Prediction of total gas content in low-resistance shale reservoirs via models fusion — Taking the Changning shale gas field in the Sichuan Basin as an example

Abstract

The low-resistivity phenomenon appeared during the development of shale gas in the Changning area of southern Sichuan. The accurate prediction of GAST is crucial in formulating optimal development plans for low-resistivity shale reservoirs. A series of formulas such as Archie in traditional electrical logging is no longer applicable to the characterization of gas-bearing properties of low-resistivity shale reservoirs. In this paper, we propose a fusion model for predicting the total gas content of low-resistivity shale reservoirs using conventional logging and total gas content data of core analysis. The model combines three popular machine learning models: random forest, extreme gradient boosting tree, and deep neural network. First, rock physics logging knowledge and the Spearman coefficient method are used to select suitable logging data types as the input of three models. Second, three models are trained by using the selected logging and core gas content data to achieve optimal performance. Finally, based on the stacking model fusion strategy, the three are fused to obtain a fusion model with better generalization performance. The experimental results show that the average absolute errors of the fusion model in predicting the total gas content of low-resistance wells and ultra-low resistance wells are 0.025 and 0.023 respectively, and the R2 coefficients reach 0.952 and 0.964. The fusion model effectively predicts the gas content of low-resistivity shale reservoirs and has high application value.