Coal bed methane (CBM) is one of realistic and reliable strategic supplementary resource for conventional natural gas in China. Intelli-gent forecasting of CBM productivity is of great significance for the development of natural gas industry. The geological and production data of CBM wells in a CBM block in Qinshui Basin, Shanxi Province were collected and preprocessed. A formula, based on produc-tion history, for calculating the productivity of a single CBM well was put forward. Using preprocessed production data and geological data, intelligent algorithms for CBM productivity forecasting based on Deep Neural Network (DNN), Support Vector Regression (SVR) and Random Forest (RF) were established. Then the single well productivity of CBM was predicted and the prediction accuracy of three machine learning models was compared. The influence of production data of different exploitation days as input parameters on the mod-el accuracy was also analyzed. Finally the sensitivity analysis of CBM productivity to dynamic parameters (daily gas production, daily water production, and bottom-hole flow pressure on start-up stage) and static parameters (buried depth of coal seam, porosity, permea-bility, thickness of coal seam, and gas content) was carried out. The results show that the average prediction accuracy of the three ma-chine learning models is 0.828, among which the prediction accuracy of DNN model is the highest, reaching 0.923. With the increase of exploitation days of production data, the prediction accuracy shows an obvious growth trend, and then the growth trend slows down and finally stabilizes. The productivity is sensitive to both dynamic and static parameters, accounting for about 48% and 52% respec-tively.
DownLoad: