逐类组合支持向量机在致密储层判识和产能预测中的应用  

Applying of the termwise-combination support vector machine to identification of reservoir and prediction of deliverability in tight reservoirs

作　　者：庞河清[1] 匡建超[1,2] 蔡左花[3] 王众[1] 黄耀综 

机构地区：[1]成都理工大学能源学院,四川成都610059 [2]成都理工大学管理科学学院,四川成都610059 [3]西南油气分公司勘探开发研究院贵阳所,贵州贵阳550004 [4]胜利油田分公司孤东采油4厂,山东东营257237

出　　处：《大庆石油学院学报》2011年第6期35-40,124-125,共6页Journal of Daqing Petroleum Institute

基　　金：教育部规划基金项目(11YJAZH043);四川石油天然气研究中心重点资助项目(川油气科SKA09-01)

摘　　要：针对致密储层判识和产能预测准确率低,提出一种新的建模方法——逐类组合支持向量机方法(TCSVM).首先应用支持向量分类机(SVC)实现储层类别判识,然后用支持向量回归机(SVR)建立气层产能预测模型,最后对未知储层进行判识和产能预测.该模型通过前期降噪、降维的属性优化,有效降低数据类别对储层判识的干扰,提高储层判识和气层产能预测的准确率.以陕甘宁盆地中部气田马五1气藏为例,选用19口井的92个已测试层位作为已知样本(其中78个训练样本,14个检验样本),以气层、含气层、干层、水层和产能赋值为目标,挑选与储层特征密切相关的10个特征参数作为输入变量,建立中部气田马五1气藏的储层判识模型和气层产能预测模型.检验结果表明:模型的预测误差较传统的建模方法和多项式自组织神经网络方法(MOSN)低,其中尤以主成分分析逐类组合支持向量机模型(PCA-TCS-VM)的预测误差最低(平均绝对误差为0.359,平均相对误差为0.036).表明逐类组合支持向量机方法减少数据类别对储层判识和产能预测的干扰,提高准确率,对油气勘探具有积极指导意义.It is difficult to identify reservoir and predict deliverability accurately in the tight reservoir. This paper proposes a new modeling method —termwise-combination support vector machine (TCSVM). Firstly, we apply support vector classification (SVC) to achieve reservoir identification, and then use the support vector regression (SVR) to establish deliverability prediction model by category. Finally, reservoir identification and deliverability prediction are implemented in the TCSVM model. After the noise reduction, dimension reduction in the earlier days, this model can reduce interference from the categories of samples, thus improve greatly the accuracy of reservoir identification and deliverability prediction. The model is applied to identify reservoir and predict deliverability in M51 reservoir, which belongs to the central gas-field of Ordos Basin. Ninety two samples were acquired by zonal testing from the selected nineteen wells to identify the gas horizon, gas-bearing horizon, dry zone, water layer and deliverability prediction. The paper chooses seventy eight samples randomly for training, and the remaining 14 to be the testing. To establish the model of M51 reservoir in the central gas-field, we select ten parameters which related closely with the reservoir characteristics as input variables. Then we use the model to achieve reservoir identification and deliverability prediction. The results show that the model prediction error is lower than the traditional modeling method and polynomial self-organizing neural networks (MOSN). Especially, the principal component analysis and termwise-combination support vector machine model (PCA-TCSVM) prediction error is the lowest (average absolute error 0.359, average relative error 0.036). Therefore, PCA-TCSVM model can generally reduce interference from the categories of samples and improve the accuracy. This will have a positive significance for oil and gas exploration.

关 键 词：逐类组合支持向量机 气层判识 气层产能预测 陕甘宁盆地马五1气藏 

分 类 号：TE122.2[石油与天然气工程—油气勘探]