A machine learning framework for intelligent prediction of ash fusion temperature characteristics

Abstract

The characteristics and analysis of coal ash fusion have always been a challenge in coal research. In this study, a hybrid data-based machine learning framework was developed and used to classify coals based on ash fusion temperature. Three main machine learning classification models, namely the Gaussian Naive Bayes (GaussianNB), Random Forest (RF) and Extreme Gradient Boosting (XGBoost), were used in this study, while three commonly used classification metrics were applied to optimize the model. In addition, the SHapley Additive exPlanations (SHAP) method was employed to interpret the selected machine learning models. The results showed that the RF and XGBoost models provided better simulation and prediction results, with all classification metrics exceeding 0.9. Among the models, for the smaller dataset, the RF model provided better simulation results with weighted accuracy (WA), unweighted average recall (UAR) and F1 score (F1) of 0.93, 0.93, and 0.95, respectively. Based on the SHAP analysis, Al2O3, SiO2 and CaO showed the most significant influence on classifications by low ash fusion temperature (LAFT) and medium ash fusion temperature (MAFT) coals. In particular, Al2O3 and SiO2 showed a negative correlation in LAFT classification, while CaO presented a positive correlation. These features had an opposite effect on the classification prediction of MAFT. However, Fe2O3 had a significant negative correlation effect on the classification of high ash fusion temperature (HAFT), second only to Al2O3. In summary, the hybrid data-driven machine learning framework is able to build a classification model of coal ash fusion characteristics with high accuracy based on features such as ash composition and ash fusion temperature. It can also perform in-depth analysis of the model based on the interpretation tool, which ultimately helps to facilitate the analysis of coals.