摘要
致密砂岩储层测井含水饱和度预测是油气藏储层评价和产量预测的关键步骤,应用机器学习模型预测含水饱和度在一定程度上缓解了常规方法预测误差大的问题。但是现有的机器学习方法通常使用有限的测井数据从头开始训练模型,导致模型能力受限,进而阻碍了它的泛化能力。为此,基于大语言模型(LLMs)出色的泛化性能及丰富的知识信息,引入LLMs进行储层测井含水饱和度预测,提出了一种基于真实关系及表格Transformer网络(REaLTabFormer)增强的LLMs对齐框架模型(RTF-LLA),最后进行了实验对比验证。研究结果表明:(1) RTF-LLA模型由数据增强、知识蒸馏和跨模态对齐3个核心模块构成;(2)数据增强模块以原始测井数据为基础,利用REaLTabFormer捕获测井参数与储层物性参数间的内在关系,生成了高信息量的测井数据;(3)知识蒸馏模块从LLMs提取主要的知识信息,指导测井数据与LLMs文本知识进行跨模态对齐,并赋予模型准确预测储层测井含水饱和度的能力;(4)跨模态对齐模块通过词元对齐、特征对齐和序列对齐,有效地降低了模型对储层含水饱和度的预测误差。结论认为:(1)RTF-LLA模型在S气田储层饱和度实验评价中的平均绝对误差和均方根误差分别为1.332和2.207,相较于其他主流机器学习算法至少降低了3.310和3.174;(2)RTF-LLA模型可为小样本测井资料储层含水饱和度准确预测提供有效技术支撑,为储层测井含水饱和度预测提供了新思路、新方法。
Prediction of the water saturation in well logging is a key step in the reservoir evaluation and production prediction of tight sand oil and gas reservoirs.The application of machine learning models to predict water saturation has,to some extent,alleviated the problem of large prediction errors in conventional methods.However,existing machine learning methods usually begin model training with limited logging data,which restricts the capacity of the model,hindering its generalization ability.Based on the excellent generalization performance and rich knowledge information of Large Language Models(LLMs),this paper introduces LLMs to predict reservoir water saturation.Then,a realistic relational and tabular transformer network(REaLTabFormer)enhanced LLMs alignment framework model(RTF-LLA)is established and experimentally compared and verified.And the following research results are obtained.First,the RTF-LLA model consists of three core modules,i.e.,data enhancement,knowledge distillation and cross-modal alignment.Second,the data enhancement module captures the intrinsic relationships between well logging parameters and reservoir physical parameters to generate large-information well logging data by using the REaLTabFormer,based on original well logging data.Third,the knowledge distillation module extracts the main knowledge information from LLMs to guide the cross-modal alignment of well logging data and LLMs text knowledge and endow the model with the ability to predict reservoir water saturation accurately.Fourth,the cross-modal alignment module effectively reduces the prediction error of reservoir water saturation through lexical unit alignment,feature alignment and sequence alignment.In conclusion,when the RTF-LLA model is used for the experimental evaluation of reservoir saturation in S Gas Field,its mean absolute error(MAE)and root mean square error(RMSE)are 1.332 and 2.207,respectively,which are at least 3.310 and 3.174 lower than those by other mainstream machine learning algorithms.What's more,the RTF-LLA mo
作者
武娟
罗仁泽
雷璨如
殷疆
陈星廷
WU Juan;LUO Renze;LEI Canru;YIN Jiang;CHEN Xingting(School of Geoscience and Technology,Southwest Petroleum University,Chengdu,Sichuan 610500,China;State Key Laboratory of Oil&Gas Reservoir Geology and Exploitation//Southwest Petroleum University,Chengdu,Sichuan 610500,China;School of Electrical Engineering and Information,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Department of Geology,Northwest University,Xi'an,Shaanxi 710069,China;State Key Laboratory of Continental Dynamics//Northwest University,Xi'an,Shaanxi 710069,China)
出处
《天然气工业》
EI
CAS
CSCD
北大核心
2024年第9期77-87,共11页
Natural Gas Industry
基金
四川省科技厅项目“新疆油田玛湖侏罗系三工河组致密薄储层地震预测方法研究”(编号:2024YFHZ0158)。
关键词
大语言模型
跨模态对齐
致密砂岩储层
测井含水饱和度预测
泛化能力
Large Language Models(LLMs)
Cross-modal alignment
Tight sandstone reservoir
Water saturation prediction in well logging
Generalization
