基于深度学习的稠油蒸汽驱汽窜时间预测方法  

Deep-learning-based prediction method of steam channeling time in heavy oil reservoirs

作　　者：崔传智[1] 陆水青山 吴忠维 盖平原 刘廷峰 CUI Chuanzhi;LU Shuiqingshan;WU Zhongwei;GAI Pingyuan;LIU Tingfeng(MOE Key Laboratory of Unconventional Oil&Gas Development,China University of Petroleum(East China),Qingdao 266580,Shandong Province,P.R.China;Petroleum Engineering Technology Research Institute,Shengli Oilfield Company,SINOPEC,Dongying 257099,Shandong Province,P.R.China)

机构地区：[1]中国石油大学(华东)非常规油气开发教育部重点实验室,山东青岛266580 [2]中石化胜利油田分公司石油工程技术研究院,山东东营257099

出　　处：《深圳大学学报（理工版）》2023年第5期622-630,共9页Journal of Shenzhen University(Science and Engineering)

基　　金：国家自然科学基金资助项目(51974343);青岛市博士后应用研究资助项目(qdyy20200084);中国石油大学(华东)自主创新科研计划资助项目(理工科)(20CX06089A);中国石油大学(华东)石油工程学院2021年研究生创新工程资助项目(YCX2021047)。

摘　　要：实时可靠的汽窜时间预测方法可以为及时采取防治措施提供支持.结合深度学习算法以及油井自身动态数据的约束,提出一种预测稠油井汽窜时间的新方法.针对单一指标波动频繁,噪声大,以及无法准确表征汽窜时间等问题,根据油田实际注采数据,以参数组合的方式构建表征蒸汽窜流通道形成时间的指标参数,并结合变异系数-G1混合交叉赋权法融合成汽窜综合判识曲线.基于标准互信息的相似性度量方法选择合适的时间序列数据作为输入特征,以相应的汽窜综合判识曲线作为输出时间序列构建学习样本.采用序列到序列深度学习框架建立汽窜时间的预测模型进行实际预测,并与传统的机器学习方法进行对比,验证模型的有效性和优越性.该方法通过数据驱动的方式模拟了注采时间序列特征与汽窜判识曲线之间的映射关系,可有效提高汽窜时间预测的效率和精度,对汽窜智能预警具有一定指导意义.A real-time and reliable prediction method for the onset of a steam breakthrough in heavy oil wells is proposed by integrating deep learning algorithms with dynamic data from the wells.To address the issues such as frequent fluctuations and high noise level in single indicator data,and inability to accurately characterize the steam breakthrough time,a set of indicator parameters is constructed based on the actual injection and production data of the oilfield,effectively representing the formation time of the steam channel.These parameters are combined with the variation coefficient-G1 hybrid cross-weighting method to fuse into a comprehensive breakthrough identification curve of steam channeling.On this basis,suitable time series data are selected as input features using the similarity measurement method based on standard mutual information,with the corresponding breakthrough identification curve as the output time series for constructing the learning samples.A sequence-to-sequence deep learning framework is used to establish a prediction model for steam breakthrough time,and the effectiveness and superiority of the model are verified by the actual predictions and comparison with the traditional machine learning methods.This method simulates the mapping relationship between the characteristics of injection production time series and the identification curve of steam channeling in a data-driven way,which can effectively improve the efficiency and accuracy of steam channeling time prediction,and has certain guiding significance for the intelligent early warning of steam channeling.

关 键 词：油藏工程 稠油油藏 汽窜判识 汽窜时间预测 自然语言处理 序列到序列深度学习框架 

分 类 号：TE34[石油与天然气工程—油气田开发工程]