-635.
[10]	Adler A I, Painsky A. Feature importance in gradient boosting trees with cross-validation feature selection[J]. Entropy, 2022, 24(5): 687.
[11]	Lundberg S M, Lee S I. A unified approach to interpreting model predictions[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. December 4 - 9, 2017, Long Beach, California, USA. ACM, 2017: 4768-4777.
[12]	Lundberg S M, Erion G G, Lee S I. Consistent individualized feature attribution for tree ensembles[EB/OL]. 2018: arXiv: 1802.03888.
[13]	Alomari Y, Andó M. SHAP-based insights for aerospace PHM: Temporal feature importance, dependencies, robustness, and interaction analysis[J]. Results in Engineering, 2024, 21: 101834.
[14]	Ponce-Bobadilla A V, Schmitt V, Maier C S, et al. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development[J]. Clinical and Translational Science, 2024, 17(11): e70056.
[15]	Jutte A, Ahmed F, Linssen J, et al. C-SHAP for time series: an approach to high-level temporal explanations[EB/OL]. 2025: arXiv: 2504.11159.
[16]	Veríssimo R F, Matias P H F, Barbosa M R, et al. Integrating machine learning and SHAP analysis to advance the rational design of benzothiadiazole derivatives with tailored photophysical properties[J]. Journal of Chemical Information and Modeling, 2025, 65(15): 7874-7886.
[17]	Vinchurkar T, Ock J, Farimani A B. Explainable data-driven modeling of adsorption energy in heterogeneous catalysis[EB/OL]. 2024: arXiv: 2405.20397.
[18]	Xu H B, Yan S W, Qin X G, et al. Interpretable intelligent fault diagnosis for heat exchangers based on SHAP and XGBoost[J]. Processes, 2025, 13(1): 219.
[19]	Luo Y P, Wang W Y, Zhang Y Y, et al. A transformer-based model for predicting and analyzing light olefin yields in methanol-to-olefins process[J]. Chinese Journal of Chemical Engineering, 2025, 83: 266-276.
[20]	Zhang Y, Jia S L, Huang H Y, et al. A novel algorithm for the precise calculation of the maximal information coefficient[J]. Scientific Reports, 2014, 4: 6662.
[21]	Li Y, Li N, Ren J Z, et al. An interpretable light attention–convolution–gate recurrent unit architecture for the highly accurate modeling of actual chemical dynamic processes[J]. Engineering, 2024, 39: 104-116.
[22]	Draper N R, Smith H. Applied Regression Analysis[M]. 3rd ed. New York: Wiley, 1998.
[23]	李文亮, 纪成, 梁晨, 等. 基于TDMN的青霉素浓度在线软测量[J]. 化工学报, 2025, 76(6): 2848-2858.
	Li W L, Ji C, Liang C, et al. On-line soft measurement of penicillin concentration based on TDMN[J]. CIESC Journal, 2025, 76(6): 2848-2858.
[24]	Maes F, Collignon A, Vandermeulen D, et al. Multimodality image registration by maximization of mutual information[J]. IEEE Transactions on Medical Imaging, 1997, 16(2): 187-198.
[25]	杨明平, 罗娟. 甲醇制低碳烯烃反应体系的热力学计算与分析[J]. 煤化工, 2008, 36(3): 44-48.
	Yang M P, Luo J. Thermodynamic calculation and analysis of methanol to olefins(MTO)[J]. Coal Chemical Industry, 2008, 36(3): 44-48.
[26]	杨余. 基于数据驱动的DMTO工艺催化剂定碳软测量研究[D]. 重庆: 西南大学, 2023.
	Yang Y. Research on soft sensing of catalyst carbon determination in DMTO process based on data-driven[D]. Chongqing: Southwest University,2023.
[27]	Bentegri H, Rabehi M, Kherfane S, et al. Assessment of compressive strength of eco-concrete reinforced using machine learning tools[J]. Scientific Reports, 2025, 15: 5017.
[28]	Bi J, Hu F C, Wang Y J, et al. A method based on interpretable machine learning for recognizing the intensity of human engagement intention[J]. Scientific Reports, 2023, 13: 2537.
[29]	Yuan C, Zhang T M, Tang Y X, et al. Fault prediction method of large forging press based on a multi scale and multi model integrated method[J]. Scientific Reports, 2025, 15: 30675.
[30]	Colombo M, Seriès P. Bayes in the brain—on Bayesian modelling in neuroscience[J]. The British journal for the philosophy of science, 2012.