基于LSTM的润滑油加氢装置产品预测和基于RF的操作优化相关性分析

基于某炼油厂润滑油加氢装置的生产工艺及实际生产数据，采用Aspen HYSYS软件建立了润滑油加氢装置的生产过程机理模型，并用随机抽样的方法验证了模型的有效性；然后根据生产工艺条件运行机理模型，扩展了装置的产品预测数据集。对比BP神经网络，采用LSTM(Long Short-Term Memory)神经网络，以加氢裂化反应器入口温度、加氢异构反应器入口温度、加氢后精制反应器入口温度、加氢裂化后常压分馏塔塔顶温度、加氢异构常压分馏塔塔顶温度、加氢裂化反应器压力、加氢异构反应器压力为输入，以轻质润滑油常温常压下的质量流量、40 ℃运动黏度、闪点和倾点4个目标为输出建立了预测精度更高的产品预测模型。通过随机森林(RF, Random Forest)算法对该装置产品预测数据集的输入特征变量与输出目标变量进行了相关性分析，确定了特征重要度排序，得到了不同生产目标对应的优化操作方案。

Product Prediction of Lubricating Oil Hydrogenation Unit based on LSTM and Optimal Operation Correlation Analysis with RF

Based on the technological process and the limited actual production data of the lubricating oil hydrogenation unit in a real-world refinery, we firstly developed a mechanism model on the Aspen HYSYS software platform and verified its effectiveness with the random sampling method. Then the product prediction data set of the lubricating oil hydrogenation unit was expanded with the mechanism model and the technological process requirement. Compared with the BP neural network, the LSTM (Long Short-Term Memory) neural network which has higher prediction accuracy was selected to model the production process of the lubricating oil hydrogenation unit. The inlet temperatures of the hydrocracking reactor, the isomerization dewaxing reactor and the hydrofining reactor; the fractionating reactor’s tower top temperatures of the hydrocracking reactor and the isomerization dewaxing reactor, and the pressure values of the hydrocracking reactor and the isomerization dewaxing reactor were selected as input characteristic variables. The mass flow rate at the ambient temperature and pressure, kinematic viscosity at 40 ℃, flash point, pour point of the light lubricating oil were selected as the output target variables. The correlation analysis among the characteristic variables and each target variable were computed with the Random Forest (RF) model. The optimized operational schemes related to different production targets have been obtained.