基于机器学习算法预测核磁共振T_(2)谱  

NMR T_(2)Spectrum Prediction Method Based on Machine Learning Algorithm

作　　者：张哲[1,2] 廖广志[1,2] 肖立志[1,2] 崔云江 王培春 李志愿 ZHANG Zhe;LIAO Guang-zhi;XIAO Li-zhi;CUI Yun-jiang;WANG Pei-chun;LI Zhi-yuan(College of Geophysics,China University of Petroleum,Beijing 102249,China;State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum,Beijing 102249,China;Exploration and Development Research Institute of CNOOC Tianjin Branch,Tianjin 300450,China)

机构地区：[1]中国石油大学(北京)地球物理学院,北京102249 [2]中国石油大学(北京)油气与勘探国家重点实验室,北京102249 [3]中海石油(中国)有限公司天津分公司勘探开发研究院,天津300450

出　　处：《科学技术与工程》2023年第17期7282-7292,共11页Science Technology and Engineering

基　　金：国家重点研发计划(2019YFA0708301);国家自然科学基金(51974337);中国石油天然气集团有限公司-中国石油大学(北京)战略合作科技专项(ZLZX2020-03);中国石油科技创新基金(2021DQ02-0403);中国石油勘探开发研究院开放基金(2022-KFKT-09)。

摘　　要：近年来,深度学习算法被广泛应用于生成各种类型的数据。通过分析测井数据与核磁共振T_(2)谱之间的映射关系,利用随机森林与长短期记忆(long short-term memory,LSTM)神经网络模型实现了对核磁共振T_(2)谱的重构。核磁共振测井在每个深度获得的T_(2)谱是在不同的时间序列中通过不同的布点数来显示形态上的变化的,随机森林算法能够处理高维度的核磁共振T_(2)谱数据且不需要做特征选择,而LSTM可以很好地控制不同深度神经元对T_(2)谱各分布点的影响,将这两种网络模型进行参数优化后,对同一口井的预测结果进行对比分析。选取了中国海上A油田的测井数据作为例子进行方法测试。首先,利用灰色关联度算法分析T_(2)几何均值与测井曲线的相关性。选取相关性高于设定值的测井曲线,将测井曲线标准化缩放在0~1后作为随机森林与LSTM模型的输入,预测同一地层T_(2)谱形态分布规律并比较算法的优劣。在比较软件处理得到的核磁共振T_(2)谱和预测结果后,分析它们之间产生差异的原因。结果显示通过LSTM神经网络模型预测的数据与地层真实数据的符合度比随机森林算法更高,符合度可达90%以上。In recent years,deep learning algorithms have been widely used to generate various types of data.The mapping relationship between logging data and NMR T_(2)spectrum was analyzed,then random forest and long short-term memory(LSTM)network model were used to realize the reconstruction of NMR T_(2)spectrum.T_(2)spectrum obtained at each depth of NMR logging shows morphological changes through different number of points in different time series.Random forest algorithm can process high-dimensional NMR T_(2)spectrum data without feature selection,while LSTM can well control the influence of neurons at different depths on T_(2)spectrum distribution points.After optimizing the parameters of the two network models,the prediction results of the same well were compared and analyzed.Well logging data of A Oilfield in CNOOC were selected as examples for method testing.Firstly,the correlation between T_(2)geometric mean and logging curve was analyzed by using gray correlation degree algorithm.The logging curve whose correlation is higher than the set value was selected,and the normalized scaling of the logging curve at 0~1 was used as the input of the random forest and LSTM models to predict the distribution law of T_(2)spectrum morphology in the same stratum and compare the advantages and disadvantages of the algorithm.After comparing the NMR T_(2)spectrum processed by the software with the predicted results,the reasons for the differences between them were analyzed.The results show that the data predicted by LSTM neural network model is more consistent with the real stratum data than the random forest algorithm,and the consistency can reach more than 90%.

关 键 词：机器学习 随机森林 LSTM 核磁共振T_(2)谱 

分 类 号：TE19[石油与天然气工程—油气勘探] P618.1[天文地球—矿床学]