论文详情
针对现有核磁共振测井渗透率模型对孔隙结构复杂的致密砂岩储层预测精度不高的问题,在分析误差反向传播神经网络的缺陷后,提出了一种利用集成正则化改进神经网络(BPNN)算法与核磁共振T2全谱预测致密砂岩储层渗透率的方法。该方法采用自构形算法自动确定隐层神经元的个数,采用自适应雨林优化算法避免BP神经网络迭代陷入局部极小值,利用L2正则化算子保证算法的稳定性,采用Adaboost集成算法串联若干BP神经网络以提高模型泛化能力。提取某区致密砂岩储层192块岩样的核磁共振T2全谱数据进行建模,并应用于非建模井的渗透率评价,认为基于集成正则化改进BPNN算法评价储层渗透率精度较高,均方误差仅有0.286。
In view of the difficulty of the existing NMR logging permeability model in the prediction of a tight sandstone reservoir with complex pore structures,after analyzing the defects of the back-propagation neural network (BPNN),we propose a method for predicting the permeability of tight sandstone reservoirs by using the
BPNN algorithm improved by integrated regularization and discrete NMR T2 full-spectrum data.For this method,a self-shaping algorithm is employed to automatically determine the number of hidden neurons,an adaptive rainforest optimization algorithm is used to avoid the BPNN iterations into local minimum values,the L2 regularization technique is used to guarantee the algorithm stability,and the AdaBoost algorithm is used to concatenate several BPNNs to improve the generalization ability of the model.Discrete NMR T2 full-spectrum data were extracted from 192 samples of tight sandstone reservoirs in a certain area for model building,which was used to predict the permeability of non-modeling wells.The results showed that it was more accurate to use the BPNN algorithm improved by integrated regularization in evaluating reservoir permeability,with a mean squared error of 0.286 only.
湖北省自然科学基金项目(2013CFB396)与中国石油天然气集团公司重大专项(2013E-38-09)共同资助。