基于机器学习的油藏生产监测数据分析与预警方法  

Reservoir production monitoring data analysis and early warning method based on machine learning

作　　者：李建荣[1] 王胜[1] LI Jianrong;WANG Sheng(Shengli Oilfield,Sinopec,Dongying,Shandong 257001,China)

机构地区：[1]中国石油化工股份有限公司胜利油田分公司,山东东营257001

出　　处：《油气井测试》2023年第5期8-13,共6页Well Testing

基　　金：中石化重大先导课题“水平井动态监测技术研究与应用”(P22024);胜利油田分公司油藏动态监测中心科研“基于监测资料的区块开发动态分析与预测技术”(ZX2022-10)。

摘　　要：对于渗流机理不明确或地质模型不确定性高的非常规油气藏,基于油藏工程和数值模拟法的油藏动态分析误差较大。利用生产监测数据,分析油藏动态,实现数据驱动的监测井网优化和生产动态预测预警。采用K均值法建立生产井聚类算法,实现生产监测数据驱动的监测井网优化;构建卷积神经网络拟合生产监测指标间非线性关系,量化指标的不确定性区间,通过计算小概率事件进行预警。经实际算例验证表明:监测井网优化算法能够有针对性地在特征空间实现聚类,得到最具代表性的井位构成监测井网;指标预测预警方法能够稳健地预测指标不确定性区间,作为预警的依据。该方法为监测井网优化和生产动态预警提供了一种不依赖渗流机理的数据驱动技术,对智能油气田建设有积极作用。For unconventional oil and gas reservoirs with unclear flow mechanisms or high geological model uncertainties,traditional reservoir engineering and numerical simulation methods can introduce significant errors in dynamic reservoir analysis.To address this issue,by using production monitoring data,the reservoir dynamics was analyzed to realize data-driven monitoring well network optimization,as well as dynamic production forecasting and early warning.By using a K-means method,a clustering algorithm for production wells was established,which achieved monitoring well network optimization driven by production monitoring data.Additionally,a convolutional neural network was constructed to model the nonlinear relationships between production monitoring indicators,as well as to quantify the uncertainty intervals of these indicators,and then,early warning was realized by calculating small probability events.Verified by actual cases,it is shown that the monitoring well network optimization algorithm can achieve targeted clustering in the feature space and obtain the most representative well locations to form a monitoring well network;the indicator prediction and early warning method can robustly predict the indicator uncertainty interval and serve as the basis for early warning.This approach provides a data-driven technique that does not rely on flow mechanisms and has a positive impact on the establishment of intelligent oil and gas fields,particularly for reservoir monitoring network optimization and dynamic production forecasting and early warning.

关 键 词：动态监测 数据驱动 井网优化 K均值聚类 卷积神经网络 指标预测 动态预警 机器学习 

分 类 号：TE353[石油与天然气工程—油气田开发工程]