基于TabPFN的汽油辛烷值预测与模型可解释性

李亚洁; 王文春; 李炜

doi:10.11949/0438-1157.20260529

您当前的位置：

首页 >

文章列表页 >

基于TabPFN的汽油辛烷值预测与模型可解释性

智能过程工程 | 更新时间：2026-06-18

- 基于TabPFN的汽油辛烷值预测与模型可解释性
- Gasoline octane number prediction and model interpretability based on TabPFN
- 化工学报 2026年
- 作者机构：
  
  1.兰州理工大学自动化与电气工程学院，甘肃兰州 730050
  2.甘肃省工业过程先进控制重点实验室，甘肃兰州 730050
  3.兰州理工大学电气与控制工程国家级实验教学示范中心，甘肃兰州 730050
- 作者简介：
  
  李亚洁（1981—），女，博士，副教授，liyaj@lut.edu.cn
  李炜（1963—），女，硕士，教授，liwei@lut.edu.cn
- 基金信息：
  
  国家自然科学基金项目(62163022);甘肃省科技重大专项计划(23ZDGA007);甘肃省产业支撑计划项目(2024CYZC-18);甘肃省博士基金项目(2021QB-044)
- DOI：10.11949/0438-1157.20260529
  中图分类号： TP 273
- 收稿：2026-04-14，
  
  修回：2026-06-17，
  
  录用：2026-06-18，
- 稿件说明：
移动端阅览
李亚洁, 王文春, 李炜. 基于TabPFN的汽油辛烷值预测与模型可解释性[J/OL]. 化工学报, 2026.

LI Yajie, WANG Wenchun, LI Wei. Gasoline octane number prediction and model interpretability based on TabPFN[J/OL]. CIESC Journal, 2026.
李亚洁, 王文春, 李炜. 基于TabPFN的汽油辛烷值预测与模型可解释性[J/OL]. 化工学报, 2026. DOI： 10.11949/0438-1157.20260529.

LI Yajie, WANG Wenchun, LI Wei. Gasoline octane number prediction and model interpretability based on TabPFN[J/OL]. CIESC Journal, 2026. DOI： 10.11949/0438-1157.20260529.

摘要

针对汽油在线调和配方优化中传统机器学习模型存在超参数调优复杂、建模效率偏低及可解释性不足等问题，提出了一种基于表格先验数据拟合网络（TabPFN）的汽油辛烷值预测与模型可解释分析方法。首先，以企业实际汽油调和组分及配比为输入、辛烷值为输出，建立基于TabPFN的辛烷值预测模型。其次，通过引入沙普利加性解释（SHAP）和偏依赖图（PDP）方法，从全局与局部、单变量与双变量多维度解析各组分及配比对辛烷值的影响机制。结果表明，本文的辛烷值预测模型无需复杂调参，测试集均方误差低至0.001，训练时间仅4 s，进而SHAP量化了各组分的贡献与影响程度，PDP揭示了各组分的非线性响应与交互作用，可为汽油在线调和配方优化提供高精度预测与决策支持。

Abstract

Gasoline octane number is a key quality index in the blending process of finished gasoline. Its accurate prediction is of great significance for improving oil quality

optimizing production efficiency and reducing production costs. Aiming at the problems of complex hyper-parameter tuning

low modeling efficiency and insufficient interpretability of existing machine learning models in octane number prediction

a gasoline octane number prediction and model interpretability analysis method based on tabular prior-data fitted network ( TabPFN ) is proposed. Firstly

based on the actual gasoline blending data of the enterprise

each blending component and its ratio are used as input variables

and the octane number is used as the output variable. Then

a gasoline octane number prediction model based on TabPFN is constructed and compared with several typical machine learning models in the verification process. Secondly

a variety of evaluation indicators are used to comprehensively evaluate the performance of the model from the three dimensions prediction accuracy

training efficiency and reasoning efficiency. Finally

the interpretability of the prediction model is analyzed by introducing Shapley additive explanation ( SHAP ) and partial dependence plot ( PDP ) methods. The experimental results show that the mean square error of the octane number prediction of the TabPFN model is as low as 0.001 on the test set without a lot of hyper-parameter tuning

and the model training takes time is only 4 s

which is better than the comparison model. SHAP analysis identified and quantified the contribution of each gasoline blending component and its ratio to the octane number prediction results. PDP analysis further revealed the marginal effect

nonlinear relationship and interaction of each blending component and its combination on the octane number prediction results. This method can achieve efficient

accurate and strong explanatory power of gasoline octane number prediction

which can provide high-precision prediction and decision support for gasoline online blending formulation optimization.

关键词

Keywords

references

何仁初 , 张朝晖 , 杨明磊 , 等 . 考虑碳排放因素的汽油调合在线优化 [J ] . 化工学报 , 2023 , 74 ( 2 ): 818 - 829 .

