Xinjiang Oil & Gas ›› 2024, Vol. 20 ›› Issue (1): 13-20.DOI: 10.12388/j.issn.1673-2677.2024.01.002

• OIL AND GAS EXPLORATION • Previous Articles     Next Articles

A Machine Learning-Based Bridging Particle Size Recommendation Method for Lost Circulation Control

  

  1. 1. CNPC Research Institute of Engineering Technology Co.,Ltd.,Changping 102206,Beijing,China;

    2. PetroChina Tarim Oilfield Company,Korla 841000,Xinjiang,China;

    3. China University of Petroleum (Beijing),Changping 102249,Beijing,China;

    4. Sinopec Northwest Oilfield Company,Urumqi 830011,Xinjiang,China.

  • Online:2024-03-06 Published:2024-03-06

基于机器学习的堵漏颗粒粒径推荐方法

  

  1. 1.中国石油集团工程技术研究院有限公司,北京昌平  102206;

    2.中国石油塔里木油田分公司,新疆库尔勒  841000;

    3.中国石油大学(北京),北京昌平  102249;

    4.中国石油化工集团股份有限公司西北油田分公司,新疆乌鲁木齐  830011 。


  • 作者简介:刘凡(1990-),2019年毕业于中国石油大学(北京)油气井工程专业,博士,高级工程师,目前从事防漏堵漏技术研究。(Tel)010-80162079(E-mail)ferman-liu@hotmail.com
  • 基金资助:

    1、中国石油重大技术现场试验项目“恶性井漏防治技术与高性能水基钻井液现场试验”(2020F-45);

    2、中国石油重大科技专项“海外大型碳酸盐岩油藏高效上产关键技术研究”(2023ZZ19)。

Abstract:

Lost circulation is a key technical challenge in oil and gas exploration,and bridge plugging is the most commonly used method for lost circulation control. As an important parameter,the size of bridging particles directly affects the plugging. At present,the size selection mainly relies on experience,lacking a scientific and effective method. In light of this,this paper investigates a particle size recommendation method for lost circulation control based on machine learning algorithms. The basic data used for this method are well-logging,mud logging,and lost circulation control data from 126 completed wells in Kuqa piedmont area of Tarim Basin. The input layer adopts 23 main parameters screened based on the Pearson algorithm,and the output layer is in 4 bridging particle size ranges of 0-750 μm,750-1 500 μm,1 500-4 000 μm and>4 000 μm. 10 commonly used machine learning algorithms are trained and tested to determine the accuracy of three types of datasets:well logging data,mud logging data,and combination of the former two types. It is found that the scores of each algorithm for the well logging + mud logging dataset are generally higher than those for the well logging and mud logging datasets. For the combined dataset,the support vector machine and extremely randomized trees algorithms have the highest F1 scores of above 0.9. The bridging particle size recommendation model based on the support vector machine and extremely randomized trees algorithms is validated twice on a well in Kuqa piedmont area. The predicted results of bridging particle size of the two algorithm models are consistent with the actual bridging effect in the field. This method exhibits good application prospects in the scientific optimization of bridging particle size.

Key words:

lost circulation control, bridging technology, bridging particle size, machine learning algorithm, Kuqa piedmont, Tarim Basin

摘要:

井漏是油气勘探领域的重大技术难题。桥接堵漏是最常用的堵漏技术手段,其中架桥颗粒粒径是关键参数,直接影响堵漏成败,目前架桥颗粒粒径的选择主要依赖经验,缺乏科学有效的方法。研究了一种基于机器学习算法的堵漏颗粒粒径推荐方法。该方法基础数据为塔里木盆地库车山前区域126口完钻井的测井、录井及防漏堵漏施工数据,其输入层为基于皮尔森算法筛选出的23项主要参数,输出层为0~750 μm、750~1 500 μm、1 500~4 000 μm、4 000 μm以上等4个架桥颗粒粒径区间。训练测试了10种常用的机器学习算法在测井数据、录井数据及测井+录井数据等三类数据集上的准确率,测井+录井数据集上各算法得分普遍高于测井和录井数据集。在测井+录井数据集上,支持向量机和极限随机树算法的F1得分最高,达到0.9以上。基于支持向量机和极限随机树算法的架桥颗粒粒径推荐模型在库车山前一口井上验证2井次,两种算法模型的架桥颗粒粒径预测结果与现场实际堵漏效果一致,在桥堵粒径级配科学优选上具有良好的应用前景。

关键词:

堵漏, 桥堵技术, 架桥颗粒粒径, 机器学习算法, 库车山前, 塔里木盆地

CLC Number: