Coal ash content estimation using fuzzy curves and ensemble neural networks for well log analysis

Abstract

Many important variables for reservoir development and production cannot be derived analytically from continuous well logs. Empirical regression and classification techniques have been widely used to predict these variables from well logs. This approach generally uses data from core analysis and well logs to train a model, which then can be used to estimate a variable where core analysis data are not available. In formation evaluation, the amount of training data is limited or costly to acquire. This may result in regression models having limited predictability. This paper addresses the problem of sparse data by using fuzzy logic and ensemble neural networks to estimate coal ash content from a collection of sparse data. Ash content is a significant parameter to evaluate coal quality and it is usually measured from proximate analysis in the laboratory. Ash content is estimated based on the components of six major oxides (Al2O3, SiO2, K2O, CaO, Fe2O3 and TiO2) by using an X-ray fluorescence technique. We first use fuzzy curve analysis to rank the relationships between well log and ash content data to determine input parameters for estimating ash content. The data sets were then sampled with a bootstrap-aggregating algorithm to create a number of training sets for building ensemble neural networks. The neural networks in the ensembles were trained individually and the outputs were combined to estimate ash content. In total 20 core samples were collected from a New South Wales (Australia) coal bed methane well in the Gloucester Basin and analyzed for ash content. The well was analyzed using density, photoelectric, gamma ray, neutron, acoustic, resistivity, spontaneous potential, and resistivity imaging logging techniques. The tested algorithm produces repeatable ash content prediction (standard deviation of repeated predictions is 0.43%) and effectively reduces the prediction variance and bias compared to the single neural network with early stopping algorithm. The workflow is data-driven and could be used to estimate other complex variables that are required when evaluating coal beds.