Calorific value prediction of coal and its optimization by machine learning based on limited samples in a wide range

Measuring the heating value through experimentation is a laborious process that demands precise instruments and a skilled technician. Due to the challenges associated with experimental determination of calorific value, a wide range of linear and non-linear models have been created as an alternative approach. Unlike the chiefly studies on calorific value estimation, this study represents hyperparameter tuning application especially within Random Forest (RF) and Artificial Neural Network (ANN) methodologies. In addition, limited amount of data in a wide range was considered to establish comprehensive and consistent models. The fundamental aim of the study is to optimize and simplify the model while maintaining satisfactory performance. When RF method was applied, equations including six parameters (fixed carbon, volatile matter, ash, carbon, hydrogen, and sulfur contents) and single parameter (carbon content) gave comparable prediction performances with R² values of 0.968 and 0.961, and mean absolute error (MAE) of 1.101 and 1.134, respectively. ANN, Decision Tree (DT), and Multiple Linear Regression (MLR) methods were also tested. It was concluded that the RF and ANN methods, which uses even a univariate equation of carbon, can provide satisfactory prediction, despite the fact that sample properties changed in wide ranges and the number of data was limited.