摘要
深度学习波阻抗预测已成为油气勘探领域的重要课题,因测井数据有限,导致可用于神经网络训练的标签数量明显不足.为此,本文采用理论岩石物理模型道集正演的方法来获得训练标签,该方法使得利用单口井数据进行深度学习储层参数反演成为可能.通过理论岩石物理模型流体替换和波动方程正演,获得与实际工区岩石物理参数一致的变孔隙度和饱和度的训练标签.为了提高模型的泛化度和精确性,实验使用逐步回归法在20多种地震属性中优选出9种最优的地震属性组合.将这9种地震属性组合和对应的测井波阻抗数据带入到深度神经网络(DNN)中进行训练,准确地预测出波阻抗,预测波阻抗与实际波阻抗的互相关系数达到了0.992,平均误差为304(m·s^(-1))·(g·cm^(-3)).在三维实际地震工区的应用中,使用该方法预测了研究工区的纵、横波阻抗和密度结果,通过建立地层模型,反演地层模型参数,反演结果显示,深度学习反演方法的预测精度比传统反演方法预测精度高,因此更有利于对薄层的识别.
Deep learning-based impedance prediction has become a significant research topic in oil and gas exploration.Due to the limited availability of well log data,the number of labels suitable for neural network training is notably insufficient.To address this,this study employs a method of forward modeling synthetic gathers using theoretical rock physics models to generate training labels,making it possible to perform deep learning-based reservoir parameter inversion with data from a single well.Through theoretical rock physics model fluid substitution and wave equation forward modeling,training labels with varying porosity and saturation consistent with the actual field's rock physics parameters are obtained.To enhance the model's generalization capability and accuracy,the experiment employs stepwise regression to select an optimal combination of 9 seismic attributes from over 20 candidates.These 9 seismic attributes,along with the corresponding well log impedance data,are fed into a Deep Neural Network(DNN)for training.The trained network accurately predicts impedance,achieving a cross-correlation coefficient of 0.992 between predicted and actual impedance,with an average error of 304(m·s^(-1))·(g·cm^(-3)).In the application to a 3D real seismic survey area,this method is used to predict P-impedance,S-impedance,and density.By constructing a formation model and inverting its parameters,the results demonstrate that the deep learning inversion method achieves slightly higher prediction accuracy than traditional inversion methods,potentially offering improved identification of thin layers.
作者
赵昭阳
赵建国
欧阳芳
肖增佳
马铭
闫博鸿
ZHAO ZhaoYang;ZHAO JianGuo;OUYANG Fang;XIAO ZengJia;MA Ming;YAN BoHong(State Key Laboratory of Petroleum Resources and Engineering,China University of Petroleum(Beijing),Beijing 102249,China;Institute of Earthquake Forecasting,China Earthquake Administration,Beijing 100036,China)
出处
《地球物理学报》
北大核心
2025年第9期3554-3574,共21页
Chinese Journal of Geophysics
基金
国家自然科学基金项目(42304141,41974120)
国家自然科学基金联合基金重点项目(U20B2015)联合资助.
