Estimating Coal Permeability Using Machine Learning Methods

Abstract

Abstract Bulk permeability of coal is a critical parameter in coalbed methane (CBM) or coal seam gas (CSG) well completion designs and field development planning. The estimation of permeability can be made by well testing either during drilling or production; however, well tests are costly, time sensitive and resource-intensive. Therefore, field-wide estimates are often dependent on production data history-matching, which has a high degree of uncertainty. In this paper, we present a new attempt to apply machine learning approach to estimate coal permeability using drilling data. We first extract important parameters from well test analyses, which are obtained using a packer element testing (PET) tool from four wells in the Surat Basin, Australia. Then drilling data from the wells are processed and fed into different artificial neural networks (ANNs), which include multi-layer perceptrons (MLP), convolutional neural network (CNN), and recurrent neural network (RNN). Two types of models are constructed: regression model with permeability values, and classification model with permeability class intervals (i.e., low, medium, and high permeability values). The evaluation metrics include R2 for regression models and confusion matrix for classification models. Results show that the drilling mud losses are generally higher in coal layers and lower in interburden formations. The predicted medium and high coal permeabilities from MLP and CNN are in good agreement with measured permeability values from PET data. For the classification task, the CNN achieved an overall accuracy of 99%. Thus, an improved coal permeability map with a higher resolution and less calibration against PET data can be developed quickly to aid production data history matching. The developed machine learning model demonstrated a potential to be applied to new wells to predict the coal permeability for the Surat Basin as well as other CSG appraisal projects. This model can allow more rapid optimisation of well spacing, improved downhole pump design, more targeted well stimulation, and overall project economics.