基于概率建模的分层产液劈分方法  

Prediction method of fluid production profiles based on a probabilistic modeling method

作　　者：辛国靖 张凯[1] 田丰 姚剑 姚传进[1] 王中正 张黎明[1] 姚军[1] XIN Guojing;ZHANG Kai;TIAN Feng;YAO Jian;YAO Chuanjin;WANG Zhongzheng;ZHANG Liming;YAO Jun(School of Petroleum Engineering in China University of Petroleum(East China),Qingdao 266580,China;Information Management Center in SINOPEC Shengli Oilfield Company,Dongying 257000,China)

机构地区：[1]中国石油大学(华东)石油工程学院,山东青岛266580 [2]中国石化胜利油田分公司信息化管理中心,山东东营257000

出　　处：《中国石油大学学报（自然科学版）》2024年第2期109-117,共9页Journal of China University of Petroleum(Edition of Natural Science)

基　　金：国家自然科学基金项目(52274057,52074340,51874335);中石油重大科技项目(ZD2019-183-008);中海油重大科技项目(CCL2022RCPS0397RSN);山东省高等学校青创科技支持计划(2019KJH002);“111”计划(B08028)。

摘　　要：传统产液劈分方法无法考虑层间干扰及注水井和邻井的影响,难以准确判断井下实际状况。同时,海上油田产液剖面测试成本高,常规的机器学习方法面临样本数量少的问题。基于此,提出一种基于贝叶斯神经网络和极限梯度提升算法的多层合采产液劈分混合学习模型。概率方法可以识别预测中的不确定性,通过将神经网络与概率建模结合,进行分层产液数据分布特征挖掘,结合主控因素分析,混合学习算法可以实现小层产液量的准确预测,可以依据较少的数据获得更为稳健的模型。为验证所提方法的有效性,将其应用于实际油田某区块进行产液剖面预测。结果表明:相比KH劈分方法在计算中劈分系数固定,不会随着生产过程波动,所提出的方法可从历史数据中学习,预测精度达到87.9%,预测结果更加逼近真实单层产液量。The traditional fluid production splitting method cannot consider the influences of interzonal interference,injection wells and adjacent wells,so it is difficult to precisely assess the actual downhole conditions.Meanwhile,due to high cost of production profile testing in offshore oilfields,the conventional machine learning methods also face the problem of small sampling numbers,which has a great limitation for their application.In this study,a hybrid learning model was proposed with Bayesian neural network and extreme gradient boosting algorithm,which can formulate a more robust model based on less data.By combining the neural network with probabilistic modeling,mining the distribution characteristics of stratified liquid production data and analyzing the main control factors,the hybrid learning algorithm can accurately predict the liquid production in different layers.The new method was applied to prediction of the liquid production profiles in a real oilfield in order to verify its effectiveness.The results show that,compared with the KH splitting method,the splitting coefficient can be fixed in the calculation and does not fluctuate with the production process.The proposed method can learn from the historical data,with an accuracy of 87.9%,and the predicted results are closer to the real liquid production of each layer.

关 键 词：多层合采 产液剖面预测 贝叶斯神经网络 极限梯度提升算法 小样本 

分 类 号：TE327[石油与天然气工程—油气田开发工程]