Iron ore pellet firing temperature prediction and abnormality diagnosis based on interpretable machine learning

Accurate control of firing temperature is essential for quality, efficiency and energy consumption in iron ore pellet production. Traditional methods that rely on measured temperatures are limited by sparse measurement points, time lag and signal fluctuations. This study proposes a method based on interpretable machine learning to predict firing temperature and identify key influencing factors and further examines whether such a framework can provide useful short-term diagnosis under real industrial operating conditions. An eXtreme Gradient Boosting model was developed and optimised using historical data from a large-scale chain grate–rotary kiln production line. The model achieved a mean absolute error of 3.10°C, a root mean square error of 4.32°C and an R2 value of 0.73 within a 10-min prediction horizon. SHapley Additive exPlanations analysis indicated the influence of process parameters on firing temperature, demonstrating that kiln head temperature, coal injection rate and cooling exhaust temperature are critical factors. This approach enables intelligent monitoring and optimisation of the firing process, supports abnormality diagnosis and improves production efficiency and energy conservation. Future research will concentrate on improving long-term prediction accuracy and integrating causal analysis for further process optimisation.