Data-driven Modeling of the Methane Adsorption Isotherm on Coal Using Supervised Learning Methods: A Comparative Study

Abstract

Methane adsorption isotherm on coals is key to the development of coalbed methane (CBM). Laboratory measurement of adsorption isotherm is time-consuming. This paper presents a comparative study on the accuracy and robustness of seven supervised learning (SL) methods in estimating the methane adsorption isotherm based on coal properties. The SL methods used include the Gaussian process regression (GPR), kernel ridge regression (KRR), classifier and regression tree (CART) and four ensemble decision tree methods (random forests (RF), Adaboost, gradient boosting decision tree (GBDT) and extreme boosting (XGBoost)). The results show that all these SL methods are capable of correlating methane adsorption amounts with the feature variables with reasonable accuracies in the training stage. However, the KRR, GBDT and XGBoost are demonstrated to outperform other SL techniques in terms of the robustness and generalization capability, which therefore are recommended for fast estimation of the methane adsorption isotherms on coals.