Identifying Types and Key Features of Typical Production Performance of Coalbed Methane with Interpretable Residual Graph Convolutional Model

Abstract

The production of coalbed methane (CBM) wells is positively correlated with their production performance, and key features of typical production performance can be applied to determine the high production exploration targets. However, accurately classifying the production types of CBM wells and rationally identifying the key controlling factors among them are challenging due to the strong heterogeneity of CBM reservoirs. The data-driven “black-box” algorithms utilized in previous studies often suffer from limited interpretability due to a lack of sufficient domain theoretical foundation. This paper proposes an interpretable residual graph convolutional neural network model (I–RGCN) for classifying the production types and for identifying key features of typical production of CBM wells from spatial relationships and attribute data. This model constructs a topological graph structure based on the spatial correlations among wells and utilizes the dynamic time warping algorithm to assess the similarity of geological feature parameters among CBM wells, incorporating these as edge weights in the model for accurate classification of CBM production types. Subsequently, the GNNExplainer was used to rank the importance of features during the model's decision-making process. Final experiments conducted on datasets from the Fanzhuang–Zhengzhuang block within the Qinshui coalfield demonstrated that the I–RGCN achieves accuracy of > 84% and F1 score of ~ 65%, and outperformed other baseline models and enhanced the interpretability of the results obtained. Thus, this paper offers a novel and interpretable research methodology for the classification of CBM production types and the identification of key features of the production performance of CBM.