基于VC-SVM与粒子群算法的卡钻智能预测方法  

Intelligent Prediction Method of Pipe Sticking Based on VC-SVM and Particle Swarm Optimization

作　　者：刘子豪 宋先知[1] 朱硕 叶山林 张诚恺 马宝东 祝兆鹏 Liu Zihao;Song Xianzhi;Zhu Shuo;Ye Shanlin;Zhang Chengkai;Ma Baodong;Zhu Zhaopeng(College of Petroleum Engineering,China University of Petroleum(Beijing);Jianghan Machinery Research Institute Limited Company of CNPC;College of Mechanical and Transportation Engineering,China University of Petroleum(Beijing))

机构地区：[1]中国石油大学(北京)石油工程学院 [2]中石油江汉机械研究所有限公司 [3]中国石油大学(北京)机械与储运工程学院

出　　处：《石油机械》2024年第10期1-11,共11页China Petroleum Machinery

基　　金：国家重点研发计划“变革性技术关键科学问题”“复杂油气智能钻井理论与方法”(2019YFA0708300);中国石油天然气集团有限公司-中国石油大学(北京)战略合作科技专项“物探、测井、钻完井人工智能理论与应用场景关键技术研究”(ZLZX2020-03);国家自然科学基金杰出青年基金项目“油气井流体力学与工程”(52125401);中国石油大学(北京)科学基金项目“深井气侵自动压井井底压力智能预测和控制方法”(2462022SZBH002)。

摘　　要：在油气钻探过程中,由于井下条件复杂、地层认识不清等因素,导致卡钻事故频发,严重制约钻井效率。目前国内外学者所研究的卡钻预测方法在准确性、时效性及迁移性等方面仍存在不足。为此提出了一种融合集成学习思想与智能优化算法的卡钻智能预测方法。该方法根据实际井场的卡钻数据,基于合理的标签标定方法,将标签准确定位于卡钻发生前而非卡死点;通过参数相关性分析、表征意义分析、时效性以及可信性分析优选了7个输入参数;使用了随机森林(RF)、支持向量机(SVM)和BP神经网络3种算法建立了卡钻预测模型,并对比了各模型在卡钻与非卡钻样本比例严重不均时(卡钻与非卡钻比例1∶117)的表现;然后使用投票分类器(VC)将多个模型集成,并分类预测,优选SVM模型作为卡钻预测基模型,使用集成学习的思想加以改进,并采用粒子群算法同时对多个SVM分类器进行超参数优化,简化了调参过程的同时实现了耦合寻优。最终以某区块10次卡钻样本进行训练测试。测试结果表明,改进后的模型可有效寻找不同类别卡钻的超平面,迁移预测虚警率可控制在9%,漏警率不到7%,有效预测了每一次卡钻的大部分数据点。研究结果有望提高现场钻井风险预警效率,为保障油气井安全高效钻进提供支撑。In the process of well drilling,pipe sticking frequently occurs due to the factors such as complex downhole conditions and unclear knowledge of the strata,seriously restricting the drilling efficiency.However,the existing prediction methods of pipe sticking are defective in aspects such as accuracy,timeliness and transferability.This paper presents an intelligent prediction method of pipe sticking based on ensemble learning idea and intelligent optimization algorithm.First,based on the actual sticking data at well site and a reasonable label calibration method,the label was accurately positioned at the point before the occurrence of sticking rather than at the freeze-in point,and by means of parametric dependence analysis,characterization significance analysis,and timeliness and creditability analysis,7 input parameters were selected.Second,three algorithms,i.e.random forest(RF),support vector machine(SVM)and BP neural network,were used to build a sticking prediction model,and the performances of each model under seriously uneven proportions of sticking and non-sticking samples(sticking to non-sticking ratio:1∶117)were compared.Third,the SVM model was selected as the basic model for the prediction of pipe sticking,and improved using the ensemble learning idea.Meanwhile,the particle swarm optimization(PSO)was used to simultaneously conduct hyperparameter optimization on multiple SVM classifiers,simplifying the parameter tuning process while achieving coupled optimization.Finally,the 10 times of sticking samples of a block were used to conduct training test.The results show that the improved model can effectively search for hyperplanes of different types of pipe sticking,with a transfer prediction false alarm rate controlled at 9%and a missed alarm rate of less than 7%,effectively predicting most of the data points of each pipe sticking.This study is expected to improve the risk warning efficiency of field drilling,and provides support for ensuring safe and efficient drilling of wells.

关 键 词：卡钻智能预测 支持向量机 BP神经网络 投票分类器 粒子群算法 迁移能力测试 耦合寻优 

分 类 号：TE28[石油与天然气工程—油气井工程]