基于机器学习与层次分子基团构建的纯组分炼化性质预测建模方法研究

Molecular management method can optimize the petroleum refining in the level of molecular composition and transition. Whether the method is utilized in petroleum industry is highly depended on the accuracy of pure component refinery properties. The state-of-art models are lacking of the accuracy since the molecular features are not appropriate to represent the refinery properties and the modeling methods cannot deal with the sparsity of the given data bases. These reasons lead to the prediction accuracy cannot meet the requirement of petroleum industry. The scope of the present project is to develop a novel set of molecular features to represent the refinery properties of pure components by combing mechanism and machine learning knowledge in the level of molecular structure. Furthermore, based on the novel molecular features, a multi-level group contribution model is developed with the capability of solving the sparse data sets by combing statistics methods. The model can precisely predict the molecular refinery properties including research and motor octane number and centane number. In order to increase the practical values, the apply domain of the model are estimated using unsupervised learning methods. In addiction, multi-objective optimization methods are utilized to simultaneously optimized the accuracy and apply domain of the developed model. The present models will precisely estimate the pure component refinery properties to meet the accuracy requirement of petroleum industry. The project will promote the understanding for the relationships between molecular structures and refinery properties. It also will complete the molecular management methods.

分子管理可从分子组成与转化层面优化石油炼化过程，对石油产物中纯组分炼化性质预测的准确与否，直接关系着分子管理能否实际应用于工业界。当前尚缺少合理表征炼化性质的分子结构特征集，以及适用于已有小样本数据库的建模方法，导致预测精度难以满足工业需求。本课题的目的是结合机理与机器学习，从分子结构级别，构建适用于分子炼化性质的特征集。并基于此，引入统计学方法构建适用于稀疏数据集的层次基团模型，以准确预测纯组分炼化性质中研究法、马达法辛烷值与十六烷值。为了提高模型的实际应用价值，引入无监督学习方法对模型适用范围做深入研究。并结合多目标优化方法，同时优化模型精度与适用范围，从而实现对石油产物中纯组分各炼化性质的精确估计以满足工业需求。本项目的成功实施将加深研究者理解分子结构对炼化性质的影响，同时完善分子管理方法，为其实际应用于炼化行业提供支撑。

随着未来对生产清洁汽油的需求逐渐增大，对评价汽油组分中分子性能的要求也趋向更为全面和精确。针对汽油流程，本研究建立起一套完整的评估体系，着重研究对高性能清洁汽油最为关键的五方面性能，包括抗爆性、清洁性、动力性、稳定性及蒸发性能。对应于此五方面性能，选择五个可定量的关键指标代表所对应性能的优劣，分别是辛烷值、生烟指数、燃烧热、不饱和双键含量以及雷德蒸气压。进而分别建立了结构性质关联模型以预测汽油分子的在各指标中的表现。综合五个指标的模型预测结果，构建汽油分子综合性能评价模型，以评价汽油分子的优劣，指导实际生产性能更为优越的产品分子。.进一步，将此综合评价模型嵌入本人读博期间所在课题组的汽油分子自动设计模型，以搜寻性能更为优越的产品分子结构，评价模型作为判别新生成分子优劣的判别依据，扮演重要角色。自动分子设计模型是为优化能源分子结构而提出的新晋技术。之前所在课题组已建立汽油加工单元所需的反应规则，完全依靠计算机，原料中的各分子结构经过一系列反应规则计算后可生成多种汽油组分分子。该模型旨在通过发现最佳的汽油组分分子结构，倒推最佳的汽油加工反应路径，进而更好地指导实际生产。使用该模型研究车用及航空汽油，优化得到最佳的分子结构。此分子结构不但拥有最佳的性能表现，同时兼具易获取性，即可通过较短的反应路径与较缓和反应条件获得。