He R C , Zhang Z H , Yang M L , et al . Online optimization of gasoline blending considering carbon emissions [J ] . CIESC Journal , 2023 , 74 ( 2 ): 818 - 829 .

He R C , Yan X Y , Hua J J , et al . Closed-loop gasoline blending scheduling based on real-time optimized slack feedback [J ] . Chemical Engineering Science , 2025 , 309 : 121426 .

Lyu D L , Cui X J , Qu Y Q , et al . Fuel octane number measurement method based on high-frequency combustion pressure and its application to fuel octane number over 100 [J ] . Energy , 2025 , 335 : 138234 .

Fernanda M A , Irawan B . The effect of different octane number on power and specific fuel consumption in gasoline compression ignition engine [J ] . Evrimata: Journal of Mechanical Engineering , 2025 , 2 ( 2 ): 32 - 36 .

Wu Y S , Liu Y S , Li X L , et al . Gasoline octane number prediction from near-infrared spectroscopy with an ANN-based model [J ] . Fuel , 2022 , 318 : 123543 .

高永红 . 中红外光谱法在汽油生产控制中的应用 [J ] . 石化技术 , 2018 , 25 ( 10 ): 135 - 137 .

Gao Y H . Application of mid-infrared spectroscopy in gasoline production control [J ] . Petrochemical Industry Technology , 2018 , 25 ( 10 ): 135 - 137 .

王拓 , 戴连奎 , 马万武 . 拉曼光谱结合后向间隔偏最小二乘法用于调和汽油辛烷值定量分析 [J ] . 分析化学 , 2018 , 46 ( 4 ): 623 - 629 .

Wang T , Dai L K , Ma W W . Quantitative analysis of blended gasoline octane number using Raman spectroscopy with backward interval partial least squares method [J ] . Chinese Journal of Analytical Chemistry , 2018 , 46 ( 4 ): 623 - 629 .

蒋书波 , 林锦国 , 程明霄 , 等 . 基于拉曼技术的汽油辛烷值测定系统设计 [J ] . 化工学报 , 2011 , 62 ( 8 ): 2188 - 2194 .

Jiang S B , Lin J G , Cheng M X , et al . System design of gasoline octane number detection based on Raman technology [J ] . CIESC Journal , 2011 , 62 ( 8 ): 2188 - 2194 .

Wang B Y , Cao R Y , Pi Z Y , et al . Displacement-imbibition coupling mechanisms between matrix and complex fracture during injecting-shut in-production process using pore-scale simulation and NMR experiment [J ] . Geoenergy Science and Engineering , 2024 , 242 : 213254 .

Cheng H R , Wei J G , Zhou X F , et al . Experimental study on the effect of microscale wettability of shale to chemical agents by using nuclear magnetic resonance method [J ] . Powder Technology , 2024 , 445 : 120084 .

Jiao L , Liu H H , Qu L , et al . QSPR studies on the octane number of toluene primary reference fuel based on the electrotopological state index [J ] . ACS Omega , 2020 , 5 ( 8 ): 3878 - 3888 .

Wu Y S , Guan C , Huang Z , et al . Fuel octane number prediction based on topological indices and active subspace method [J ] . Fuel , 2021 , 293 : 120494 .

Al-Fahemi J H , Albis N A , Gad E A M . QSPR models for octane number prediction [J ] . Journal of Theoretical Chemistry , 2014 , 2014 : 520652 .

Issa H M . Prediction of octane numbers for commercial gasoline using distillation curves: a comparative regression analysis between principal component and partial least squares methods [J ] . Petroleum Science and Technology , 2024 , 42 ( 10 ): 1233 - 1249 .

仇爱波 , 周如金 , 邱松山 , 等 . 基于原子类型电拓扑状态指数的 QSPR 模型预测链烷烃辛烷值 [J ] . 天然气化工—C1化学与化工 , 2014 , 39 ( 5 ): 52 - 55 .

Qiu A B , Zhou R J , Qiu S S , et al . QSPR modeling with electrotopological state index for atom type for predicting octane number of paraffins [J ] . Natural Gas Chemical Industry , 2014 , 39 ( 5 ): 52 - 55 .

Mohan B , Alsaleh F , AlRamadan A S , et al . In silico estimation of density, vapor pressure, distillation curve, and octane numbers of gasoline fuels using machine learning [J ] . Fuel , 2025 , 399 : 135640 .

Wang C C , Chien C H . Machine learning for industrial optimization and predictive control: a patent-based perspective with a focus on Taiwan’s high-tech manufacturing [J ] . Processes , 2025 , 13 ( 7 ): 2256 .

