基于机器学习的煤层气产能标定智能算法及影响因素分析

Forecasting and influencing factor analysis of coalbed methane productivity utilizing intelligent algorithms

煤层气是我国常规天然气现实可靠的战略补充资源之一,智能化标定煤层气产能对于天然气工业的发展具有重要意义.对山西沁水盆地某煤层气区块的煤层气井进行了实际地质、生产数据的收集及数据预处理,提出了基于生产井史的煤层气单井产能计算公式.利用预处理后的生产数据及储层数据,建立了基于深度神经网络、支持向量回归机、随机森林的煤层气产能标定智能算法,对煤层气单井产能进行了预测,比较了三种机器学习模型的预测结果,分析了不同排采天数的生产数据作为输入参数对模型精度的影响.基于预测效果最好的机器学习模型,进行了动态参数(排采前期的日产气、日产水和井底流压)和静态参数(煤层埋深、孔隙度、渗透率、煤层厚度和含气量)对煤层气产能标定模型的重要性程度分析.结果表明:三种机器学习模型标定煤层气单井产能的平均决定系数为0.828,其中深度神经网络模型决定系数最高,达0.923;增加生产数据前期排采天数,决定系数增长趋势明显,之后增长趋势减缓并最终趋于平稳;动态参数和静态参数对产能的影响都较强,这两类参数对于产能预测模型的贡献度分别为48%和52%.

Coalbed methane(CBM)is one of the realistic and reliable strategic supplementary resources of conventional natural gas in China,and intelligent calibration of CBM production capacity is of great importance for developing the natural gas industry.Actual geological and production data were collected and preprocessed for CBM wells in a CBM block in the Qinshui Basin,Shanxi Province.For CBM wells that have been exploited for a long time,a calculation formula based on production well history was proposed for single well productivity based on formation pressure in static data,daily gas production,and bottom-hole flow pressure in production data.The accuracy of the calculation formula was determined using the maximum daily gas production curve of CBM wells.Using the preprocessed production and reservoir data,an intelligent algorithm was developed for CBM productivity calibration based on deep neural network(DNN),support vector regression machine,and random forest for predicting the productivity of a single CBM well.A comparison of the prediction results of three machine learning models was performed,and the effect of production data for different discharge days as input parameters on model accuracy was examined.Based on the machine learning model with the best prediction effect,the importance of dynamic parameters(daily gas production,daily water production,and bottom-hole flow pressure in the early stage of drainage and production)and static parameters(coal seam depth,porosity,permeability,coal seam thickness,and gas content)to CBM was explored.The findings revealed that the fluctuation coefficient for 100 CBM wells is approximately zero,and the fluctuation range is small,indicating high calculation accuracy.The average coefficient of determination of CBM well productivity calibrated using the three machine learning models is 0.828,and the coefficient of determination of the DNN model is the highest,attaining 0.923,with mean absolute error and root mean square error values of 194.44 and 214.66 m3·d-1,respectively.With increasing days of production data collection in the early stage,the coefficient of determination clearly increases,and then the growth trend slows down and finally becomes stable.Daily gas production,daily water production,production pressure difference,gas content,and permeability are important factors affecting coalbed methane productivity in the early stage of drainage and production.Productivity is highly sensitive to dynamic and static parameters in the early stage of drainage and production,and these two types of parameters contribute 48%and 52%,respectively,to the capacity prediction model.