摘要
针对大数据下抽油机井采油生产系统特点以及目前抽油机井工况识别研究中存在的问题,为进一步提高抽油机井工况识别精准率和工程实用性,提出一种基于Hessian正则化多视角学习的抽油机井工况识别新方法。首先通过先验知识和专家经验,选择实测地面示功图、电功率和井口温度信号3个视角并进行特征提取,然后利用log损失函数建立Hessian正则化多视角logistic回归工况识别模型,最后采用交替优化算法求取最优解并进行分类识别。应用该方法对胜利油田某区块11种抽油机井典型工况进行识别,其识别效果分别比传统的基于实测地面示功图、实测电功率及特征连接多源识别方法提高了2.4%、11%和13.8%,而在少量标记训练样本下该方法识别效果更优,从而验证了该方法的有效性。
To resolve the problems in working condition recognition of sucker-rod pumping wells and to further improve the accuracy and practicality,a novel method based on multi-view learning and Hessian regularization to identify the working condition was proposed. Firstly,the measured dynamometer cards,electrical power and wellhead temperature data were characterized based on the prior information and empirical knowledge. Then a multi-view logistic regression model with log loss function and Hessian regularization for working condition recognition was established. Finally,the working condition was classified and recognized by an alternating optimization algorithm. The proposed method was applied to eleven cases of typical working condition recognition in a block in Shengli Oilfield,and the results were compared with traditional recognition methods based on measured dynamometer cards,electrical power data and multi-sources of feature connection,respectively. The comparison shows that the recognition rates are improved by 2. 4%,11% and 13. 8%,respectively. The performance is even much better with a small amount of marked training samples.
作者
周斌
王延江
刘伟锋
刘宝弟
ZHOU Bin;WANG Yanjiang;LIU Weifeng;LIU Baodi(College of Information and Control Engineering in China University of Petroleum（East China）, Qingdao 266580, Chin)
出处
《中国石油大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2018年第3期154-161,共8页
Journal of China University of Petroleum(Edition of Natural Science)
基金
国家自然科学基金项目(61671480)
