基于稀疏诱导特征选择神经网络的储层预测方法研究

STUDY ON THE RESERVOIR PREDICTION METHOD BASED ON SPARSE INDUCED FEATURE SELECTION NEURAL NETWORK

常规储层预测方法对地震属性之间的隐含关系挖掘不充分、地震属性种类繁多难以选择。针对以上问题,为提高储层岩性的分类预测精度,提出一种结合特征选择与神经网络的储层预测方法。以DenseNet与SENet为基础,使用正则惩罚项进行网络输入层稀疏化,得到每个输入节点权重,进一步使用ReLU激活函数构建特征选择层,实现地震属性的筛选。实验使用胜利油田的地震属性数据与钻井岩性数据,结果表明,该方法对储层岩性分类效果具有明显的提升,分类准确率超过70%,验证了神经网络模型与该特征选择方法的有效性。

Conventional reservoir prediction methods is not sufficient to explore the implicit relationships between seismic attributes,and the type of seismic attributes is so numerous that is difficult to choose.Thus,in order to improve the classification prediction accuracy of reservoir lithology,a reservoir prediction method combining feature selection and neural network is proposed.Based on DenseNet and SENet,the network input layer was thinned using a regular penalty term to get the weight of each input node.The feature selection layer was further constructed using the ReLU activation function to achieve the selection of seismic attributes.Using seismic attribute data and lithology data of Shengli Oilfield,the experimental results show that this method significantly improves the reservoir lithology classification effect,and the classification accuracy rate is over 70%,which proves the effectiveness of the neural network model and the feature selection method.