基于CNN和LSTM的机器学习模型在测井岩性识别的应用  

Application of Machine Learning Model Based on CNN and LSTM in Well Logging Lithology Identification

作　　者：张凤博 马雪玲 董珍珍 邹路 王茜 李伟荣 吴磊 ZHANG Fengbo;MA Xueling;DONG Zhenzhen;ZOU Lu;WANG Xi;LI Weirong;WU Lei(Petroleum Exploration and Development Research Institute of Xiasiwan Oil Production Plant,Yanchang Oilfield Co.,Ltd.,Yan’an,Shaanxi 716000,China;College of Petroleum Engineering,Xi’an Shiyou University,Xi’an,Shaanxi 710065,China)

机构地区：[1]延长油田下寺湾采油厂勘探开发研究所,陕西延安716000 [2]西安石油大学石油工程学院,陕西西安710065

出　　处：《西安石油大学学报（自然科学版）》2024年第5期96-103,133,共9页Journal of Xi’an Shiyou University（Natural Science Edition）

基　　金：中国石油科技创新基金项目(2022DQ02-0201)。

摘　　要：在油气田勘探和开发中,测井解释是表征储层物性参数和评价油气储量的重要手段之一。其中,岩性识别是测井解释的主要任务。针对用于储层岩性识别的机器学习方法普遍存在参数优化难、训练时间长、容易过拟合等问题,导致测井解释精度低及岩性相近易混淆等。本文将卷积神经网络(CNN)有利于特征提取的优点与长短期记忆神经网络(LSTM)可考虑测井曲线随深度变化的趋势性信息的优点相结合,提出CNN-LSTM混合神经网络构建测井数据与岩性类别之间的非线性模型,并采用遗传算法(GA)优化混合神经网络模型的超参数,提高识别效率。基于4069组样本数据评估了该混合模型的性能。研究结果表明,与传统的机器学习方法相比,CNN-LSTM-GA混合神经网络优化模型有效地克服了储层岩性识别研究中的问题,取得更好的岩性识别效果,对油藏精细描述和储量评价具有一定的实用价值。In the exploration and development of oil and gas fields,well logging interpretation is one of the important means to characterize reservoir physical parameters and evaluate oil and gas reserves.Lithology identification is the main task of well logging interpretation.It is of great significance to use machine learning methods for well logging data interpretation and reservoir lithology identification.The current machine learning methods used for reservoir lithology identification generally have problems such as difficult parameter optimization,long training time,and easy overfitting,resulting in low accuracy of logging interpretation and easy confusion of similar lithologies.In this study,by combining the advantage of convolutional neural network(CNN)for feature extraction with the advantage of long short-term memory neural network(LSTM)for considering the varying trend information of well logging curves with depth,it is proposed to construct a nonlinear model between well logging data and lithology categories using a CNN-LSTM hybrid neural network,and the hyperparameters of the hybrid neural network model are optimized using genetic algorithm(GA)to improve lithology identification efficiency.The performance of the hybrid model was evaluated based on 4069 sets of sample data.The results show that compared with traditional machine learning methods,the CNN-LSTM-GA hybrid neural network optimization model effectively solves the problems in reservoir lithology identification research,achieves better lithology identification results,and has certain practical value for fine reservoir description and reserve evaluation.

关 键 词：岩性识别 卷积神经网络 长短期记忆网络 遗传算法 混合神经网络模型 

分 类 号：TE928[石油与天然气工程—石油机械设备]