【作者】 邢兰昌；

【导师】 耿艳峰；

【作者基本信息】 中国石油大学 ， 检测技术与自动化装置， 2008， 硕士

【摘要】 凝析天然气的在线计量属于多相计量的一个分支,是石油天然气工业迫切需要解决的问题之一。本文将其简化为低含液率气液两相流计量问题,基于课题组自行开发的凝析天然气流量计样机和大量的实验数据。本文为流量计开发一种基于软测量模型的快速、高效的数据处理方法,扩大流量计的测量范围、提高测量精度和实时性,为流量计计量算法的完善奠定基础。研究工作分为两部分,第一部分对单个槽式孔板测量数据进行处理,先后完成了基于差压平均值的孔板压降倍率特性研究和基于差压波动部分的特征量提取;第二部分为基于双槽式孔板组合的两相流参数检测技术研究,分别为基于相关技术的相含率测量和基于神经网络的流型辨识和流量测量。基于差压信号的平均值,详细分析了不同孔径比的槽式孔板ΦG随X、Frg和γ变化的规律,并将X、Frg和γ作为变量对Murdock相关式进行修正,提出了一种新的适用于不同孔径比槽式孔板的压降相关式;与前人基于标准孔板的相关式相比,由于修正了γ对ΦG的影响,在文中实验条件下其流量计算精度最高。基于差压信号具有强烈的非线性、非平稳性和非高斯性,应用Hilbert-Huang变换和高阶统计分析方法对差压信号波动部分进行处理,提取了一组新的特征量:IMF分频段能量及能量分数、IMF熵、HHT熵和双谱熵;分析表明:上述特征量对气液两相流动参数的变化非常敏感,为建立气液两相流参数检测的软测量模型提供了依据。基于流量计的双槽式孔板结构,应用小波变换和经验模式分解对上下游差压信号进行处理,分别将分解结果作为相关量计算相关速度,研究了相关速度与X及液相含率之间的关系;结果表明:原始差压信号得到的相关速度不能很好地反映两相流动参数的变化,而由分解后的某些分量获得的相关速度与相含率之间具有稳定的关系。提出了一种新的信号特征选择方法——mRMR+BP,分别以mRMR+BP方法和主成分分析法对差压信号特征量进行预处理;以流型和气液分相流量作为BP网络输出,以从上下游差压信号中提取的两组特征量及其组合作为网络输入,建立了三个独立的子网络;通过对子网络的输出进行集成建立了基于集成BP神经网络的软测量模型;在文中实验条件下,软测量模型的流型辨识准确度高于93%,气液分相流量测量的平均相对误差分别低于5%和15%。更多还原

【Abstract】 Wet gas metering is a subset of multiphase flow metering, and it is one of the major unsolved problems for oil and gas industry. In this thesis it is simplified to the metering of gas-liquid two-phase flow with low liquid fractions. Based on a self developed prototype of wet gas meter and experimental data set, a series of works have been carried out. The main target of this thesis is to develop a fast and efficient data processing method, and to improve the performance of wet gas meter, such as expending the metering range, high accuracy, fast response and so on.This thesis consists of two parts. The first part is mainly focus on the metering characteristics of a single slotted orifice, which included the two-phase multiplier analysis based on the mean value of differential pressure, and the dynamic feature extraction based on fluctuation of the differential pressure. The second part is mainly focus on the metering algorithm development, which included the phase fraction measurement based on correlation analysis, flow regime identification and flow rate measurement based on neural networks.Based on the mean value of differential pressure, detail analysis about how two-phase multipliers changing with X、Frg andγwas made for slotted orifices with different beta ratios. A new correlation for slotted orifice was put forward based on the modification to the Murdock correlation. The proposed correlation shows a more accurate calculation results compared with the existing correlations for standard orifice plates as the effect ofγhas been considered.Based on the recognition that the differential pressure is nonlinear, non-stationary and non-Gaussian, Hilbert-Huang transform and high-order statistics were employed to analyze the differential pressure. A group of features such as the energy and fraction of IMFs, entropy of IMF, HHT and bispectrum were extracted and qualitative analysis shows that the features are sensitive to the variation of flow parameters. The results provide the basis for the building of soft-sensing models for flow parameters measurement.Based on the dual slotted orifices of wet gas meter, the upstream and downstream differential pressures were processed through wavelet analysis and empirical mode decomposition approaches. Correlation velocities were obtained based on the corresponding components of a, d and IMFs, and the relationship between correlation velocities and X, phase fraction was investigated. The results show that the relationship between X, phase fraction and correlation velocities calculated by the above signal components are more stable than those by the original differential pressures.A novel feature selection approach called mRMR+BP was put forward and used to preprocess the original features together with the principal component analysis method. Flow regimes and gas/liquid individual phase flow rates were taken as outputs of the BP network, and those selected features as inputs, then three sub networks were built. Soft-sensing models for flow regimes identification and flow rates measurement were built based on neural networks ensemble. The results show that the neural network ensemble based model is more accurate than any sub network, and the accuracy rate of flow regimes identification is above 93%, and the mean relative errors are below 5% and 15% for gas and liquid flow rates respectively.更多还原