Accurate calculation of adsorbed shale gas content is critical for gas reserve evaluation and development.However, gas adsorption and desorption experiments are expensive and time-consuming, while physics-based models and empirical correlations are unable to accurately capture the adsorption characteristics fordifferent shales. Langmuir adsorption is one of the most commonly used model for calculating the adsorbedgas content in shale gas reservoirs. However, most existing correlations for the Langmuir pressure andLangmuir volume in the model are oversimplified based on limited experimental data points. Thus theyare not representative of key geological parameters and are far from accurate for prediction in many cases.We developed a variety of machine learning models that are multivariable controlled to quantify shale gasadsorption. The data-driven method subdivides into two procedures: data compilation and machine learningregression. Over 700 data entries, composed of reservoir temperature (T, °C), total organic carbon (TOC,wt%), vitrinite reflectance (Ro,%), Langmuir pressure, and Langmuir volume are compiled from shale gasplays mainly in USA, Canada, and China. Data have been consistently curated, then machine learningapproaches, including multiple linear regression (MLR), support vector machine (SVM), random forest(RF) and artificial neural network (ANN), have been built, trained and tested by partitioning the datainto 75%:25%. For SVM, RF and NN models, 1000 simulations were run and averaged for performancecomparison. MLR identifies non-negligible parameters and general trends for shale gas adsorption. Nonetheless, thecorrelation coefficients from MLR are far from satisfactory. For Langmuir pressure, RF models fit best tothe data entries and the other models follow the order of SVM > ANN > MLR. Particularly, RF modelsshow the highest performance stability with the averaged R-squared value of 0.84 and the maximum of 0.87,indicating a very strong relationship constructed for these 213 data entries. For 485 Langmuir volume dataentries, RF models also perform best while the other three regression methods are comparable. It should benoted that altering machine learning model structure and parameters could significantly affect the regressionresults.
展开▼
