【作者】 王玉婷；

【导师】 闫长辉；

【作者基本信息】 成都理工大学 ， 油气田开发工程， 2021， 硕士

【摘要】 页岩气田储层低孔低渗,气体赋存状态多样,采用水平井多段压裂技术开采,在气藏、工程等多因素影响下页岩气井生产动态特征与常规气井存在差异,常规的产量递减研究方法对其适用性低、误差大。随着数字油田与生产自动化的蓬勃发展,机器学习方法实现了油气田数字化建设、智能生产、运行、管理等;由于油气田开发的数据具有多源、多维、关系复杂、模式隐蔽的特点,通过机器学习可得到数据的变化规律,建立可预测的数据模型,从而提高油气藏动态分析效率及预测精度。本文采用机器学习方法研究页岩气产量递减,提高页岩气田产量预测精度。本文以CN页岩气田为研究对象,首先采用相关性方法分析影响页岩气井自然递减阶段产气量的气藏因素、工程因素,以产气量作为标签值,结合特征重要性确定的重要因素作为机器学习特征值,并采用K-Means聚类算法按投产初期产量进行气井产能分类;其次用非线性回归算法(随机森林、支持向量机、长短时记忆网络)对不同类井分别建立多个产量递减模型;然后以均方误差、45°等值线拟合精度为模型评估指标,优选不同类井产量递减模型;最后将优选模型和常规递减模型进行对比,并在建立模型基础上编制产量预测软件,将其在单井中进行产量预测应用。结果表明:(1)在众多气藏因素、工程因素中,地层压力、孔隙度、吸附气量和用液强度、加砂强度为影响产量递减的重要因素,以此作为机器学习输入特征;(2)页岩气井按产能可分为低产井(<15.99×10~4m~3/d)、中产井(15.99×10~4m~3/d～27.25×10~4m~3/d)、高产井(>27.25×10~4m~3/d)三类,且同类产能井自然递减阶段的递减规律相似:低产井初期产能低,产量递减速度慢;中产井产量递减较快;高产井初期天然能量充足,产量递减相对低产井、中产井更快;(3)对不同类页岩气井进行模型优选表明:低产井适用随机森林模型,中、高产井适用长短时记忆网络模型,且相较于页岩气常规递减模型,建立的机器学习产量递减模型能准确识别产量递减趋势,提高预测产量精度。本文研究拓展了页岩气井的产量递减研究方法。更多还原

【Abstract】 Due to the influence of multiple factors such as low porosity and low permeability of reservoirs,different occurrence states of gas,and production with multi-stage fracturing technology of horizontal wells in shale gas fields,the production performance characteristics of shale gas wells are different from conventional gas wells,which could make the conventional decline method with low applicability and large errors.With the vigorous development of digital oil fields and production automation,machine learning methods have facilitated the digital construction,intelligent production,operation,and management of oil and gas fields.As oil and gas field development data has the characteristics of multiple sources,multiple dimensions,complex relationships,and hidden mode,the law of data changes can be obtained through machine learning and a predictable data model can be established to improve the efficiency of oil and gas reservoir dynamic analysis and prediction accuracy.In this paper,machine learning method is used to study the decline of shale gas production and improve the production forecasts.CN shale gas field is picked to be the paper research object.First,the correlation method is used to analyze the gas reservoir factors and engineering factors that affect the production of shale gas wells in the natural decline stage and the production is used as the tag value,which the important factors are determined as the machine learning feature values in combination with the weight of features.Besides,the K-Means clustering algorithm is used to press classification of gas well productivity at the initial stage of production.Secondly,the nonlinear regression algorithms(Random Forest,Support Vector Machine,Long Short-Term Memory)are used to establish multiple production decline models for different types of wells;Then,the mean square error and45°contour fitting accuracy are used as evaluation indexes to optimize the production decline model for different types of wells;Finally,the optimal model is compared with the conventional decline model,and the production forecasting software is developed on the basis of the established model,which is applied to the production prediction in a single well.The results show that:(1)Among many gas reservoir factors and engineering factors,formation pressure,porosity,adsorbed gas volume,liquid strength and sand strength are most important factors affecting production decline,which are taken as the input features of machine learning.(2)According to the productivity,shale gas wells can be divided into three types:low productivity wells(<15.99×10~4m~3/d),middle productivity wells(15.99×10~4m~3/d～27.25×10~4m~3/d),and high productivity wells(>27.25×10~4m~3/d).The decline law of the natural decline stage of similar productivity wells is similar:low productivity wells have low initial productivity and slow production decline,middle productivity wells decreases rapidly,high productivity wells have sufficient natural energy at the initial stage,and their production decline is faster than that of low productivity wells and middle productivity wells.(3)Model optimization for different types of shale gas wells shows that:low productivity wells are suitable for random forest model,medium and high productivity wells are suitable for long short-term memory model.Compared with the conventional decline models of shale gas,the production decline models established by machine learning method can accurately identify the production decline trend and improve the accuracy of the forecast production.This study expands the research methods of the decline of shale gas production forecast.更多还原