节点文献
基于多属性神经网络的冲积扇构型识别
Alluvial Fan Configuration Recognition Based on Multiple Attributes Neural Network
【作者】 王伟;
【导师】 李少华;
【作者基本信息】 长江大学 , 矿产普查与勘探, 2012, 硕士
【摘要】 地震属性分析技术是近几年迅速发展起来的一种储层预测技术,它主要是应用多种数学分析(如神经网络等)方法,从地震数据体中提取有关储层物性、岩性信息的多种属性,同时结合工区内测井资料,建立井旁道目标曲线与地震属性的关系,使用相关度最大的属性预测和估算整个地震数据体的目标曲线特征,来达到储层预测目的。概率神经网络是以径向基函数为内核的一种神经网络,采用了局部化的作用函数,具有最佳逼近特性,且没有局部极小值。概率神经网络方法可以看作线性聚类分析法的非线性扩展,是一种利用神经网络结构实现的数学插值方法。概率神经网络方法具有高度的容错性,即使某个井旁道地震参数或某个网络连接有缺陷,也可以通过联想得到全部或大部分信息。因此,用概率神经网络建立地震属性和测井特征属性之间的映射关系可靠性高。概率神经网络方法还具有动态适应性,当地质岩性类别变化或地震参数修改时,网络可自动适应新的变量,调整权系数,直到收敛。对于受岩性控制的储层,概率神经网络是描述其地震属性参数与岩性参数关系的有效方法。储层构型亦称为储层建筑结构、结构单元等,是指储层内部不同级次构成单元的形态、规模、方向及其相互之间的叠置关系。构型的概念最初由美国沉积学家Miall于1985年提出,之后很多学者开展了储层构型研究。构型研究突出体现了储层内部沉积单元的层次性和结构性,这一研究在河流相储层分析中取得了较大的进展,而且在地下储层研究中得到了较好的应用。然而,冲积扇储层构型研究程度很低,虽然近年来有学者对冲积扇储层构型进行了初步的探索,但在建立判别模式的预测精度和三维构型建模方法的研究方面仍需开展深入的研究。本文综合分析了克拉玛依油田六中东区密井网区的静、动态资料,在前人研究的基础上,重点针对冲积扇储层内部构型测井参数判别模式,单元的几何学特征及空间组合关系开展研究,利用多属性神经网络的建模方法对砂砾岩储层内部构型建模的适用性进行研究,研究冲积扇储层内部构型单元的定量几何特征与叠置模式,建立反映多级次构型界面的三维储层构型模型。完成的工作主要包括以下五大方面:(1)测井曲线三维数据场的建立对六中东区的277口井的井头、井斜、分层和测井曲线数据进行了整理,优选了了与构型相关度较大的7条测井曲线,采用序贯高斯模拟方法建立了7条测井曲线的三维数据场,并将其转换为SEGY格式的地震数据体。(2)神经网络方法的选择综合分析了BP神经网络、多层前馈神经网络和概率神经网络方法的基本原理,对这三种方法在六中东区的构型识别精度进行了分析比较,通过对比,选择了概率神经网络用于六中东区的构型识别。(3)神经网络训练应用多属性线性回归方法对提取的24种叠后地震属性和7种测井属性进行了优选,对概率神经网络的预测参数进行了多次实验,综合比较后选择了预测误差小并且训练时间相对较小的参数设置。(4)未取心井构型的划分利用八口密闭取心井的构型曲线和测井曲线进行神经网络训练,建立构型曲线和测井曲线的判别模式,将此判别模式应用到未取心井上,识别未取心井的构型曲线,从而进行单井构型要素划分与对比。(5)六中东区的构型建模以六中东区的8口小井距密闭取心井的构型曲线为目标曲线,采用优选出的最佳属性以及概率神经网络参数对六中东区的地震体进行了反演,得到了六中东区构型模型。构型模型较好的表征了六中东区扇根、扇中和扇缘的构型单元的构成特点及叠置关系,基于多属性神经网络的冲积扇构型建模为储层的非均质性研究提供了依据,同时也为构型模型的建立提供了新的思路。
【Abstract】 Seismic attribute analysis is a rapid development of a reservoir prediction technology in recent years, which is the application of a variety of mathematical analysis (such as neural network) methods to extract about the reservoir properties and variety attributes of lithology from seismic data, combined with logging data of the work area to establish the target curve and the relationship of the seismic attributes of the well. Then use the maximum related properties of the seismic data to forecasts and estimates the target curve characteristics, eventually reach the reservoir prediction purposes.Probabilistic neural network is a neural network based on radial basis function kernel, it is using localized functions, the best approximation characteristics, and no local minima. The probabilistic neural network approach can be seen as a nonlinear extension of linear cluster analysis. It is a mathematical interpolation method which implemented the neural network structure. The probabilistic neural network has a high degree of fault tolerance, even if the seismic parameters of a well or a network connection is defective, all or most of the information available through Lenovo. Thus, the mapping between the property of seismic attributes and logging features is of high reliability by using the probabilistic neural network. The probabilistic neural network approach also has the dynamic adaptability, the network can automatically adapt to the new variable to adjust the weight coefficients, until convergence, when geological lithology category or seismic parameters changed. The probabilistic neural network is an effective method to describe the relationship between seismic attributes and lithology parameters for the reservoirs controlled by lithology.Reservoir architecture, also known as the reservoir architectural, structural unit, and means of different levels of the reservoir internal superposition between cell shape, size, direction, and their mutual relations. Configuration concept originally developed by the American sedimentologists Miall1985, followed by many scholars to carry out