Coalbed methane content prediction based on deep belief network

Coalbed methane (CBM) is considered an unconventional gas resource. Accurate determination of CBM content can provide potential disaster warnings and guide exploration and development. Direct measurement and statistical analysis of CBM content are common techniques. Hoverer，direct measurement methods have high accuracy, but they are time consuming, labor intensive, and inefficient; statistical methods have a limited ability to solve complicated nonlinear problems, for example, CBM content prediction commonly used computational methods do not have high enough accuracy due to the small amount of training data and the shallow model structure. 3D seismic exploration has been widely used in CBM exploration and development due to its small grid size and high resolution. It will improve the accuracy of coalbed methane prediction to combine 3D seismic data with coalbed methane content. Machine learning techniques are a set of computational methods that can learn from data and make accurate predictions. In recent years，many applications of machine-learning techniques for CBM content prediction are found to be more reliable，however the results from traditional machine learning models have errors to some extent. A CBM content model based on Deep Belief Network (DBN) has been developed in this paper, with the input as continuous real values and the activation function as the rectified linear unit. Firstly, various seismic attributes of the target coal seam were calculated to highlight its features, then the original attribute features were preprocessed, and finally the performance of the DBN model was developed using the preprocessed features. Different from conventional DBN models, the proposed model uses continuous real values as the input and the rectified linear unit (ReLU) as the activation function. Training process includes pre training and fine-tuning. Pre training gives the model good initial parameters by training with unlabeled data, and fine-tuning uses a standard supervised method with labeled data to optimize the model. This paper successfully applied a DBN model to predict CBM content from a CBM 3D seismic  prospecting district. With more layers pre trained, the average error decreased from 3.69% to 2.16% and from 2% to 5.76% for the maximum error. Using a pre training strategy to initialize the model’s parameters can improve the accuracy of the model results. Compared with the typical multilayer perceptron（MLP）and the support vector regression（SVR）models, the DBN model has the smallest error, which means it is more accurate in predicting CBM content than the other two models.