A novel heterogeneous ensemble approach to variable selection for gas-liquid two-phase CO2 flow metering
A novel heterogeneous ensemble approach to variable selection for gas-liquid two-phase CO2 flow metering
复制标题
DOI：
10.1016/j.ijggc.2021.103418
复制
发表时间：
2021-09
影响因子：
3.9
通讯作者：
Caiying Sun;Lijuan Wang;Yong Yan;Wenbiao Zhang;Ding Shao
中科院分区：
文献类型：
--
作者：
Caiying Sun;Lijuan Wang;Yong Yan;Wenbiao Zhang;Ding Shao
变量选择是开发有效的数据驱动模型以测量碳捕获和存储系统中二氧化碳流量的重要预处理步骤。为了有效量化潜在输入变量对期望输出的重要性，提出了集成学习并将其纳入变量选择方法中。本文提出了一种基于树的异构集成方法来进行变量选择及其在气液两相 CO2 流量测量中的应用。每个变量的重要性是通过组合四种基于树的算法的重要性分数来确定的，包括决策树回归、回归树的引导聚合、梯度提升决策树和梯度提升随机森林。然后应用向后消除算法来删除相对不太重要的变量，从而删除数据驱动模型的一小组输入变量。选择结果表明，CO2质量流量测量的重要变量包括表观质量流量、时移、压差和压降，同时观测密度、密度降、观测流速和出口温度用于预测气体体积分数。为了评估所选变量的有效性，开发了基于梯度增强随机森林的数据驱动模型。结果表明，以所选变量作为模型输入，CO2质量流量测量的模型输出相对误差大多在1%以内，气体体积分数预测的相对误差大多在5%以内。
Variable selection is an important preprocessing step in the development of effective data-driven models for CO2flow measurement in carbon capture and storage systems. In order to effectively quantify the importance of potential input variables to the desired output, ensemble learning is proposed and incorporated into variable selection methodology. This paper presents a tree-based heterogeneous ensemble approach to variable selection and its application to gas-liquid two-phase CO2flow measurement. The importance of each variable is determined through combining the importance scores from four tree-based algorithms, including decision tree regression, bootstrap aggregating of regression trees, gradient boosting decision tree and gradient boosting random forest. Then the backward elimination algorithm is applied to remove the relatively less important variables and hence a small set of input variables for data-driven models. The selection results demonstrate that the significant variables for CO2mass flow measurement includeapparent mass flow rate, time shift, differential pressureandpressure dropwhileobserved density, density drop, observed flow velocityandoutlet temperaturefor prediction of gas volume fraction. To assess the validity of the selected variables, data-driven models based on gradient boosting random forest are developed. Results suggest that the relative error of the model output is mostly within 1% for CO2mass flowrate measurement and 5% for gas volume fraction prediction by taking the selected variables as model inputs.