渗透率是确定油气田储层特性的最重要参数之一。在未取心的层段或井中,能采用测井数据进行油藏描述和油藏评价的方法在技术和经济上都有很大优势,因为它可以在不取心的情况下提供一口井的连续记录。但用测井数据确定非均质地层渗透率时还存在着一些需要用常规统计方法解决的复杂问题。近年来,采用人工神经网络(ANN)方法成功地解决了储层渗透率估算中出现的许多问题。不过,在将神经网络用于映射复杂的非线性关系时,尽管其功能性很强,但仍然暴露出一些技术问题。本文提出了一种基于多项式神经网络(PNN)的数据处理组方法(GMDH),它根据测井数据预测渗透率,避免了常规神经网络方法的某些缺陷。PNN综合了网络大小、连通性、处理元素的类型,以及训练后得到的全局优化结构系数。这种自组织方法以多项式形式自动显示数据间的内在关系,增强了最终数据学习模型的数据逼近能力和解释能力。证明该方法的数据取自韩国海上油田。与常规神经网络的对比研究表明,虽然PNN模型的预测精度会受到测量数据误差的影响,但仍然能提供相对可靠的性能。PNN是根据测井数据预测非均质地层储层渗透率的一种非常实用、有效的工具。 Permeability is one of the most important parameters in determining reservoir properties. In un-cased intervals or wells, methods that can be used for reservoir characterization and reservoir evaluation with log data are technically and economically significant because they provide a well without cogging Continuous recording. However, there are still some complex problems that need to be solved by conventional statistical methods when using logging data to determine the heterogeneity of formation permeability. In recent years, artificial neural network (ANN) method has successfully solved many problems in reservoir permeability estimation. However, while using neural networks to map complex non-linear relationships, despite their high functionality, some technical problems are still exposed. This paper presents a data processing group method (GMDH) based on polynomial neural network (PNN), which predicts permeability based on log data and avoids some of the drawbacks of conventional neural network methods. The PNN combines the network size, connectivity, types of elements handled, and global optimization structure coefficients obtained after training. This self-organizing approach automatically displays the internal relationships between the data in a polynomial fashion and enhances the data approximation and interpretation capabilities of the final data-learning model. The data that prove this method are taken from the offshore oilfield in Korea. Compared with the conventional neural network, it shows that although the prediction accuracy of PNN model is affected by the error of measurement data, it can still provide relatively reliable performance. PNN is a very practical and effective tool to predict the permeability of heterogeneous formations based on log data.