A novel explainable modeling method for cleaned coal quality evaluation in jigged fluidized bed

Abstract

ABSTRACT Rapid monitoring of cleaned coal quality with soft sensors is essential for efficient production and control of the jigged fluidized bed. In this study, a new explainable modeling method for refined coal quality prediction in coal-jigged fluidization is proposed. The models were developed based on the 47 process variables of air pressure, buoy height, belt weight, water flow, gate opening, bucket level, hydraulic, and feed frequency. By comparing the performance of the two ensemble learning models, the results show that XGBoost predicts lower MAE (1.16%) and RMSE (1.72%) for cleaned coal quality. In addition, the SHapley Additive explanation (SHAP) is employed to interpret the contributions and analyze the importance of each process variable. The SHAP results indicate that nine important variables (Maximum buoy height of the 3- and 4-bed area, amplitude of gate opening of the 3- and 4-bed area, Gangue belt weight, Gangue bucket level, hydraulic of the 3- and 4-bed area and Feed frequency) contributed the most to ash content prediction of cleaned coal, which Gini coefficient higher than 0.2. Then, the XGBoost predictor is optimized again based on the 9 variables selected based on SHAP contributions. The results show that the MAE and RMSE of the XGBoost model are reduced by 0.24% and 0.40%, respectively. In contrast, the method in this paper effectively addresses the prediction of cleaned coal quality in the jigged fluidized bed, where the contribution of variables obtained by SHAP can be used as a theoretical basis for developing optimal control for jigged fluidized bed.