Operating Key Factor Analysis of a Rotary Kiln Using a Predictive Model and Shapley Additive Explanations

Abstract

The global smelting business of nickel using rotary kilns and electric furnaces is expanding due to the growth of the secondary battery market. Efficient operation of electric furnaces requires consistent calcine temperature in rotary kilns. Direct measurement of calcine temperature in rotary kilns presents challenges due to inaccuracies and operational limitations, and while AI predictions are feasible, reliance on them without understanding influencing factors is risky. To address this challenge, various algorithms including XGBoost, LightGBM, CatBoost, and GRU were employed for calcine temperature prediction, with CatBoost achieving the best performance in terms of MAPE and MLSE. The influential factors on calcine temperature were identified using SHAP from XAI in the context of the CatBoost model. SHAP effectively assesses model impacts, accounting for variable interdependencies, and offers visualization in high-dimensional contexts. Given the correlation and dimensionality of variables predicting calcine temperature, SHAP was preferred over Feature Importance or PDP for the analysis. By incorporating seven out of twenty operational factors like burner fuel and reductant feed rate, combustion conditions inside of the rotary kiln and RPM, the calcine temperature increased from 840 °C in 2023 to 910 °C by October 2024, concurrently reducing the electricity unit consumption of the electric furnace by 7.8%. Enhancements to the CatBoost algorithm will enable the provision of guidance values after optimizing key variables. It is expected that managing the rotary kiln’s calcine temperature according to the predictive model’s guidance values will allow for autonomous operation of the rotary kiln through inputting guidance values to the PLC.