摘要
针对套管损坏影响因素多、数据复杂、跨越时间范围和空间范围大、存在非线性、不确定性和时变性等特点,基于数据驱动的理念,采用机器学习技术开展相关研究。首先围绕注水压力、注水量、注水强度、压差等套损影响因素采集油水井历史生产数据,建立单井套损评价指标;其次通过数据预处理、特征参数计算、相关性分析等技术构建单井套损样本集;然后对特征参数的重要性进行评估,形成4套特征组合方案;最后分别采用随机森林和支持向量机算法建立单井套损预测模型,并给出不同特征组合下模型性能参数。试验结果表明:高压注水是影响研究断块注水井套损的主要因素,采用MDA特征组合方案建立的支持向量机模型能够较好地预测套损,准确率达到93.3%。
Casing damage can be the result of a number of factors during a long process of oilfield production, which involves complicated and interaction of various factors. Regarding to this characteristics, a data-driven machine learning prediction approach of casing damage was proposed. Firstly, the evaluation indexes of casing damage were designed from the massive historical field data related to multiple factors, including water injection pressure, injection volume and injection and production pressure difference. Then, samples of individual well casing damage were collected after the data preprocessing, feature extraction and correlation analysis. Finally, the algorithms of random forest and support vector machine were applied to build the predictions models of casing damage based on different feature combinations. The results of case studies show that high-pressure water injection is the major factor causing casing damage, and the prediction accuracy of the model is 93.3% using the MDA feature combination and the support vector machine algorithm.
作者
赵艳红
姜汉桥
李洪奇
刘洪涛
韩大伟
王英男
刘灿超
ZHAO Yanhong;JIANG Hanqiao;LI Hongqi;LIU Hongtao;HAN Dawei;WANG Yingnan;LIU Canchao(College of Petroleum Engineering in China University of Petroleum(Beijing),Beijing 102249,China;Petroleum Engineering and Data Mining Laboratory,China University of Petroleum(Beijing),Beijing 102249,China;College of Artificial Intelligence in China University of Petroleum(Beijing),Beijing 102249,China;The 7th Oil Production Plant of Daqing Oilfield,Daqing 163517,China)
出处
《中国石油大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2020年第4期57-67,共11页
Journal of China University of Petroleum(Edition of Natural Science)
基金
国家自然科学基金项目(41130417)。
关键词
套管损坏
套损预测
数据驱动
机器学习
时间序列
casing damage
casing damage prediction
data driven
machine learning
time series
