摘要
传统地震储层预测技术已无法满足储层精细评价的需求,深度学习具有强大的特征提取和高维数据处理能力,近年来广泛应用于地震储层预测并取得了较好的效果。为此,本文深入讨论深度学习技术在地震储层预测中的应用、进展及它在实际工作中面临的挑战,并提出未来的发展方向。主要认识有:(1)在烃类定性检测方面,深度学习技术有助于综合利用多属性地震数据去提高效率和预测结果的准确率;在定量预测方面,深度学习技术可以更精准地逼近地震数据与目标之间复杂的非线性关系,实现储层的精细定量评价。(2)深度学习技术的应用面临的挑战主要是标签数据不足和样本不均衡等容易导致模型过拟合,泛化能力差;模型复杂,计算成本高;模型的“黑匣子”特征使预测结果缺乏物理可解释性;缺乏定性预测模型的评价标准和高精度的不确定性量化算法。(3)未来的研究方向应致力于克服数据可用性的不足和深度学习的局限性等,构建地球物理知识图谱,实现多源数据与知识的有效融合、共享,将深度学习与反馈强化学习等其他机器学习算法相结合,为油气勘探和开发提供更可靠的技术支撑。
Traditional seismic data‐based reservoir prediction technology fails to meet the demands of refined reservoir evaluation.Deep learning has strong feature extraction and high‐dimensional data processing capabili‐ties and has been extensively applied in seismic data‐based reservoir prediction with promising results in recent years.This paper delved into the application and progress of deep learning technology in seismic data‐based reservoir prediction,analyzed the challenges encountered during practical implementation,and proposed future research directions.The conclusions are as follows:①In terms of qualitative hydrocarbon detection,deep learning technology facilitates the comprehensive utilization of multi‐attribute seismic data to improve the effi‐ciency and accuracy of prediction results.In terms of quantitative prediction,it enables a more precise approxi‐mation of the intricate nonlinear relationship between seismic data and targets,thereby achieving a refined quan‐titative evaluation of reservoirs.②The application of deep learning technology faces several challenges.The is‐sues such as insufficient label data and unbalanced samples lead to overfitting and poor generalization ability of the model;the complex model results in high computational costs;the“black box”feature of the model makes the prediction results lack physical interpretability;there is no evaluation criteria for qualitative prediction model and high‐precision quantization algorithm for uncertainty.③Future research should prioritize addressing challenges related to insufficient data availability and limitations of deep learning,such as constructing geophysi‐cal knowledge maps,effectively integrating and sharing multi‐source data and knowledge,and combining deep learning with other machine learning algorithms such as feedback reinforcement learning,so as to provide more reliable technical support for hydrocarbon exploration and development.
作者
骆迪
王宏斌
蔡峰
吴志强
孙运宝
李清
LUO Di;WANG Hongbin;CAI Feng;WU Zhiqiang;SUN Yunbao;LI Qing(Key Laboratory of Gas Hydrate,Ministry of Natural Resources,Qingdao Institute of Marine Geology,China Geological Survey,Qingdao,Shandong 266237,China;Laboratory for Marine Mineral Resources,Laoshan Laboratory,Qingdao,Shandong 266237,China)
出处
《石油地球物理勘探》
EI
CSCD
北大核心
2024年第3期640-651,共12页
Oil Geophysical Prospecting
基金
崂山实验室科技创新项目“西太典型边缘海盆地水合物运聚成藏过程研究”(LSKJ202203501)
中国地质调查项目“渤海等海域新生界油气地质条件与碳封存选区”(DD20230401)联合资助。
关键词
地震储层预测
深度学习
地震反演
地震烃类检测
有监督学习
无监督学习
seismic data‐based reservoir prediction
deep learning
seismic inversion
seismic hydrocarbon detec‐tion
supervised learning
unsupervised learning
作者简介
王宏斌,山东省青岛市即墨区观山路596号青岛海洋地质研究所,266237。Email:luodi0927@sina.com;骆迪,副研究员,1982年生,2006年获山东科技大学计算机科学与技术专业学士学位,2009年获西华大学计算机应用技术专业硕士学位,2013年获中国石油大学(华东)地质资源与地质工程专业博士学位,目前就职于中国地质调查局青岛海洋地质研究所,主要从事海洋综合地球物理及地震储层预测工作。
