Machine learning method for shale gas adsorption capacity prediction and key influencing factors evaluation

Abstract

Shale gas plays a pivotal role in the global energy landscape, emphasizing the need for accurate shale gas-in-place (GIP) prediction to facilitate effective production planning. Adsorbed gas in shale, the primary form of gas storage under reservoir conditions, is a critical aspect of this prediction. In this study, a machine learning Gaussian process regression (GPR) model for methane adsorption prediction was established and validated using published experimental data. Five typical variables, i.e., total organic carbon (TOC), clay minerals, temperature, pressure, and moisture were considered, which were derived from the Marine shale of the Longmaxi formation in the Sichuan Basin through correlation analysis. The performance of the GPR model was compared with the widely used an extreme gradient boosting model. It turned out that our GPR model had better accuracy for predicting methane adsorption in shale with an average relative error of less than 3%. Furthermore, a variance-based sensitivity analysis method in conjunction with kernel density estimation theory was employed to conduct a global sensitivity analysis, quantifying the nonlinear influence of each variable methane adsorption. The findings indicate that TOC is the most significant factor affecting methane adsorption, while clay minerals have a limited direct impact but can enhance their influence through interactions with other influencing factors. Finally, based on the GPR model, a GIP prediction method was proposed that eliminates the need for calculating the density of the adsorbed phase. These findings are expected to extend the shale gas reserve assessment methodologies and offer valuable insight for further exploring the adsorption mechanisms of shale gas.