论文部分内容阅读
由未取心井段的测井信息估算渗透率是一个重要问题,也是很多地学学科遇到的难题。估算渗透率的最常用方法不是应用经验公式,就是应用某种形式的多元线性回归(MLR)技术。比较复杂一点的多元非线性回归(MNLR)技术,由于涉及到近似数学模型的选择和所选模型对输入变量的敏感性分析等困难而很少采用。但新近发展的一类非线性优化人工神经网络(ANNs)技术比较好地克服了这些困难。 我们应用反向传播ANN(BP—ANN)模拟了空间位置和六种不同的测井曲线与渗透率之间的关系。取自Venture气田的四口井中数据形成BP—ANN模拟所需的训练数据集和监督学习数据集。在同一气田,保留五分之一的井中数据作为独立数据集用作试验。用这些试验数据作试验时,训练后的BP—ANN产生的渗透率和取心井段实测的渗透率匹配得很好。由训练后的BP—ANN计算出的渗透率剖面揭示出许多低渗透率层、并可在井间进行对比。这些层很可能是储层内流体受阻的原因,对进一步油藏开发计划有着重要意义。 为便于讨论,我们采用了同样数据集的常规统计法预测方程(即MLR和MNLR)。这些例子着重强调了BP—ANN方法在解决多变量、非线性模拟,诸如渗透率估算等难题中的效用。
The estimation of permeability from the logs of un-cored wells is an important issue and also a difficult problem for many geosciences. The most commonly used method of estimating permeability is not to apply empirical formulas or to apply some form of multivariate linear regression (MLR). The more complex multivariate nonlinear regression (MNLR) technique is seldom used due to the difficulties associated with the selection of approximation math models and the sensitivity analysis of selected models to input variables. However, a newly developed class of nonlinear optimization artificial neural network (ANNs) technology overcomes these difficulties well. We use backpropagation ANN (BP-ANN) to simulate the relationship between spatial location and six different well logs and permeability. The data from four wells taken from the Venture field formed the training datasets and supervised learning datasets needed for BP-ANN simulations. In the same field, one fifth of well data was retained as a separate dataset for testing. When tested with these experimental data, the permeability produced by the trained BP-ANN matched well with the measured permeability of the coring well. The permeability profiles calculated from the trained BP-ANNs reveal many low-permeability layers and can be compared between wells. These layers are likely to be the cause of obstructed fluid in reservoirs and have important implications for further reservoir development planning. For the purposes of this discussion, we use the same statistical set of equations (MLR and MNLR) for the same data set. These examples highlight the utility of the BP-ANN method in solving problems such as multivariable nonlinear simulations such as permeability estimation.