基于卷积神经网络和迁移学习的特高含水油井生产预测  

Production prediction of extra high water cut oil well based on convolution neural network and transfer learning

作　　者：姜春雷[1,2] 方硕 刘伟[1,2] 邵克勇[1] 陈朋[1] JIANG Chunei;FANG Shuo;LIU Wei;SHAO Keyong;CHEN Peng(School of Electrical Information Engineering,Northeast Petroleum University,Daqing 163318,China;Sanya Offshore Oil&Gas Research Institute,Northeast Petroleum University,Sanya 572024,China)

机构地区：[1]东北石油大学电气信息工程学院,黑龙江大庆163318 [2]东北石油大学三亚海洋油气研究院,海南三亚572024

出　　处：《中国石油大学学报（自然科学版）》2023年第6期162-170,共9页Journal of China University of Petroleum(Edition of Natural Science)

基　　金：黑龙江省自然科学基金项目(LH2021F008);海南省重点研发项目(ZDYF2022SHFZ047);控制科学与工程团队专项(2022TSTD-04)。

摘　　要：油井的实时生产监测对油田的辅助生产和精细化管理有重要意义。然而,针对仅有小样本生产数据、数据波动大且有缺失的特高含水期油井,传统的机器学习算法无法实现良好的生产预测。提出一种基于卷积神经网络和迁移学习的多任务生产预测方法。该方法不仅可以实现时间和空间上特征的自适应提取,还可以改善模型在小样本数据上的预测性能。结果表明:相比于基准模型,产液量和动液面的平均绝对误差分别降低31.26%和60.81%,决定系数分别提高1.89%和7.59%。基于迁移学习的MTCNN模型提高小样本数据油井的生产预测精度,实现了特高含水油井产液量和动液面的实时预测,对抽油机系统的效率优化、油井边缘设备智能化有参考意义。The real-time production monitoring of oil wells is of great significance for enhancing auxiliary production and fine management in oil fields.However,the traditional machine learning algorithms struggle to provide accurate production predic-tions for ultra-high water cut oil fields due to limited sample production data,substantial data fluctuations,and missing data.This paper proposes a multi-task production forecasting scheme based on convolutional neural networks and transfer learning to address these challenges.This model not only enables adaptive extraction of temporal and spatial features,but also enhanc-esprediction performance on small sample data.The experimental results demonstrate notable improvements over the bench-mark model.Specifically,the average absolute percentage errors of liquid production and dynamic liquid level are reduced by 31.26%and 60.81%respectively.Additionally,and the determination coefficient increases by 1.89%and 7.59%respec-tively.The MTCNN model,based on transfer learning,enhances the prediction accuracy of oil wells with limitedsample data,enabling real-time prediction of liquid production and dynamic liquid level inultra-high water cut oil wells.It holds significant implications for the efficiency optimization of pumping unit systems and the intelligence of oil well edge equipment.

关 键 词：卷积神经网络 迁移学习 特高含水油井 小样本数据 多任务 动态生产预测 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]