Interpreting coal component content in logging data by combining grey relational analysis and hybrid neural network

The coal component content is an important parameter during the coal resources exploration and exploitation. Previous logging curve regression and single neural network methods have the disadvantages of low accuracy and weak generalization ability in calculating coal component content. In this study, a GRA-HNN method was proposed by combining grey relational analysis (GRA) and hybrid neural network (HNN) to predict coal component content in logging data. First, the correlation degree between different conventional logging data and coal components was calculated using the GRA method, and logging curves with a correlation degree = 0.7 were selected as the input training data set. Then, a back propagation neural network (BPNN), support vector machine (SVM) neural network, and radial basis function (RBF) neural network of different coal components were constructed based on the selected optimal input logging data, and the weighted average strategy was used to form a HNN prediction model. Finally, the GRA-HNN method was used to predict the coal component content of coalbed methane production wells in Panji mining area. The application results showed that the coal component content predicted by the GRA-HNN method has the highest accuracy compared to the logging curve regression method and its single neural network model, with a maximum average relative error of 13.4%. Besides, the accuracy of coal component content predicted by some single intelligent models is not always higher than the logging curve regression method, indicating that the neural network model is not necessarily suitable for all coal component content predictions. The proposed GRA-HNN method not only optimizes the prediction performance of a single neural network model by selecting effective input parameters, but also comprehensively considers the prediction effect of several neural network models, which strengthens the generalization ability of neural network model and increases the log interpretation accuracy of coal component content.