Deep learning with TabNet: rapid coal ash content estimation via X-ray fluorescence

Abstract

As a paramount index of coal quality, the quick and accurate prediction of ash content bears substantial significance for the coal preparation plant. In this study, an innovative approach that integrates SHapley Additive exPlanations with a deep learning model Tabular Networks has been introduced, which efficiently predict the ash content by employing the compositional data obtained from rapid elemental analysis. Pictures of the relationship between elemental and ash content reveal a significant correlation. To further elucidate this relationship, Polynomial regression, Random Forest regression, and eXtreme Gradient Boosting are used as comparison prediction models to verify the accuracy of the deep learning model. The R2 for TabNet, Poly, XGBoost, and RFR are 0.9867, 0.9755, 0.9797, and 0.9714, respectively. The optimization is continued based on SHAP averages, and the best combination of elemental feature contributions obtained is used to continue the prediction of the experiments. The experimental results show that the TabNet model has better predictive performance than other models (RMSE, MAE, and R2 are 1.5126, 1.0953, and 0.9903). An examination and exploration of the contributions and roles of elements to the ash prediction model has been conducted. The case study presented in this paper illustrates that the interpreted deep learning model holds promising prospects for industrial ash fraction prediction.