基于斑马算法优化支持向量回归机模型预测页岩地层压力  

Prediction of shale formation pore pressure based on Zebra Optimization Algorithm-optimized support vector regression

作　　者：赵军[1] 李勇[1] 文晓峰 徐文远 焦世祥 ZHAO Jun;LI Yong;WEN Xiaofeng;XU Wenyuan;JIAO Shixiang(School of Geoscience and Technology,Southwest Petroleum University,Chengdu 610500,China;Changqing Branch,China Petroleum Group Logging Co.,Ltd,Xi’an 710201,China)

机构地区：[1]西南石油大学地球科学与技术学院,成都610500 [2]中国石油集团测井有限公司长庆分公司,西安710201

出　　处：《岩性油气藏》2024年第6期12-22,共11页Lithologic Reservoirs

基　　金：中国石油集团测井有限公司重点攻关项目“页岩油工程品质评价方法研究”(编号:ZYCJ-CQ-2023-JS-2337)资助。

摘　　要：针对陇东地区三叠系延长组7段(长7段)页岩孔隙结构复杂、非均质性强、地层压力预测精度较低等问题,提出了一种基于斑马算法优化支持向量回归机(ZOA-SVR)模型预测地层压力的方法,并在实际钻井中进行了应用,将预测结果与基于机器算法的模型和常规地层压力预测方法结果进行了对比。研究结果表明:①ZOA-SVR模型以实测地层压力数据为目标变量,优选与陇东地区长7段页岩地层压力数据关联度达到0.70以上的深度、声波时差、密度、补偿中子、自然伽马、深侧向电阻率、泥质含量等7个参数作为输入特征参数,设置训练样本数为40,交叉验证折数为5,初始化斑马种群数量为10,最大迭代次数为70,对惩罚因子和核参数进行优化并建模,参数优化后拟合优度指标R2达到0.942,模型预测的地层压力数据在训练集和测试集上的绝对误差均低于1 MPa,预测测试集地层压力数据与实测压力数据的平均相对误差为2.42%。②ZOA-SVR模型在研究区长7段地层压力预测中优势明显,比基于粒子群优化算法、灰狼算法和蚁群算法的模型具有更好的参数调节及优化能力,R2分别提高了0.209,0.327,0.142;比等效深度法、Eaton法、有效应力法预测的地层压力精度更高,相对误差分别降低了32.53%,15.31%,5.91%。③ZOA-SVR模型在实际钻井中的应用结果显示,研究区长7段地层压力在垂向上分布较稳定,泥页岩段的地层压力高于砂岩段,地层压力系数主要为0.80~0.90,整体上属于异常低压环境,与实际地层情况相符。To address the complexity and heterogeneity of the pore structure of seventh member of Triassic Yanchang Formation(Chang 7 member)in Longdong area,a method based on the Zebra Optimization Algorithm-optimized Support Vector Regression(ZOA-SVR)model was proposed.This method utilizes formation pressure test data and conventional well logging curves.It was applied in actual wells,and compared with other machine learning models,as well as conventional methods for predicting formation pressure.The results show that:(1)The ZOA-SVR model uses actual formation pressure data as target variables,selecting seven input feature parameter,such as depth,sonic transit time,density,compensated neutron,natural gamma,deep resistivity,and clay content,associated with shale formation pressure data correlation above 0.7 in the study area.The model was trained with 40 samples,validated with 5-fold cross-validation,and optimized with 10 initial Zebra populations and a maximum of 70 iterations.After optimizing penalty factors and kernel parameters,the model achieved a fit indicator R2 of 0.942.Predicted formation pressure data had absolute errors below 1 MPa on both training and test sets,with an average relative error of 2.42%compared to measured data.(2)The ZOA-SVR model demonstrated significant advantages in predicting of Chang 7 Member formation pressure in the study area compared to models based on Particle Swarm Optimization,Grey Wolf Algorithm,and Ant Colony Algorithm.It shows a better parameter adjustment and optimization capabilities,with coefficient of determination increases of 0.209,0.327,and 0.142 respectively.It also exhibited higher accuracy in pressure prediction compared to Equivalent Depth Method,Eaton Method,and Effective Stress Method,reducing relative errors by 32.53%,15.31%,and 5.91%respectively.(3)Application of the ZOA-SVR model in actual wells indicated stable vertical distribution of chang 7 Member formation pressure in the study area.Pressure in mud shale sections is higher than the sandstone sections,with press

关 键 词：页岩 地层压力 斑马优化算法 支持向量回归机 机器学习 测井曲线 长7段 三叠系 陇东地区 

分 类 号：P618.13[天文地球—矿床学] TP18[天文地球—地质学]