首页|基于概率建模的分层产液劈分方法

基于概率建模的分层产液劈分方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
  • 维普
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.
Prediction method of fluid production profiles based on a probabilistic modeling method
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 sam-pling 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 da-ta.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 produc-tion 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.

multi-layers productionprediction of production profileBayesian neural networkextreme gradient boosting al-gorithmsmall sampling number

辛国靖、张凯、田丰、姚剑、姚传进、王中正、张黎明、姚军

展开 >

中国石油大学(华东)石油工程学院,山东青岛 266580

中国石化胜利油田分公司信息化管理中心,山东东营 257000

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

国家自然科学基金国家自然科学基金国家自然科学基金中国石油科技重大专项中国海油重大科技项目山东省高等学校青创科技支持计划高等学校学科创新引智计划(111计划)

522740575207434051874335ZD2019-183-008CCL2022RCPS0397RSN2019KJH002B08028

2024

10.3969/j.issn.1673-5005.2024.02.012

中国石油大学学报(自然科学版)
中国石油大学

中国石油大学学报(自然科学版)

CSTPCD北大核心
影响因子：1.169
ISSN：1673-5005
年,卷(期)：2024.48(2)
  • 28