摘要
由于加氢裂化装置中柴油产量影响因素多且复杂,常因关键指标测量困难造成生产效率低、产品质量差。为此提出MLP-RF(多层感知器-随机森林)组合模型对某石化公司芳烃厂加氢裂化装置现场数据进行预处理和模型训练,以期有效预测柴油产量。首先,采用滑动平均滤波法对原始数据降噪处理,借助Z-score标准化法消除特征数据间的尺度差异,筛选了装置中原料性质和操作条件的特征变量。其次,在模型内部构建权重量化器,为组合模型动态分配权值,实现优化。结果显示该组合模型在加氢裂化工艺中柴油产量预测时,比单一MLP、RF模型在原始数据降噪和预测精度上都具有显著优势,为生产过程控制与优化提供理论依据。
The factors affecting productivity of diesel in hydrocrackers are quite a lot and complex,resulting in low efficiency and inferior quality in production due to the difficult measurement of key indicators.Therefore a combined MLP-RF model(maltilayer perceptron-rondom forest algorithm)was proposed to effectively predict the diesel production by preprocessing field data and model training of the hydrocracking unit of aromatics plant of a petrochemical company.At first,the noise reduction of the original data was carried out by sliding average filtering,then the scale difference among the characteristic data was smoothed by Z-score normalization.After that,the characteristic variables of the properties of raw materials and operating conditions in the facility were screened.And then,a weight quantizer was constructed inside the model to dynamically allocate the weights to optimize the combined model.The results show that the combined model presents a more prominent advantage in terms of noise reduction of original data and prediction accuracy compared with the single models of MLP or RF for the application of the prediction of diesel yield in hydrocracking process,which provides theoretical basis on the control and operation of the production process.
作者
谢洋
陆新建
邵晓雯
吴永红
张兵
XIE Yang;LU Xinjian;SHAO Xiaowen;WU Yonghong;ZHANG Bing(School of Petroleum and Chemical Engineering,Shenyang University of Technology,Liaoyang 111003,Liaoning Province,China;Nanjing ChemCyber Technology Company Ltd.,Nanjing 210899,Jiangsu Province,China)
出处
《化学工程》
CAS
CSCD
北大核心
2024年第11期66-70,共5页
Chemical Engineering(China)
基金
辽宁省自然科学基金资助项目(2021-MS-238)
辽宁省教育厅科研项目(LJGD2020002)
沈阳市中青年科技创新人才项目(RC200325)。
关键词
MLP-RF模型
软测量
加氢裂化
特征选择
预测
机器学习
MLP-RF model
soft measurement
hydrocracking
feature selection
forecasting
machine learning
