利用聚类分析方法进行模型优选

Application of clustering analysis in model optimization

戴危艳¹ 李少华¹ 王军² 宋道万² 史敬华² 陈苏²

长江大学地球科学学院，湖北 武汉430100 中国石化胜利油田分公司地质研究院，山东 东营 257015

School of Geosciences, Yangtze University, Wuhan 430100, China Geological Research Institute, Shengli Oilfield Company, SINOPEC, Dongying 257015, China

运用储层随机模拟技术能够产生多个等概率的实现，而在油藏数值模拟中，考虑到成本计算，通常只能够对有限的几个实现进行模拟计算。为了从多个模型中快速有效地选出具有代表性的模型，文中提出利用K-means聚类分析方法进行模型优选。以WZ油田西区为例，采用相控物性参数建模技术，利用顺序高斯模拟方法建立渗透率的三维模型，计算每两个模型之间相对应的每一网格节点的渗透率值差的平方和，然后取平方根，得到一个表征各模型之间差异的相异性矩阵；应用多维尺度分析技术对矩阵进行降维，实现在二维空间中用向量来可视化模型的相似性；用K?鄄means聚类分析方法对模型进行聚类，采用Dunn指标对聚类结果进行评价，结果显示，k值为5时的聚类效果最好；最后，通过对比所选模型计算的储量与用蒙特卡洛模拟法得到的P10，P50和P90储量，说明了方法的可行性。

Reservoir stochastic simulation can generate multiple realizations in equal probability. While in reservoir simulation, considering the computational cost, it can only simulate a limited number of realizations in practice. In order to select several representative models from the multiple models quickly and efficiently, this paper presents a method to introduce K-means clustering for model optimization. Take the western WZ Oilfield as an example, the modeling technique of facies controlled physical parameters is used to establish 3D permeability models with Gauss simulation method. The square sums of permeability value difference of each grid node in two models are calculated, and taking the sum square root, a dissimilarity matrix which characterizes the differences between models is obtained. In order to to visualize the similarity of model in 2D space by vectors, the dimension of matrix through multidimensional scaling analysis is reduced. The models in K-means cluster algorithm are clustered, and the clustering results are evaluated by Dunn index, the results show that the clustering effect is best when k is 5. Finally, comparing the geological reserves calculated by this model with the reserves for P10, P50 and P90 by Monte Carlo simulation method shows that this method is practical.

10.6056/dkyqt201504019