The viability of enhanced coalbed methane recovery (ECBM) and enhanced shale gas recovery (ESGR) are abundantly explored in various studies since they present a solution for ongoing energy demand and environmental crisis. As a matter of challenge, the prediction of the gas sorption profile poses a significant obstacle in the development of such resources. In this regard, mathematical models, numerical simulations, empirical correlations, and experimental measurements suffer from over-simplification, computationally extensive calculations, time-consuming, and costly processes. This work examines different machine learning methods, from shallow to deep learning, to investigate their capability to model 3804 data points of methane (CH4) and/or carbon dioxide (CO2) sorption capacity in shale and coal at different thermodynamic conditions. Percentage of CO2 in injection gas, rock type, total organic carbon (TOC), moisture, temperature, and pressure were taken as input. Hyperparameter tuning was executed by Optuna optimizer. According to the results, the random forest (RF) was the most proficient predictor for both test subsets of CH4 (MAE = 0.0864, MSE = 0.0231, RMSE = 0.1520) and CO2 (MAE = 0.0529, MSE = 0.0533, RMSE = 0.2308) sorption capacity. Based on the trend prediction simulation, RF predictors were able to properly capture the single and binary sorption capacities. More importantly, they successfully predict the abnormal descending behavior of CO2 sorption in high pressures. In addition, a sensitivity analysis was carried out to acquire the importance of input features and model hyperparameters in the training process. Input feature impacts were consistent with technical expectations of geology and reservoir engineering aspects. Both models are capable of being applied to lab scale and further developed for their application in reservoir scale simulators.
Read full abstract