Prediction of gross calorific value from coal analysis using decision tree-based bagging and boosting techniques

Area of Science:

• Computational chemistry and materials science
• Machine learning applications in energy resources
• Analytical chemistry and fuel characterization

Background:

• The bomb calorimeter is the standard for determining coal's gross calorific value (GCV), but it is time-consuming and labor-intensive.
• Developing alternative methods using proximate and ultimate analyses is crucial for efficient coal quality assessment.
• Existing empirical models often lack accuracy due to their simplicity and location-specific constraints.

Purpose of the Study:

• To develop and evaluate novel machine learning models for predicting coal's GCV.
• To compare the performance of new tree-based algorithms (Extra trees, Bagging, Decision tree, Adaptive boosting) against established ones (Random forest, Gradient boosting, XGBoost).
• To identify the most significant coal features for accurate GCV modeling.

Main Methods:

• Utilized 7430 data points from the U.S. Geological Survey Coal Quality (COALQUAL) database, including proximate and ultimate analyses.
• Developed and tested Extra trees, Bagging, Decision tree, and Adaptive boosting models for GCV prediction.
• Investigated Random forest, Gradient boosting, and XGBoost, comparing them with empirical models (Schuster, Mazumdar, Channiwala and Parikh, Parikh et al., CFRI).
• Tuned models using exhaustive grid search and evaluated performance using statistical indexes (R², MSE, MAE, MAPE).

Main Results:

• Bagging and boosting techniques achieved a coefficient of determination (R²) over 0.97.
• XGBoost outperformed all other models, yielding an R² of 0.9974 and the lowest error metrics (MSE: 14703.3, MAE: 89.2).
• Oxygen and carbon content were identified as the most significant features for GCV prediction, while volatile matter and sulfur were least significant.

Conclusions:

• Machine learning models, particularly XGBoost, provide a highly accurate and efficient alternative to bomb calorimetry for GCV determination.
• The developed models can significantly reduce the time, cost, and complexity associated with traditional laboratory analyses.
• This approach enables rapid and accurate coal quality assessment, aiding engineers and operators in fuel grading and utilization.