摘要: |
目的 解决油田目的层钻井过程中完井液受盐水、残酸等污染后不能高效识别污染类型的问题。方法 对完井液进行不同质量占比的盐水、残酸污染测定,采用K-means聚类订正不同污染等级数据样本的标签。根据数据样本特征的获取难易度、隐藏层数目,训练不同的BP神经网络模型,并由留一交叉验证法检验模型的分类准确率。结果 数据样本拥有的特征越多,训练的BP神经网络分类准确率越高,隐层数目越多,分类准确率反而越低。选择包含“流变+老化+滤失+井名”4类特征的数据样本建立1隐藏层的BP神经网络模型,其平均分类准确率达到93.18%。结论 由流变、滤失等特征训练的BP神经网络模型可快速应用于试油现场,解决完井液污染类型识别问题,避免了试油现场因缺少大型仪器而无法鉴别完井液污染类型的难题。 |
关键词: 完井液 污染类型 计算机模拟 K-means聚类 神经网络 留一交叉验证 |
DOI:10.3969/j.issn.1007-3426.2023.06.018 |
分类号: |
基金项目:中国博士后科学基金“页岩气藏转向重复压裂可降解暂堵剂及其降解控制机理”(2018M631100);四川省重点研发项目“绿色环保热固性低密度压裂支撑剂研制及连续化生产关键技术”(2021YFG0112) |
|
A BP neural network-based identification method on the type of completion fluid contamination |
Cheng Xin1, Zhang Tailiang1, Yang Lanping2, Yang Qingzheng1, Bai Yi1
|
1. College of Chemistry and Chemical Engineering, Southwest Petroleum University, Chengdu, Sichuan, China;2. CCDC Drilling Fluid Technical Service Company, Chengdu, Sichuan, China
|
Abstract: |
Objective The aim is to solve the problem that the contamination type of the completion fluid can not be effectively identified after being contained by brine and residual acid during the drilling of the target layer. Methods The contamination of brine and residual acid with different mass fractions of the completion fluid was measured, and the labels of the data samples with different contamination degrees were revised by K-means clustering algorithm. Different BP neural network models were trained according to the difficulty of obtaining data sample features and the number of hidden layers, and the classification accuracy of the models was tested by leave-one-out cross validation method. Results It is found that the more features the data samples possess, the higher classification accuracy of the trained BP neural network could be achieved, while more hidden layers would lower the classification accuracy. The BP neural network model with one hidden layer was subsequently established with data samples that contain four kinds of features including "rheology+aging+filtration loss+well name". The average classification accuracy rate reached as high as 93.18%. Conclusion sThe BP neural network model trained by rheology and filtration loss features can be quickly deployed in the oil-testing sites to solve the problem of failing to identify the type of completion fluid contamination due to the lack of special equipment in the field. |
Key words: completion fluid contamination type computer simulation K-means clustering algorithm neural network leave-one-out cross validation |