基于机器学习的汽油精制工艺操作方法优化

降低辛烷值损失是石化企业催化裂化汽油脱硫精制工艺过程中的重要目标之一。针对该工艺过程中控制变量维度高且存在非线性和强耦联性等问题，研究利用皮尔森、斯皮尔曼、最大信息系数三种方法，对操作变量进行相关性分析及特征降维，选取与辛烷值损失强相关的21个主要变量，建立了基于XGBoost辛烷值损失预测模型，交叉验证平均准确率达96.54%；然后，提出以硫含量不大于5 ug/g 为约束的工艺操作方法优化模型实现辛烷值损失最小，并通过遗传算法-聚类递归的方法进行求解，确定主要变量取值。以133号样本为例的模型可视化结果表明，所提出的优化模型可以在主要变量逐步调整过程中实现硫含量降至最低点，辛烷值损失接近最小。这既验证了模型的有效性和可移植性，也为汽油精制工艺提供了量化科学优化支撑。

Optimization of Gasoline Refining Process Based on Machine Learning

A main target of catalytic cracking gasoline desulfurization by petrochemical enterprises is to reduce the loss of octane value. Aiming at problems existing in the refining process, including high dimension, nonlinearity and strong coupling of control variables, Pearson, Spearman and maximum information coefficient were adopted for correlation analysis and feature dimension reduction of operating variables. Twenty-one main variables strongly correlated to octane value loss were selected for the establishment of an octane value loss prediction model based on XGBoost, and the average accuracy rate of cross-validation reached 96.54%. Then a technology optimization model with a constraint of sulfur content not exceeding 5 ug/g was proposed to minimize the loss of octane value. Furthermore, a genetic algorithm-recursive clustering method was employed for a solution so as to determine the value of main variables. The model visualization results of sample No.133 indicate that the proposed optimization model can reduce the sulfur content and octane value loss to a minimum through the gradual adjustment of major variables. This not only verifies the effectiveness and portability of the model, but also provides quantitative scientific optimization for gasoline purification.