基于机器学习的酸性气藏地下储气库硫化氢含量预测方法  

Prediction of H2S content in sour gas reservoir underground gas storage based on machine learning

作　　者：冯国庆[1,2] 杜勤锟 周道勇 蔡家兰 程希[1] 莫海帅 FENG Guoqing;DU Qinkun;ZHOU Daoyong;CAI Jialan;CEHNG Xi;MO Haishuai(State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Petroleum Engineering School,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Chongqing Division,PetroChina Southwest Oil&Gas Field Company,Chongqing 400021,China)

机构地区：[1]油气藏地质与开发全国重点实验室·西南石油大学,四川省成都市610500 [2]西南石油大学石油与天然气工程学院 [3]中国石油西南油气田公司重庆气矿

出　　处：《天然气工业》2025年第2期159-169,共11页Natural Gas Industry

基　　金：油气藏地质及开发工程全国重点实验室(西南石油大学)开放课题“基于物理模拟和数据模型的过套管地层真电阻率反演方法研究”(编号:PLN2022-14)。

摘　　要：地下储气库(以下简称储气库)中含有硫化氢等有害气体,不仅影响储气库的安全运行,还直接对环境造成严重污染,准确预测储气库采出气组分中H2S的含量具有重要意义。目前,常采用油藏数值模拟的组分模型来预测H2S含量,但其计算过程复杂且耗时较长,不能方便快捷地用于储气库单井H2S的含量预测。为此,以HCX储气库为研究对象,在建立储气库的机理模型并开展数值模拟的基础上,以机理模型计算的储气库多周期H2S预测结果为样本集,应用多输出支持向量回归(MSVR)、长短期记忆网络(LSTM)、人工神经网络(ANN)3种机器学习算法建立了硫化氢含量的智能代理模型,并对3种模型预测精度进行对比分析。研究结果表明:①长短期记忆网络模型具有适中的训练时间、较好的预测精度,可将该模型作为HCX储气库的H2S预测智能代理模型;②进一步对LSTM模型的训练数据和过渡拟合问题进行优化,确定最佳训练数集1500组,最佳丢弃率为0.2,隐含层设置范围可控制在层数1~2层,节点数30~60个;③经HCX储气库的实例应用表明,建立的LSTM智能代理模型能够准确预测储气库采出气中H2S的含量。结论认为,经过优化的LSTM算法智能代理模型具有较好的外推性,该研究成果可为含H2S储气库的建设和安全高效运行提供技术支持。Underground gas storage(UGS)contain harmful gases such as H2S,which may impact its safe operation,and may also directly cause serious environmental pollution.Therefore,it is of great significance to accurately predict the content of H2S in the gas produced from UGSs.At present,the component model of reservoir numerical simulation is often used to predict H2S content,but its calculation process is complex and time-consuming,so it cannot predict H2S content in a single well of gas storage conveniently and quickly.Taking the HCX UGS as the research object,this paper establishes a UGS mechanism model and carries out numerical simulation.Then,with the multi-cycle H2S prediction results of the HCX UGS calculated by the mechanism model as the sample set,an intelligent surrogate model of H2S content is established by means of three machine learning algorithms,i.e.,multi-output support vector regression(MSVR),long short-term memory network(LSTM),and artificial neural network(ANN).In addition,the prediction accuracy of the three models are comparatively analyzed.And the following research results are obtained.First,the LSTM model has moderate training time and better prediction accuracy,and it can be used as the intelligent surrogate model of H2S content prediction for HCX UGS.Second,the training data and transition fitting of LSTM model are further optimized.It is determined that the optimal number of training sets is 1500,with an optimal discard rate of 0.2,and the setting range of hidden layers can be 1 or 2 with 30-60 nodes.Third,the practical application of the established LSTM intelligent surrogate model in the HCX UGS demonstrates that it can accurately predict the content of H2S in the gas produced from UGSs.In conclusion,the optimized LSTM intelligent surrogate model has good extrapolation.These research results can provide technical support for the construction and safe and efficient operation of sulfurous UGSs.

关 键 词：含硫储气库 数值模拟 组分模拟 硫化氢含量预测 机器学习 长短期记忆网络模型 机器学习模型优化 

分 类 号：TE863[石油与天然气工程—油气储运工程]