Limitations Associated with Proximate Analysis-Based Gross Calorific Value Modeling for Coals

This study aims to investigate the limitations associated with proximate analysis-based gross calorific value (GCV) modeling for coals. Toward this, a dataset comprising proximate analysis and GCV data of 4792 coal samples collected from various Indian coal basins was generated, and then a GCV prediction model was developed using the popular multivariate linear regression (MLR) technique. Although the developed model appeared to be acceptable in terms of a prediction R2 value of 0.934, through rigorous statistical analysis, it has been shown that grade misclassification and source-specific biases are inherent limitations associated with such GCV prediction models. It was found that the grade classification accuracy associated with a GCV prediction model was inversely and linearly proportional to the model's associated mean absolute error (MAE) value. It has further been demonstrated that even well-validated GCV prediction models available in literature may perform sub-optimally when utilized for grade classification tasks. The analysis presented in this study also confirms that a source-specific bias can be introduced in the GCV prediction models developed using coal samples from varied geographical sources. It has further been shown that the incorporation of a categorical representation of the sample sources in the GCV prediction model could successfully eliminate the source-specific biases.