Dynamic measurement of gas volume fraction in a CO2 pipeline through capacitive sensing and data driven modelling
Dynamic measurement of gas volume fraction in a CO2 pipeline through capacitive sensing and data driven modelling
复制标题
DOI：
10.1016/j.ijggc.2019.102950
复制
发表时间：
2020-03
影响因子：
3.9
通讯作者：
Ding Shao;Yong Yan;Wenbiao Zhang;Shijie Sun;Caiying Sun;Lijun Xu
中科院分区：
文献类型：
--
作者：
Ding Shao;Yong Yan;Wenbiao Zhang;Shijie Sun;Caiying Sun;Lijun Xu
CO2气液两相流气体体积分数（GVF）的测量是碳捕集与封存（CCS）技术应用的关键。提出了一种利用12电极电容传感器测量CO2两相流GVF的新方法。利用电容数据分别建立了基于反向传播神经网络（BPNN）、径向基函数神经网络（RBFNN）和最小二乘支持向量机（LS-SVM）的数据驱动模型。在数据预处理阶段，Copula函数用于选择特征变量并为数据驱动模型生成训练数据集。在CO2气液两相流实验装置上进行了稳态流动条件下的实验研究，实验条件为液体CO2的质量流量为200 ~ 3100 kg/h，GVF为0%~ 84%。由于具有CCS能力的发电公司的灵活操作，还在测试台上进行了GVF快速变化的动态实验，以评估数据驱动模型的实时性能。稳态流场测量结果表明，RBFNN的相对误差在±7%以内，优于其他两种模型。在动态流场条件下的仿真结果表明，该神经网络能够跟踪GVF的快速变化，误差在± 16%以内。
Gas volume fraction (GVF) measurement of gas-liquid two-phase CO2flow is essential in the deployment of carbon capture and storage (CCS) technology. This paper presents a new method to measure the GVF of two-phase CO2flow using a 12-electrode capacitive sensor. Three data driven models, based on back-propagation neural network (BPNN), radial basis function neural network (RBFNN) and least-squares support vector machine (LS-SVM), respectively, are established using the capacitance data. In the data pre-processing stage, copula functions are applied to select feature variables and generate training datasets for the data driven models. Experiments were conducted on a CO2gas-liquid two-phase flow rig under steady-state flow conditions with the mass flowrate of liquid CO2ranging from 200 kg/h to 3100 kg/h and the GVF from 0% to 84%. Due to the flexible operations of the power generation utility with CCS capabilities, dynamic experiments with rapid changes in the GVF were also carried out on the test rig to evaluate the real-time performance of the data driven models. Measurement results under steady-state flow conditions demonstrate that the RBFNN yields relative errors within ±7% and outperforms the other two models. The results under dynamic flow conditions illustrate that the RBFNN can follow the rapid changes in the GVF with an error within ±16%.