基于机器学习的气井产气产水动态预测方法及应用  

A ML-based method to dynamically predict gas and water production in gas wells

作　　者：朱秋琳 梁艳萍 赵玉琴 杨静 曾健 卢俊卿 张祥熙 ZHU Qiulin;LIANG Yanping;ZHAO Yuqin;YANG Jing;ZENG Jian;LU Junqing;ZHANG Xiangxi(No.1 Gas Production Plant,PetroChina Qinghai Oilfield Company,Golmud,Qinghai 816000,China;Chengdu Compressor Branch,CNPC Ji-chai Power Company Limited,Chengdu,Sichuan 610051,China)

机构地区：[1]中国石油青海油田公司采气一厂,青海格尔木816000 [2]中国石油济柴动力公司成都压缩机分公司,四川成都610051

出　　处：《天然气技术与经济》2025年第2期22-31,共10页Natural Gas Technology and Economy

摘　　要：为了解决气井产气、产水数据获取滞后制约单井预警时效的问题,基于机器学习算法构建了气井产气、产水动态预测方法,提升了异常生产井单井预警的及时率。研究结果表明:①通过特征参数相关性分析,优选出工作制度、油压等7项相关性高的核心特征参数建立多维数据集;②采用3种机器学习算法对模型进行训练与测试(训练集与测试集按2∶1划分),其中随机森林模型稳定性及准确性表现最优;③基于随机森林算法,结合气井生产特征构建分类模型与“一井一策”的特色预测模型,该模型测试集的平均相对误差在20%以内,采用Python程序搭建Windows系统服务稳定运行1年后,预测准确率稳定在80%以上。结论认为:①模型误差主要来源于下述6个方面,数据本身质量、数据取值规则、特征参数集的划分、模型自身设计、气井特殊生产特征、气井流体运动规律改变;②该方法能够实时预测单井产气、产水数据,进而提升了异常生产井的预警及时率,有助于气田生产管理从数字化向智慧化的转型。The lagging data acquisition of gas and water production in gas wells restricts the validity of early warning for an individual well.Thus,based on machine learning(ML)algorithms,a method to dynamically predict gas and water production was established to improve the validity in abnormal producers.Results show that(i)with high correlation,seven core characteristic parameters including working system and tubing pressure are selected after the correlation analysis to construct a multi-dimensional data set;(ii)three ML algorithms are adopted for model training and testing(the ratio of training and testing sets is 2:1),among which the best stability and the highest accuracy are achieved in the random forest model;and(iii)combined with production characteristics,one classification model and another distinguished prediction model of"one strategy per well"are established according to the random forest algorithm.For these two models,the mean relative error in testing set is within 20%.After running stably in Windows'Python program for one year,the models remains their prediction accuracy over 80%.In conclusion,for all models,the error mainly results from data quality,data-setting rules,division of characteristic parameter sets,model design,and special production characteristics and changes of fluid flowing laws in gas producers.What's more,this ML-based method can predict gas and water production in the individual well in real time,and improve the early-warning validity in those abnormal producers,which is conducive to the transformation of gasfield production and management from digitalization to intelligence.

关 键 词：随机森林 机器学习 动态预测 预测模型 特征参数 单井预警 涩北气田 

分 类 号：TE37[石油与天然气工程—油气田开发工程]