Estimation of coal elemental composition from proximate analysis using machine learning techniques

Abstract

ABSTRACT Knowing the properties of coal, which is still the most widely used among primary energy sources, is critical for determining the application area and the technology to be applied. The ultimate analysis results contain important information for the estimation of the gas product composition to be released to the environment as a result of the combustion process. This study aims to make a comparative study using different machine learning models, namely, Support Vector Regression (SVR), Regression Trees (RT), the Ensemble of Trees (ET), Gaussian Process Regression (GPR) for predicting the elemental composition of coal. The moisture, ash, volatile matter, and fixed carbon contents were used as input variables to predict carbon, hydrogen, and oxygen contents of coal. 70% of 6339 coal samples (4437 data) containing different quality coal types were for the training of the models, whereas the remaining 30% sets of data (1902 data) were used to evaluate the prediction performance of the developed models. Analysis results proved that the GPR-exponential model performs best among all models. In the training stage, the correlation of determination (R2), mean absolute percentage error (MAPE), root mean square error (RMSE) and mean absolute deviation (MAD) values were obtained as 0.9970, 0.7525%, 0.6492, 0.4635 for Carbon content; 0.9900, 0.5712%, 0.0646, 0.0297 for Hydrogen content; 0.9951, 4.0664%, 0.7910, 0.4152 for Oxygen content, respectively. Reported indices showed that the GPR-exponential model is a promising procedure for achieving high accuracy and can be used as a reliable model to predict the coal elemental components successfully.