reservoir architecture studies. Configuration study highlights the level of the reservoir internal sedimentary units and structural, this research has made great progress in fluvial reservoir analysis, and a better application in the underground reservoir. However, the degree of structure of the alluvial fan reservoir is very low, although in recent years some scholars have conducted a preliminary exploration of the alluvial fan reservoir architecture, but much work remains to be done in-depth study in terms of prediction accuracy to develop the discriminate model and three-dimensional configuration of modeling methods.The paper analysis the static and dynamic data of the Eastern zone of dense well in six of the Karamay oil field, on the basis of previous research, focusing on the alluvial fan reservoir internal configuration of logging parameters distinguish the mode unit geometry characteristics and spatial combination related research, the use of multi-attribute neural network modeling method to study the applicability of sandy conglomerate reservoir internal architecture modeling, internal configuration unit of alluvial fan reservoir quantitative geometrical features and overlay mode, set up to reflect the multi-level configuration interface,3D reservoir model configuration. The main contents include the following five areas:(1) Establish of logging curve three-dimensional data fieldFinishing277wells in the Eastern of the Sixth District, well inclined, layered and well logs data organized, preferably7well logs in larger configuration, the use of sequential Gaussian simulation method7three-dimensional data field of the well logs and seismic data in SEGY format.(2) The choice of the neural networkComprehensive analysis of the basic principle of BP neural network, multilayer feedforward neural network and probabilistic neural network method, the three methods in pattern recognition accuracy of the Sixth Eastern, by contrast, chose the probabilistic neural network six in the Eastern District of pattern recognition.(3) Neural network trainingUsing multi-attribute linear regression method to preferred24kinds of post-stack seismic attributes and seven kinds of logging property, carried out several experiments to forecast parameters of the probabilistic neural network, the comprehensive comparison of the prediction error is small and the training time relatively small set of parameters.(4) Configuration division of the un-coring wellsEight sealed coring structure curves and log curve for neural network training, the establishment of the configuration curves and log-curve determination mode, this discriminant model applied to the coring Inoue, recognition does not take the structure of the core hole curve, and thus the division and correlation of a single well configuration elements.(5) Configuration modeling of Eastern District in the six Taken to the Sixth Eastern8SLIM from the closed configuration curve of the core hole is the target curve, The optimized properties, and probabilistic neural network parameters inversion of seismic body of the Sixth East, six in the Eastern configuration model.Configuration model is better characterization the constituent characteristics and the superimposition of the configuration unit of the fan root, fan and fan fringe of the six Middle East alluvial fan. Alluvial fan configurations modeling based on multi-attribute neural network provide a new method for the reservoir heterogeneity, and also a new approach for the configuration modeling.
【Key words】 Multi-attribute analysis; Neural network; Three-dimensional data field oflogging curve; Configuration model;
- 【网络出版投稿人】 长江大学 【网络出版年期】2013年 01期
- 【分类号】P618.13;P631.4
- 【被引频次】1
- 【下载频次】439