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

Abstract

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.

Keywords

• coalbed methane
• qinshui basin
• single well productivity
• intelligent algorithm
• machine learning