摘要
中国低煤阶煤层气资源丰富,且煤层气作为一种清洁能源,其开发和利用可有效地缓解我国天然气资源短缺问题,但是商业化规模开发稍显不足,亟需系统研究。煤层气高效开发的前提是有利区优选,但目前煤层气开发选区评价均涉及一定的主观人为因素,会间接影响或干扰预测效果。将黄陇煤田彬长矿区大佛寺井田低煤阶作为研究对象,以生产实际数据为基础,采用机器学习中的随机森林算法对该区煤层气开发选区做出预测。结果表明:①Pearson关联系数(PCC)分析表明含气量、灰分、煤层净厚度、构造位置、顶板厚度、渗透率、储层压力和埋深这8个煤层气产出相关参数相互独立,可用于模型建立;②随机森林算法将煤层气开发选区划分为5类不同程度的区域,其中Ⅰ类(极高)和Ⅱ类(高有利)区占整个研究区域的13.88%,主要分布在井田的中部,西部存在Ⅱ类(高有利区)分布,后续开发部署井位也可着重考虑,而井田的东南部不适于后续部署井位;③由接受者操作特征曲线(ROC)可知,一般成功率曲线与预测率曲线下的面积值(AUC)为0.961,表明随机森林模型预测精度较高,结果可靠。通过机器学习算法对煤层气开发选区进行综合预测,可规避传统算法中的人为主观因素,且具有较高的精度,可为后续非常规油气开发选区提供一定的理论参考依据。
In China,low-rank coal and coalbed methane resources are abundant,meanwhile,as a kind of clean energy,the development and utilization of coalbed methane(CBM)can effectively alleviate the shortage of natural gas resources,but the commercial scale development is slightly insufficient,and systematic research is urgently needed.The premise of efficient CBM development is the selection of favorable areas,but the current CBM development evaluation involves certain subjective human factors,which will indirectly affect or interfere with the prediction effect.Taking the low-rank coal in the Dafosi minefield in the Binchang mining area of Huanglong Coal Field as the research object,based on the actual production data,the random forest algorithm in machine learning is used to predict the favorable area of coalbed methane in the area.The results show that:①Pearson correlation analysis shows that the gas content,ash content,net thickness of coal seam,structural position,roof thickness,permeability,reservoir pressure and burial depth are eight mutually independent CBM output-related parameters and can be used for model establishment;②The Random Forest algorithm divides the CBM development area into five types of areas with different degrees,of which typeⅠ(extremely high)to Ⅱ(highly favorable)areas account for 13.88% of the entire study area,mainly distributed in the middle of the well field.The southeast is not suitable for subsequent deployment of well locations,and there is a distribution of highly favorable areas in the west,so the well locations for subsequent development and deployment should also be considered.③It can be obtained from the receiver operating characteristic(ROC)curve,and the area under the ROC curve(AUC)is 0.961,indicating that the Random Forest model has high prediction accuracy and reliable results.Using machine learning algorithms for comprehensive prediction of CBM favorable areas can avoid human subjective factors in traditional algorithms,and can provide a certain theoretical reference for
作者
陈跃
王丽雅
李国富
张林
杨甫
马卓远
高正
CHEN Yue;WANG Liya;LI Guofu;ZHANG Lin;YANG Fu;MA Zhuoyuan;GAO Zheng;无(College of Geology and Environment,Xi’an University of Science and Technology,Xi’an,Shaanxi 710054,China;Key Laboratory of coal and coalbed methane co-mining technology,Jincheng,Shanxi 048000,China;Shaanxi Coalbed Methane Development Corp.Ltd.,Xi’an,Shaanxi 710065,China;Key Laboratory of Coal Resources Exploration and Comprehensive Utilization,Ministry of Natural Resources,Xi’an,Shaanxi 710021,China)
出处
《油气藏评价与开发》
CSCD
2022年第4期596-603,616,共9页
Petroleum Reservoir Evaluation and Development
基金
国家自然科学基金项目“润湿性制约下低阶煤不同煤岩组分甲烷解吸机制”(41902175)
山西省揭榜招标项目“基于气藏工程的煤层气井人工智能排采技术与示范”(20201101002)
中国博士后科学基金项目“低煤阶镜煤与暗煤润湿性差异及对甲烷解吸的影响”(2019M653873XB)。
关键词
随机森林
开发选区预测
煤层气
低阶煤
黄陇煤田
Random Forest
development constituency forecast
coalbed methane(CBM)
low rank coal
Huanglong Coal Field