Sun X Y , Zhang F , Liu J P , et al . Prediction of gasoline research octane number using multiple feature machine learning models [J ] . Fuel , 2023 , 333 : 126510 .

Zhang F Y , Su X C , Tan A L , et al . Prediction of research octane number loss and sulfur content in gasoline refining using machine learning [J ] . Energy , 2022 , 261 : 124823 .

Lin X C , Yin S T , Li X . A trustworthy framework for predicting gasoline octane loss using sparse autoencoder and stacking ensemble learning in petrochemical processes [J ] . Results in Engineering , 2025 , 28 : 107519 .

Hettikankanamage N , Shafiabady N , Chatteur F , et al . eXplainable artificial intelligence (XAI): a systematic review for unveiling the black box models and their relevance to biomedical imaging and sensing [J ] . Sensors , 2025 , 25 ( 21 ): 6649 .

Zhao C , Liu J , Parilina E . The Shapley value contribution to explainable artificial intelligence: a comprehensive survey [J ] . Dynamic Games and Applications , 2025 : 1 - 38 .

Pelosi D , Cacciagrano D , Piangerelli M . Explainability and interpretability in concept and data drift: a systematic literature review [J ] . Algorithms , 2025 , 18 ( 7 ): 443 .

李炜 , 郑明杰 , 李亚洁 , 等 . 融合可解释机器学习的成品汽油调和配方质量预测评价与致因分析 [J ] . 石油学报(石油加工) , 2024 , 40 ( 1 ): 126 - 136 .

Li W , Zheng M J , Li Y J , et al . Predictive evaluation and cause analysis of finished gasoline blending formulation quality by interpretable machine learning [J ] . Acta Petrolei Sinica (Petroleum Processing Section) , 2024 , 40 ( 1 ): 126 - 136 .

信自成 , 张江山 , 张军国 , 等 . TabPFN与SHAP融合的LF精炼Si元素收得率预测模型 [J ] . 中国冶金 , 2025 , 35 ( 11 ): 178 - 186 .

Xin Z C , Zhang J S , Zhang J G , et al . Prediction model of Si element yield in LF refining based on TabPFN and SHAP fusion [J ] . China Metallurgy , 2025 , 35 ( 11 ): 178 - 186 .

郑可欣 , 江雨欣 , 毕可鑫 , 等 . 用于蒸汽裂解产物成分预测的集成迁移学习框架 [J ] . 化工进展 , 2024 , 43 ( 5 ): 2880 - 2889 .

Zheng K X , Jiang Y X , Bi K X , et al . Ensemble transfer learning framework for outflow compositions prediction in steam cracking process [J ] . Chemical Industry and Engineering Progress , 2024 , 43 ( 5 ): 2880 - 2889 .

陈霖 , 祖思雅 , 胡方杰 , 等 . 炼化含油污水全流程急性生物毒性演变与关键影响因子识别 [J/OL ] . 化工进展 . https://doi.org/10.16085/j.issn.1000-6613.2025-1280 https://doi.org/10.16085/j.issn.1000-6613.2025-1280 .

Chen L , Zu S Y , Hu F J , et al . Evolution of acute ecotoxicity and identification of key influencing factors in full-process refinery oily wastewater [J/OL ] . Chemical Industry and Engineering Progress . https://doi.org/10.16085/j.issn.1000-6613.2025-1280 https://doi.org/10.16085/j.issn.1000-6613.2025-1280 .

Huang W Y , Li K Q , Miao S S . Real-time process design enabled by an interpretable GBDT model for high-fidelity prediction of pyrolysis products from crop residues [J ] . Bioresource Technology , 2026 , 442 : 133742 .

Panicker T F , Sarkar D , Krishna R , et al . A machine learning-assisted prediction of potential biochar yield subjected to the physicochemical properties of biomass [J ] . Fuel , 2026 , 409 : 137836 .

Yun J H , Yu Y , Tao C L , et al . Machine learning-based optimization of enhanced nitrogen removal in a full-scale urban wastewater treatment plant with ecological combination ponds [J ] . Water Research , 2025 , 285 : 123976 .

Hollmann N , Müller S , Purucker L , et al . Accurate predictions on small data with a tabular foundation model [J ] . Nature , 2025 , 637 ( 8045 ): 319 - 326 .

Wang H Y , Pei J , Yan F Q , et al . Application of TabPFN model on the energy performance improvement of high-power multistage centrifugal pump [J ] . Energy , 2025 , 336 : 138399 .

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

燃料电池数字孪生系统综述

基于机器学习算法的生物质热解产物产率预测研究

饱和气液相密度关联外推的新通用对比态方程

基于物理信息神经网络的石化污水处理过程智能预测研究

融合自适应证据的深度学习框架预测PFAS环境迁移性质