摘要
变量筛选是现代工业过程产品质量预测研究中的热点问题之一。过滤式变量选择方法因其计算速度快且不易造成过拟合得到了广泛应用,但其存在容易忽略变量相关性且不能准确反映工况信息的问题,在高维数据维度灾难问题日渐突出的当今不再适用。针对这一问题,提出一种分步约简的敏感变量选择方法。该方法在明确敏感变量和关键敏感变量的基础上,根据变量对工况的描述能力和辅助变量与主导变量的净相关性定义了敏感性指标,实现敏感变量的初选;接着,构建加权余弦马田系统以解决变量冗余性问题,实现敏感变量的精选。所提方法应用于加氢裂化产品质量预测,实际工业应用结果表明,该方法不仅可以提高模型的预测精度,而且可以有效地降低模型复杂性。
Variable selection has always been a hotspot in product quality prediction of modern industrial production.Due to the fast calculation speed and do not easily cause overfitting,filter methods have been widely used.However,because the correlation of variables is easily ignored and no information of working condition reflected in the filter methods,it is no longer suitable for solving the disaster problems of high-dimensional data.To cope with these problems,a fractional step reduction method for sensitive variable selection is proposed in the paper.First,two concepts of sensitive variables(SV)and key sensitive variables(KSV)are first identified.Then,a sensitivity index(SI)is developed to select the sensitive variables preliminarily.After that,a weighted cosine Mahalanobis-Taguchi system(WCMTS)is proposed to precisely select the key sensitive variables.Finally,the proposed method is applied to an industry hydrocracking process.Practical industrial application results show that the method can not only improve the prediction accuracy of the model,but also effectively reduce the complexity of the model.
作者
李灵
王雅琳
孙备
LI Ling;WANG Yalin;SUN Bei(School of Automation,Central South University,Changsha 410083,Hunan,China)
出处
《化工学报》
EI
CAS
CSCD
北大核心
2020年第5期2173-2181,共9页
CIESC Journal
基金
国家自然科学基金重大项目(61590921)
国家重点研发计划项目(2018AAA0101603)。
关键词
敏感性指标
变量选择
预测
系统工程
石油
sensitivity index
variable selection
prediction
systems engineering
petroleum
