基于机器学习融合模型的页岩气藏测井曲线构建方法  

Machine learning fusion model-based logging curve construction method for shale gas reservoir

作　　者：王鸣川[1] 岳明 WANG Mingchuan;YUE Ming(Sinopec Petroleum Exploitation&Production Research Institute,Beijing 102206,China;School of Civil and Resource Engineering,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]中国石化石油勘探开发研究院,北京102206 [2]北京科技大学土木与资源工程学院,北京100083

出　　处：《中国科技论文》2025年第3期255-266,共12页China Sciencepaper

基　　金：中国石化重点基础前瞻研究项目(P22205)。

摘　　要：针对页岩气藏多数开发井台测井曲线缺失给页岩气藏的精细表征与建模带来困难、单一机器学习模型构建测井曲线的精度无法满足需求且泛化性低的问题,提出了一套基于机器学习融合模型的页岩气藏测井曲线构建方法。在已知测井曲线预处理的基础上,建立训练数据集,输入到深度神经网络(deep neural net⁃work,DNN)、卷积神经网络(convolutional neural network,CNN)、长短期记忆神经网络(long short-term memory,LSTM)和随机森林(random forest,RF)4个个体学习器进行初步训练,从测井数据的序列特征、空间信息、细粒度特征等获取数据间的非线性映射关系。接着根据验证集数据进行模型参数调整,并获取各模型的预测精度。然后基于预测效果为各学习器分配权重,并对预测结果进行加权融合,从而形成精度高且泛化性强的测井曲线并构建融合模型。选取四川盆地X区块4口盲井进行应用效果验证,4口井构建的测井曲线,与原始的测井曲线相比,平均精度在90%以上。结果表明,新方法不仅能准确构建不同性质的测井曲线,而且泛化性强,能为页岩气藏的精细表征与建模提供较为可靠的测井曲线数据。The accuracy of constructing well logging curves and the low generalization ability by using a single machine learning model are insufficient to meet the requirements for the difficulty in finely characterizing and modeling shale gas reservoirs.This challenge arises due to the lack of well logging curves of most development wells in the platform.To address this issue,a method for shale gas reservoir well logging curve construction based on machine learning fusion models was proposed.Building upon preprocessed well logging curves,a training dataset was established and fed into four individual learners:deep neural networks(DNN),convolutional neural networks(CNN),long short-term memory neural networks(LSTM),and random forests(RF).These learners were preliminarily trained to capture the nonlinear mapping relationships among well logging data’s sequence features,spatial information,and fine-grained features.Subsequently,model parameters were adjusted using the validation dataset to evaluate each model’s prediction accuracy.Based on the prediction results,weight was assigned to each learner,and their predictions are fused through a weighted combination to generate a high-precision and strong generalization well logging curve prediction model.The effectiveness of the proposed approach was demonstrated by applying it to four blind wells in Block X of the Sichuan Basin.The constructed well logging curves for these four wells achieved an average accuracy of over 90%compared to that of the original well logging curves.The results indicate that the new method not only accurately constructs well logging curves of different properties but also exhibits strong generalization ability,providing reliable well logging curve data for the fine characterization and modeling of shale gas reservoirs.

关 键 词：页岩气藏 测井曲线 机器学习 融合模型 数据处理 构建方法 

分 类 号：TE122[石油与天然气工程—油气勘探]